## Introduction

Artificial Intelligence (AI) has been advancing rapidly, with models like Claude Opus 4 representing new frontiers in capability. Alongside these advances, there arises a critical need to ensure that AI technologies are deployed safely and securely, especially to prevent potential misuse in sensitive and dangerous areas such as chemical, biological, radiological, and nuclear (CBRN) weapons development.

To address these concerns, Anthropic has activated the AI Safety Level 3 (ASL-3) protections according to its Responsible Scaling Policy (RSP). This blog post unpacks what ASL-3 entails, why these enhanced protections are necessary, and how they are implemented.

---

## Background: Why AI Safety Levels Matter

Anthropic’s Responsible Scaling Policy (RSP) establishes clear guidelines linking the capability levels of AI models to required safety and security measures. As models become more powerful, the potential risks associated with their misuse grow—especially regarding dangerous applications like CBRN weapons.

Two core areas are addressed:

- **Deployment Measures:** These focus on restricting the AI’s assistance with certain high-risk queries or tasks, particularly those related to CBRN threats.
- **Security Controls:** These protect the AI’s internal core, specifically the model weights, which are the key numerical parameters encoding the model’s intelligence.

Before activating ASL-3, models were typically deployed under AI Safety Level 2 (ASL-2) standards, which include basic measures like refusal to answer dangerous CBRN queries and defenses against opportunistic attempts to steal model weights. ASL-3 introduces more sophisticated protections for both deployment and security to address risks posed by more capable models and increasingly sophisticated attackers.

---

## What Triggers ASL-3 Protections?

ASL-3 protections are triggered when an AI model’s capabilities approach defined thresholds indicating potential misuse risks, or when those risks cannot be conclusively ruled out.

In the case of Claude Opus 4, it was activated under ASL-3 measures as a precaution. Although it has not been definitively established that Claude Opus 4 surpasses the capabilities requiring ASL-3, the rapid progress in CBRN-related knowledge warranted stronger preemptive safeguards.

This cautious approach aligns with the RSP’s philosophy of erring on the side of safety. If, after further evaluation, Claude Opus 4 is found to pose lower risk, the ASL-3 protections can be adjusted or removed.

---

## Deployment Measures Under ASL-3

### Focused Restriction on CBRN-Related Risks

ASL-3 deployment measures do **not** aim to broadly restrict model use but specifically target misuse related to CBRN weapon development. This includes preventing the model from assisting with complex, multi-step workflows that could facilitate creating or acquiring these weapons.

Importantly, these measures do not generally block responses to simple or commonly available CBRN-related information, such as basic chemical formulas.

### Preventing Universal Jailbreaks

A universal jailbreak is a sophisticated method that attackers use to bypass AI guardrails and extract harmful information systematically. ASL-3 aims to limit such exploits through a three-pronged approach:

1. **Making Jailbreaking Difficult:** Anthropic uses Constitutional Classifiers—real-time guards trained on examples of harmful and harmless prompts and completions related to CBRN. These classifiers monitor inputs and outputs to detect and block harmful content.

2. **Detecting Jailbreaks:** Beyond prevention, extensive monitoring systems and a dedicated bug bounty program encourage stress-testing of defenses. This helps identify new jailbreak methods quickly.

3. **Iterative Improvement:** When jailbreaks are detected, synthetic examples are created to train improved classifiers, enabling rapid defense updates.

While these measures might occasionally block legitimate queries (false positives), they represent a significant advancement in mitigating catastrophic risks.

---

## Security Controls Under ASL-3

### Protecting Model Weights

Model weights are large numerical datasets that encode an AI’s learned intelligence. Protecting these weights is crucial because if stolen, the model could be run and potentially misused outside controlled environments.

Anthropic employs over 100 security controls that include:

- **Two-party authorization:** Ensures multiple parties approve access to sensitive resources.
- **Enhanced change management:** Tracks and controls all changes to model access and infrastructure.
- **Endpoint software controls:** Uses binary allowlisting to prevent unauthorized software execution.

### Egress Bandwidth Controls

A unique and effective security measure involves limiting outbound network traffic (egress bandwidth) from secure environments where model weights reside. Because model weights are large, rapidly transferring them out of secure systems is detectable as unusual bandwidth usage.

By imposing rate limits, the system can block suspicious data transfers, making stealthy exfiltration very difficult even if other security defenses fail.

### Continuous Improvement

Security controls evolve with the threat landscape. Anthropic plans ongoing refinement of egress limits, insider threat mitigations, and overall security posture to maintain robust protection.

---

## Looking Ahead: Continuous Evaluation and Collaboration

Determining which safety and security standards to apply to frontier AI models is an ongoing challenge. By deploying Claude Opus 4 under ASL-3, Anthropic gains valuable operational experience to identify and address both technical and policy issues.

Collaboration with the AI industry, government, civil society, and end users is essential for developing best practices. Sharing detailed reports and methodologies can help the wider AI community prepare for the responsible scaling of even more capable AI systems in the future.

---

The activation of AI Safety Level 3 protections reflects a responsible, precautionary approach to managing evolving AI capabilities. Through targeted deployment restrictions and comprehensive security controls, Anthropic seeks to minimize the risks of AI misuse, especially in critical areas like CBRN weapon development.

As AI technologies continue to advance, such layered safety strategies will be key to unlocking their benefits while safeguarding against potential harms.

Robin “rob90” Doetsch

Understanding AI Safety Level 3 Protections: Ensuring Responsible Deployment of Advanced AI Models

## Introduction

The rapid advances in large language models (LLMs) and generative AI (GenAI) have led to a broader range of teams engaging with these technologies beyond traditional data science. However, this increased accessibility brings new challenges around experimentation, trust, and observability. Insights from Google’s Director of AI, Dr. Ali Arsanjani, and Datadog’s VP of Engineering, Sajid Mehmood, shed light on how organizations can effectively approach AI and LLM adoption, especially on platforms such as Google Cloud.

---

## 1. Emphasizing Observability Across the AI Lifecycle

Unlike traditional data science teams familiar with continuous evaluation and experimentation, many software engineers now working with LLMs may lack experience with machine learning workflows. This discrepancy underscores the essential role of observability tooling in the successful deployment and maturity of AI systems.

> As Sajid Mehmood explains, "The strength of an organization’s observability tooling becomes that much more important."

Observability enables teams to trace and understand model behavior, which is critical when performing prompt engineering or retrieval augmented generation (RAG). Without appropriate data and insight, experiments risk becoming inefficient and unproductive.

Dr. Arsanjani adds that observability should be integrated at every stage — from initial prompt engineering and experimentation with various models to advanced tuning and grounding outputs with search capabilities.

**Action 1:** Leaders should ensure AI teams are trained in classic machine learning paradigms of continuous evaluation and provide them with comprehensive observability tools to trace model behavior throughout development and production.

---

## 2. Setting Realistic Expectations: Master Simpler Use Cases First

Many organizations aspire to build complex multi-agent systems, where large models orchestrate smaller specialized models to achieve tasks more efficiently. However, without a solid foundation in fundamental capabilities like prompt engineering and RAG, these large-scale projects risk failure.

Dr. Arsanjani advises organizations to assess current capabilities: skills, platform maturity, and architecture, then set attainable goals aligned with these capabilities rather than rushing into complex architectures.

Moreover, as model capabilities evolve — exemplified by Google’s Gemini team pushing boundaries on context window sizes — foundational skills like evaluation and experimentation remain critical, regardless of shifts in techniques such as RAG.

**Action 2:** Leaders should thoroughly evaluate their organization’s current LLM capabilities and set realistic maturity targets, charting a progressive path that mitigates risks and accelerates deployment.

---

## 3. Building Trust Through Transparent Model Outputs

Confidence in LLM outputs is essential for real-world application. Users must trust the data to make informed decisions; otherwise, even the most advanced AI models are ineffective.

During Datadog’s development of Bits AI, a priority was to provide users with accessible evidence explaining model conclusions — such as relevant telemetry, logs, and traces. This transparency mimics human dialogue, where interlocutors provide reasoning or evidence to support claims.

Techniques like factual grounding — where models cite legitimate sources — further reduce risks of hallucinations and increase reliability. Google’s Gemini team’s foundational efforts to minimize hallucinations complement these post-hoc verification methods.

**Action 3:** Beyond ML optimizations, leaders should prioritize user experience aspects that foster trust, such as clearly cited sources and interactive dialogues enabling users to explore the basis of model outputs.

---

## 4. Leveraging Simple Use Cases for Immediate Impact

While advanced multi-agent systems attract attention, simpler applications like summarization can deliver substantial productivity benefits quickly.

Summarization is easy to evaluate and builds user trust, as users can compare outputs directly with source material. Datadog’s use of incident summarization exemplifies this: automated summaries of incidents are shared in collaboration channels, improving situational awareness and response efficiency.

Additionally, multimodality plays an increasing role, with models like Google’s Gemini capable of processing diverse data types—text, images, video, and audio—to generate comprehensive summaries.

**Action 4:** Technical leaders should identify straightforward LLM use cases, such as summarization, to unlock productivity gains that enhance cross-team communication and operational efficiency.

---

## 5. Evolving Tooling Alongside Organizational Maturity

Organizations initially adopting LLMs might rely on simple prompting with preferred models. However, as they progress toward production-grade applications, robust controls, evaluation mechanisms, and human oversight become necessary.

Google Cloud’s AI Suite aims to provide a unified platform that supports varying levels of maturity—from basic applications to advanced techniques like vector databases for RAG, grounding, and reinforcement learning-based model tuning.

This integrated tooling fosters seamless maturity progression without disruption.

**Action 5:** Leaders should ensure continuous access to appropriate technologies and evaluative tools as teams advance in LLM maturity, minimizing friction during transitions.

---

## 6. Balancing Compute Cost, Performance, and Quality with Observability

Model quality alone is insufficient for practical AI deployment. Cost and performance aspects require monitoring and optimization, which demands specialized observability capabilities.

At Datadog, cluster utilization metrics — including GPUs and TPUs — are closely tracked to optimize resource use. As AI workloads scale, monitoring multi-node GPU clusters and network performance becomes critical.

Datadog’s LLM Observability platform complements their full-stack solution by supporting quality, safety evaluations, and tracing of model chains, providing a holistic view needed to balance compute time, cost, and model effectiveness.

**Action 6:** To optimize AI operations, organizations must implement modern observability solutions that track performance metrics comprehensively and support model quality evaluation.

---

## 7. Defining Metrics that Matter to Drive Effective AI Adoption

Effective AI deployment requires precise understanding and management of metrics that truly impact business outcomes. Overabundance of irrelevant data or focusing on marginal signals can distract teams and increase costs.

Sajid Mehmood emphasizes the importance of establishing clear service level objectives (SLOs) and actionable priorities such as server costs, training efficiency, or user-facing degradation.

Properly scoped metrics help reduce noise and allow organizations to concentrate efforts on improvements that deliver tangible benefits.

Datadog supports this approach by enabling straightforward definition and monitoring of SLOs.

**Action 7:** Leaders should cultivate and communicate a clear understanding of critical metrics tailored to their business, ensuring AI initiatives focus on meaningful impact and cost-effectiveness.

The journey of adopting AI and LLMs involves more than just acquiring the latest models. It requires a nuanced approach that balances technical capabilities, organizational maturity, observability, trust, cost management, and clear business priorities.

By following these seven key insights — covering observability, realistic goal setting, trust-building, practical use cases, tooling evolution, performance monitoring, and metric clarity — leaders can guide their organizations toward successful and sustainable AI integration on platforms like Google Cloud.

For organizations starting or advancing their AI initiatives, aligning strategy with these principles offers a roadmap to maximize value while minimizing risks and inefficiencies.

The Future of AI, LLMs, and Observability on Google Cloud: 7 Key Insights for Leaders

OpenAI has announced the acquisition of io, an AI hardware startup founded by Sir Jony Ive, the renowned former Apple design chief, along with several of his former Apple engineering colleagues. The deal, valued at nearly $6.5 billion, involves approximately 55 employees joining OpenAI, including experienced hardware engineers, software developers, and manufacturing experts. Although Ive himself will not join OpenAI, his design company, LoveFrom, will oversee the design of all OpenAI products, including software.

## Background on io and Jony Ive’s Role

Jony Ive is widely recognized for his influential work at Apple, where he led the design of iconic products such as the iPhone, iPad, and MacBook. After departing Apple, Ive founded io, focusing on creating new AI hardware that rethinks how users interact with technology. He formed the company alongside former Apple engineers Scott Cannon, Evans Hankey, and Tang Tan, aiming to push the boundaries of product design and AI integration.

In a recent interview, Ive criticized some current AI hardware products like the Humane Pin and Rabbit R1, labeling them as "very poor products" and pointing out a general lack of innovation in product design within the AI hardware space.

## Collaboration Details and Strategic Goals

Although Ive will not be an OpenAI employee, his design firm LoveFrom will retain independence and take responsibility for the design of OpenAI's entire product line, including software elements. This collaboration builds on a two-year relationship between OpenAI CEO Sam Altman and LoveFrom, focusing on creating new categories of AI-driven devices.

About 55 members of the io team will transition to OpenAI, contributing their expertise as engineers and manufacturing specialists. These team members include notable figures like Cannon, Hankey, and Tan, bringing decades of collective experience in product development.

Altman emphasized that the forthcoming hardware will represent “a totally new kind of thing” rather than attempting to replace existing products such as smartphones. He drew a parallel to how smartphones did not eliminate laptops, indicating that these new devices will complement rather than compete with current technologies.

## The Upcoming AI Hardware Product

While specific details about the upcoming hardware have not been fully disclosed, both Ive and Altman have expressed enthusiasm about its potential impact. The product is expected to launch in 2026, following several years of research and development by the combined teams.

Altman described receiving a prototype from Ive and living with it, calling it “the coolest piece of technology that the world will have ever seen.” This sentiment underscores the ambition behind the project and the confidence in its transformative capabilities.

According to reports, the product may involve features previously considered, such as headphones or devices with camera capabilities. However, the final design aims to transcend existing categories, emphasizing novel interactions powered by AI.

<iframe width="100%" height="355" src="https://www.youtube.com/embed/W09bIpc_3ms?si=t-Gmqkex_XPgB_gL" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>


## Intersection of Technology, Design, and Human Experience

Both Ive and Altman have highlighted the importance of merging technology, design, and an understanding of human needs to develop truly impactful AI tools. Altman noted that "great tools require work at the intersection of technology, design, and understanding people and the world," praising Jony Ive and his team for their meticulous care throughout the design process.

Ive echoed this vision, expressing that his 30 years of experience have prepared him for this moment. He acknowledged the responsibility ahead but expressed optimism about collaborating with OpenAI's teams, which share values and a vision aligned with creating inspiring and empowering technology.

## Industry Implications

OpenAI’s acquisition of io signals a strategic expansion from software-centric AI development to integrated AI hardware solutions. By onboarding top hardware talent from the design world, OpenAI aims to create AI experiences that are seamless, intuitive, and thoughtfully designed.

This move could redefine how AI is embedded in consumer devices, potentially leading to new product categories that leverage AI in ways not previously realized. The anticipated 2026 launch allows time for rigorous development and refining of the hardware-software integration.

The acquisition of io by OpenAI represents a significant convergence of advanced AI research and industrial design excellence. By combining these strengths, OpenAI aims to pioneer innovative AI-powered devices that enhance how people interact with technology.

While Jony Ive himself will remain independent, his design firm LoveFrom’s involvement ensures the aesthetic and functional aspects of future OpenAI products will benefit from his visionary approach. The tech community will be closely watching as this collaboration unfolds toward the debut of new AI hardware platforms set to launch within the next few years.

OpenAI Acquires Jony Ive’s AI Hardware Startup io in a $6.5 Billion Deal

## Introduction

Large language models (LLMs) have revolutionized many fields by performing tasks such as summarizing text, generating code, and ideating creatively. Building upon these capabilities, AlphaEvolve, a Gemini-powered coding agent, takes this further by targeting complex algorithmic problems in mathematics and computing. It melds the generative creativity of LLMs with automated verification to discover and optimize algorithms across a broad array of domains.

## What is AlphaEvolve?

AlphaEvolve is an evolutionary AI agent that integrates large language models with automated evaluators, enabling it to generate, test, and iteratively improve algorithmic solutions expressed in code. It leverages an ensemble of Gemini models:

- **Gemini Flash**: Designed to maximize idea diversity and breadth.
- **Gemini Pro**: Provides in-depth, high-quality coding suggestions.

Together, these models propose candidate programs, which AlphaEvolve then verifies and scores against evaluation metrics to establish accuracy and quality objectively. This feedback-driven process mimics evolutionary improvement by selecting and refining the most promising algorithms.

## Applications in Google's Computing Ecosystem

AlphaEvolve is not just a research tool but has practical impacts on Google’s infrastructure and AI development.

### Enhancing Data Center Efficiency

One notable success is optimizing Borg, Google’s cluster management system. AlphaEvolve developed a heuristic that improved task scheduling efficiency, recovering roughly 0.7% of global compute resources continuously. This means more workload can be completed with the same computational footprint, translating to cost savings and environmental benefits. Importantly, its human-readable output supports interpretability and ease of maintenance.

### Assisting Hardware Design

In hardware, AlphaEvolve contributed by refining a Verilog description of an arithmetic circuit used for matrix multiplication, a critical operation in AI workloads. After thorough verification ensured functional correctness, Google incorporated this refinement into an upcoming Tensor Processing Unit (TPU). This collaboration between AI-driven proposals and human chip designers accelerates the development of specialized AI hardware.

### Accelerating AI Training and Kernel Optimization

AlphaEvolve has helped optimize fundamental AI operations:

- It improved matrix multiplication kernel performance within the Gemini model’s architecture by 23%, reducing overall training time by around 1%.
- It achieved up to a 32.5% speedup in GPU kernel implementations such as FlashAttention, a key component in Transformer models.

These advancements save substantial computing resources and allow researchers to focus on innovation rather than lengthy manual tuning.

## Advancing Mathematical Research and Algorithm Discovery

AlphaEvolve’s ability to evolve entire algorithms extends beyond computing infrastructure into foundational math problems.

- It developed new matrix multiplication algorithms, surpassing classical results like Strassen’s algorithm for 4x4 complex matrices.
- The system tackled over 50 open mathematical problems spanning analysis, geometry, combinatorics, and number theory.

Remarkably, AlphaEvolve rediscovered known solutions in approximately 75% of cases and improved on previous best-known results in 20%. For instance, it advanced the *kissing number problem* by establishing a new lower bound in 11-dimensional space—a question studied by mathematicians for centuries.

## How AlphaEvolve Works: Technical Overview

The core workflow of AlphaEvolve involves:

1. **Proposal Generation:** Gemini models produce candidate algorithm implementations in code.
2. **Automated Evaluation:** Each proposal is compiled and executed against defined benchmarks or test cases to assess correctness and performance.
3. **Scoring and Selection:** Automated metrics quantify solution quality, guiding selection.
4. **Evolutionary Improvement:** Top-performing solutions are refined through mutations or recombination by the models, iteratively enhancing the algorithm.

This loop repeats until obtaining solutions that meet or exceed established performance thresholds.

## Future Directions and Accessibility

With ongoing improvements in large language model coding skills, AlphaEvolve is expected to discover even more sophisticated algorithms. Google is working on user-friendly interfaces and plans to offer early academic access to foster collaboration. Longer-term goals include wider availability and expansion into diverse fields such as material science, drug discovery, and sustainability, where verified algorithmic solutions could yield transformative benefits.

## Conclusion

AlphaEvolve exemplifies the synergy between advanced AI models and automated verification frameworks to accelerate scientific computing and technology innovation. By evolving algorithms with a human-understandable and verifiable approach, it strengthens both the foundation and application of computational problem-solving. This system not only powers efficiency improvements in large-scale data centers and AI model training but also propels progress in challenging mathematical frontiers.

For those interested, Google invites academic users to register for early access and explore opportunities to collaborate on this cutting-edge technology.

---

For more details and updates, visit the [AlphaEvolve project page](https://ai.googleblog.com/).

AlphaEvolve: Harnessing Gemini-powered AI for Advanced Algorithm Discovery and Optimization

## Introduction

Large Language Models (LLMs), such as GPT-4, have revolutionized the way machines process and generate human-like text. To understand how these models work, it is crucial to start with the foundational concept of **conditional probability**, which helps explain how LLMs predict the next word in sentences.

---

## What is Conditional Probability?

Conditional probability is a statistical measure that calculates the likelihood of an event occurring given that another event has already occurred. It is denoted as **P(A|B)**, which reads "the probability of event A happening given event B has happened."

### Example:

Imagine a group of 14 people, some of whom like Tennis 🎾, some Football ⚽️, a few like both, and some like neither.

- If event A is "liking Tennis" and event B is "liking Football," then P(A|B) measures the probability that a person likes Tennis given that they like Football.

Similarly, in weather forecasting, if event A is "rain today" and event B is "cloudy skies," then the probability of rain given clouds is represented as P(A|B). Since rain is more likely when it is cloudy, this conditional probability would be relatively high.

This idea of updating the likelihood of events considering other related events forms the backbone of how LLMs predict text.

---

## How Large Language Models Use Conditional Probability

An LLM predicts the next word in a sequence by calculating the probability of possible next words **given the words that have come before**. This is effectively a problem of conditional probability.

### Step-by-step prediction process:

1. The model reads a sequence of words (this is called the **context**).
2. For each word in its vocabulary, it calculates the probability that this word could come next, given the sequence so far.
3. The model predicts the next word as the one with the highest calculated probability.

LLMs don't just learn probabilities for individual words but rather a **high-dimensional probability distribution over sequences of words**. The knowledge captured during training is encoded in the model's internal parameters, known as **weights**.

### What are these weights?

- They store the learned patterns and relationships between words after training on massive amounts of text.
- They allow the model to assign probabilities to word sequences based on learned linguistic contexts.

The training process usually involves **supervised learning**, where the model is shown text sequences and learns to predict the next word correctly.

---

## The Challenge of Repetition and Lack of Creativity

If a model were to always select the word with the highest probability, it would often produce repetitive, dull, and predictable text. This behavior limits the usefulness of LLMs for applications like storytelling or creative writing.

To mitigate this problem, LLMs introduce a concept called **temperature** in the prediction process.

## What is Temperature in LLMs?

Temperature is a parameter that controls the randomness of the next word prediction.

- A **low temperature** (<1) makes the model more confident and deterministic. It amplifies the difference in probabilities making the most likely word much more probable.
- A **high temperature** (>1) smooths out the probabilities, making the selection process more random and diverse.

### How temperature works technically:

LLMs use a function called **softmax** to convert raw scores (logits) into probabilities. Temperature adjusts these logits before applying softmax.

The formula is:

\[
P(\text{word}) = \frac{\exp(\text{logit}/T)}{\sum \exp(\text{logits}/T)}
\]

where:
- *logit* is the original score for a word
- *T* is the temperature

When *T* is low, the highest score dominates after applying softmax.
When *T* is high, the probabilities become nearly uniform.

### Sampling vs. Greedy Selection

- **Greedy selection** always chooses the word with the highest probability.
- **Sampling** randomly picks the next word according to the probability distribution.

By sampling with temperature, the model introduces creativity and variation into the generated text.

### Example:

- At temperature 0.1, the model almost always picks the top word, generating very repetitive text.
- At temperature 1.0, the probability distribution is used as is, balancing predictability and diversity.
- At temperature 2.0 or higher, the text may become incoherent due to very random word choices.

---

## Summary

To put it all together:

- Large Language Models rely on **conditional probability** to predict the next word based on context.
- During training, the model learns a probability distribution over sequences of words, captured in its weights.
- The **temperature** parameter controls the randomness and creativity of word prediction.
- Sampling from the probability distribution with temperature prevents repetitive and dull outputs.

Understanding these concepts helps demystify how LLMs generate human-like and varied text.

---

## Further Learning and Resources

If you found this explanation helpful and want to dive deeper into AI, Machine Learning, or LLM training processes, consider following experts like @akshay_pachaar for tutorials and insights.

For foundational reading on softmax and language modeling, these resources are recommended:

- [The Illustrated Transformer](http://jalammar.github.io/illustrated-transformer/)
- [Deep Learning Book (Chapter on Language Modeling)](https://www.deeplearningbook.org/contents/sequence.html)

Feel free to explore these materials to build a more technical understanding of LLMs!

---

Thanks for reading!

Understanding How Large Language Models Work

Duolingo recently announced the launch of 148 new language courses developed with the assistance of generative AI, marking the largest expansion in the company's history. This initiative effectively doubles the platform's existing course offerings, highlighting the transformative impact of AI technology on educational content creation.

## AI-Driven Course Development

Traditionally, building a single language course on Duolingo could take several years. However, the integration of generative AI has dramatically accelerated this process. According to co-founder and CEO Luis von Ahn, while developing the first 100 courses took approximately 12 years, the company managed to create and launch nearly 150 courses in less than one year using AI and automation tools.

These new courses primarily target beginner language learners and include features like Stories to improve reading comprehension and DuoRadio for listening skills enhancement. Duolingo plans to introduce more advanced levels of these courses in the near future.

## The Shift in Workforce Strategy

With the adoption of AI, Duolingo revealed an organizational shift encouraging reduced reliance on human contractors. The CEO shared that the company would "gradually stop using contractors to do work that AI can handle," signaling a move to become an “AI-first” company. Hiring new staff will be contingent on the inability to further automate tasks, emphasizing productivity gains through AI.

This approach is positioned as a strategic move to scale content rapidly and fulfill the company’s mission to make language learning accessible to more users swiftly.

## User Backlash and Quality Concerns

While AI enables impressive scaling, the announcement prompted criticism from users and commentators. Some users expressed concerns about potentially lower-quality and inaccurate content generated by AI, affecting the learning experience adversely. This dissatisfaction has led to some users deleting the app and voicing opposition on social media platforms.

Balancing AI's efficiency and maintaining content quality remains a central challenge for Duolingo as it continues to expand its offerings with AI technologies.

Duolingo's deployment of generative AI to expand its language courses sets a notable precedent in the intersection of education and AI technology. The company demonstrates how AI can drastically reduce the time and labor required to develop educational content. However, the transition also raises important considerations regarding workforce impacts and content quality, which will be critical to address as AI increasingly shapes educational tools and platforms.

Duolingo's AI-Powered Expansion: Launching 148 New Language Courses Amid Controversy

Remember the Great GPU Shortage? It Feels Like Déjà Vu!
If you've tried to buy a graphics card (or GPU) anytime in the last few years, especially during the height of the pandemic, you probably remember the pain. Prices were through the roof, cards were impossible to find, and the dream of building a powerful gaming PC felt out of reach for many.

Well, buckle up, because it looks like we're heading back into choppy waters. Reports from retailers and online marketplaces are showing significant price hikes on the latest graphics cards from both Nvidia and AMD.

Just recently, a local computer store known for sometimes having stock sent out an email: a specific AMD card, the Radeon RX 9070 XT, with a **MSRP** of $599, was available... but for $850. That's a $250 markup right off the bat!

And it's not just one store. Checking around major US retailers and online marketplaces like eBay paints a similar, frustrating picture. Many popular AMD cards are selling for $100, $200, or even $280 over their MSRP. Nvidia cards aren't immune either; the street price for an RTX 5080 is now reportedly over $1,500, a whopping $500 above its $999 MSRP. And the top-tier RTX 5090, with a $1,999 MSRP? Good luck finding one for under $3,000.

Wait, What Exactly is a GPU and Why Is It So Important?
Before we dive deeper into the price madness, let's quickly touch on what a GPU is. GPU stands for **Graphics Processing Unit**. Think of it as the artist of your computer. While your main processor (CPU) handles all the general thinking and calculations, the GPU is specifically designed to handle the complex math needed to create images, videos, and animations on your screen.

For things like playing modern video games, editing high-resolution videos, or working with 3D models, a powerful GPU isn't just nice to have – it's essential. Without a good GPU, your computer would struggle to display detailed graphics smoothly, making games look choppy or even unplayable.

The Data Doesn't Lie: Prices Are Way Up
Let's look at some examples of how street prices compare to the official MSRP:

**AMD Radeon RX 9070 XT:** MSRP $599, Average Street Price ~$957 (eBay), Best Retail Price Seen ~$880*
**AMD Radeon RX 9070:** MSRP $549, Average Street Price ~$761 (eBay), Best Retail Price Seen ~$835*
**Nvidia RTX 5090:** MSRP $1,999, Average Street Price ~$3,871 (eBay), Best Retail Price Seen ~$3,140*
**Nvidia RTX 5080:** MSRP $999, Average Street Price ~$1,533 (eBay), Best Retail Price Seen ~$1,390*
**Nvidia RTX 5070 Ti:** MSRP $749, Average Street Price ~$1,052 (eBay), Best Retail Price Seen ~$825*
**Nvidia RTX 5070:** MSRP $549, Average Street Price ~$715 (eBay), Best Retail Price Seen ~$610

As you can see, the markups are significant across the board, especially for the higher-end cards. Even cards originally marketed closer to their MSRP are now showing up with much higher price tags at major retailers.

So, What's Causing This Price Jump?
It's rarely just one thing, and the situation is complex. Here are a few factors likely playing a role:

<ul>
<li>Retailer Markups (aka "Scalping"): While scalpers buying cards in bulk to resell at inflated prices on sites like eBay were a big problem during the last shortage, it seems some retailers themselves are now listing cards well above MSRP. This was observed even back in early March with supposedly entry-level models.</li>
<li>Not Just Simple High Demand: Interestingly, the volume of cards changing hands on eBay isn't incredibly high right now (around 1,100 new Nvidia and 266 new AMD GPUs listed in the last month). This suggests it might not *just* be overwhelming consumer demand driving prices up, but perhaps limited supply reaching the market or other factors at play.</li>
<li>the Impact of Tariffs: This is a big one that's looming. **Tariffs** are essentially taxes placed on imported goods. Currently, there's something called a **de minimis exemption** which means shipments valued at $800 or less can often enter the country without being subject to these taxes. However, this exemption is set to change on May 2nd. After this date, many more shipments, even those under $800, could be taxed. This means the cost for retailers to import goods, including electronics components like GPUs, could increase. Stores that rely heavily on importing goods, like Shein and Temu, have already started raising prices in anticipation.</li>
</ul>

While tariffs might not be the *sole* reason for the *current* price hikes, the upcoming change could certainly make things worse and push prices even higher in the near future.

What Does This Mean for You?
nIf you've been planning to build a new PC or upgrade your graphics card, this news is undoubtedly frustrating. It means you'll likely have to pay significantly more than the official price, or face long waits for stock.

During the last shortage (roughly December 2020 to July 2022), prices got so bad that some GPUs were selling for *triple* their MSRP! While we're not quite there yet, the current trend is worrying.\

It leaves consumers in a tough spot: Do you bite the bullet and pay the inflated price now, hoping it doesn't get worse? Or do you wait it out, risking prices going even higher or stock remaining scarce?

For those of us who track this market, the question is whether this is a temporary blip or the start of another prolonged period of high prices, potentially exacerbated by trade policies like tariffs. It feels like we're all just waiting to see when it's "safe to enter the water again" without getting hit by a tidal wave of markups.

Hopefully, supply chains stabilize, demand balances out, and the impact of tariffs is manageable. Until then, buying a new graphics card might require patience, a bigger budget, or perhaps exploring alternatives like cloud gaming or integrated graphics if your needs allow.

Uh Oh, Here We Go Again: Why Graphics Card Prices Are Skyrocketing (Again!)

Demis Hassabis, the CEO of Google's DeepMind, has been immersed in the world of Artificial Intelligence (AI) for over 30 years. In a recent interview with 60 Minutes, he unveiled some truly mind-blowing facts about the future of AI, particularly its impact on health and society. These revelations paint a picture of a future dramatically different from today, and it's a future that's arriving faster than many realize.

Let's dive into the key takeaways from Hassabis's interview and explore what they mean for you and me.

<iframe width="560" height="315" src="https://www.youtube.com/embed/1XF-NG_35NE?si=Zafc0hD5TJlzHQKv" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

## Human-Level AI: Coming Sooner Than You Think

Hassabis predicts that we'll achieve Artificial General Intelligence (AGI) – AI with all the intellectual capabilities of a human being – by **2030-2035**. That's just around the corner! This isn't just about faster computers; it's about creating AI that can reason, learn, and understand the world like we do.

Think of it like this: current AI can play chess exceptionally well, but it can't understand the rules of a new board game without being explicitly programmed. AGI, on the other hand, would be able to learn new games, strategize, and even invent its own games, just like a human.

This accelerated timeline means we have less time to prepare for the societal and ethical implications of AGI. It's crucial to start thinking about how we can ensure this powerful technology is used for the benefit of all humanity.

## The End of Scarcity?

Hassabis's most significant revelation is that AI has the potential to **eliminate scarcity**. This might sound like science fiction, but consider the implications. Resources that are limited today – in healthcare, energy, food, and economics – could become abundant tomorrow thanks to AI-driven innovations.

Imagine AI optimizing energy production to the point where clean energy is virtually free. Or AI designing new crops that are resistant to climate change and produce significantly higher yields. The possibilities are truly transformative.

## From Analytical to Creative AI

Today's AI excels at analyzing data and performing tasks based on existing knowledge. However, it struggles with imagination and creativity. As Hassabis points out, current AI systems "can't ask novel questions or form new hypotheses."

But this is about to change. Within the next 5-10 years, AI will likely develop the ability to generate original ideas and scientific theories. This shift from analytical to creative AI will revolutionize entire industries overnight.

Think of AI as a research assistant that never sleeps, constantly generating new hypotheses and experiments. This could lead to breakthroughs in fields like medicine, materials science, and even art.

## AI Revolutionizing Drug Discovery: A Real-World Example

One of the most compelling examples of AI's transformative power is in drug discovery. Before DeepMind, scientists had mapped less than 1% of all protein structures. Proteins are the workhorses of our bodies, controlling virtually every biological process. Mapping just *one* protein used to take years of painstaking research.

DeepMind's AlphaFold AI changed everything. It mapped **200 million protein structures in a single year!** This unlocked a treasure trove of new drug targets and accelerated the development of life-saving treatments.

This breakthrough has earned Hassabis significant recognition, and it's a testament to the power of AI to solve complex scientific problems.

## Targeted Treatments and the End of Major Diseases

With the ability to see exactly how diseases attack cells at the molecular level, scientists can now design precise, targeted treatments that were previously impossible. This is a game-changer for treating diseases like cancer, Alzheimer's, and HIV.

According to Hassabis, AI could help us **cure all major diseases within the next decade or so.** This is because AI can:

- Predict disease before symptoms appear.
- Design perfect molecular interventions.
- Optimize treatment for each individual.

In fact, 9 out of 10 top pharmaceutical companies are already using AI to accelerate their drug discovery efforts.

## Robot Healthcare Workers: A Glimpse into the Future

Get ready for robot healthcare workers! Hassabis predicts a breakthrough in robotics within the next couple of years. 60 Minutes showcased robots that can understand context and adapt in real-time, suggesting they'll soon be able to administer treatments when human providers aren't available.

These robots could be particularly valuable in remote areas or during emergencies, providing access to healthcare for those who need it most.

## The Challenge of AI Consciousness

Perhaps the most intriguing and potentially unsettling revelation is Hassabis's suggestion that AI consciousness may emerge "implicitly." The challenge is that we might not even recognize it because AI systems "are running on silicon" while "we're made of carbon."

This raises profound ethical and philosophical questions. If AI becomes conscious, what rights should it have? How do we ensure that conscious AI is aligned with human values?

We currently have no legal framework to address the possibility of AI consciousness, highlighting the urgent need for discussion and policy development.

DeepMind CEO Predicts AI Revolution: Curing Diseases and Ending Scarcity

## Introduction

Drug discovery is a lengthy and complex process, and effective treatments are available for only a fraction of human diseases. Artificial intelligence (AI) is poised to revolutionize this field, offering the potential to accelerate the identification and development of new therapeutics. A significant breakthrough in this area was the development of AlphaFold, which accurately predicts protein structures for the entire human genome, enabling structure-based drug design.

## AI-Powered Drug Discovery for Liver Cancer

A recent study published in *Chemical Science* details the application of AlphaFold within an end-to-end AI-powered drug discovery platform (Pharma.AI) to identify a novel drug for hepatocellular carcinoma (HCC). HCC is the most prevalent type of primary liver cancer. The Pharma.AI platform includes:

-   **PandaOmics:** A biocomputational engine for target identification.
-   **Chemistry42:** A generative chemistry platform for molecule design.

This study marks the first instance of using AlphaFold to identify a confirmed hit for a novel target in early drug discovery, particularly one without a known crystal structure. The collaborative research was led by Alán Aspuru-Guzik, PhD, Michael Levitt, PhD, and Alex Zhavoronkov, PhD.

> "We decided to go after a project where AI would be used to identify a target for a disease without an existing crystal, use AlphaFold to get the crystal, use another form of generative AI to generate the molecules for this crystal, and then synthesize and test the compounds," said Zhavoronkov. "And it worked!"

## The Power of AI Integration

Michael Levitt emphasized the transformative potential of AI in drug discovery:

> "This paper is further evidence of the capacity for AI to transform the drug discovery process with enhanced speed, efficiency, and accuracy. Bringing together the predictive power of AlphaFold and the target and drug design power of Insilico Medicine’s Pharma.AI platform, it’s possible to imagine that we’re on the cusp of a new era of AI-powered drug discovery."

Alán Aspuru-Guzik highlighted the synergistic effect of combining AI technologies:

> "This paper demonstrates that for healthcare, AI developments are more than the sum of their parts. If one uses a generative model targeting an AI-derived protein, one can substantially expand the range of diseases that we can target. If one adds self-driving labs to the mix, we will be in uncharted territory."

## Overcoming Computational Challenges

Bud Mishra, PhD, commented on the significance of this achievement in the context of computational biology:

> "In 1969, Cyrus Levinthal despaired that with large number of degrees of freedom in an unfolded polypeptide chain, it will be intractable to sift through the molecule’s astronomical number of possible conformations as would be necessary in computational drug design. But imitating nature one can get around this paradox by focusing on ‘only selected easy instances of the hard problem’ as guided by evolution," said Bud Mishra, PhD. "By using large molecular datasets and powerful computers, it has now become possible to engineer AI’s like Alphafold, AlphafoldDB AlphaDesign and RosettaFold, which have enabled Zhavoronkov et al, to recently design CDK20 inhibitors, purely in silico. Their work marks a milestone in computational biology, which will inspire others in taming human suffering, diseases and aging!"

## Identifying and Validating the Drug Candidate

The researchers utilized PandaOmics to pinpoint the HCC protein target and Chemistry42 to generate molecules based on the AlphaFold-predicted structure. They synthesized and tested seven molecules, leading to the identification of a small molecule hit compound for cyclin-dependent kinase 20 (CDK20) within a month. This process typically takes months or years using traditional methods.

A second AI cycle resulted in the identification of an even more potent compound (ISM042-2-048) with enhanced binding (Kd, 566.7 ± 256.2 nM) and inhibitory (IC50, 33.4 ± 22.6 nM) activity against CDK20.

### The Role of CDK20 in HCC

Previous research has shown that CDK20 is often overexpressed in HCC tumor cell lines, promoting cell cycle progression through a positive feedback loop involving the androgen receptor (AR), CDK20, and β-catenin.

> "Therefore, higher binding affinity or enzymatic inhibitory activity in a cell-free system will translate to better anti-proliferation effect in those HCC cell lines with relatively high CDK20 expression," said Zhavoronkov.

Functional assays confirmed that the newly identified molecule selectively inhibited the proliferation of the Huh7 HCC cell line, which expresses high levels of CDK20, compared to the HEK293 non-HCC cell line.

## Insilico Medicine's AI-Driven Approach

Insilico Medicine focuses on leveraging AI to streamline every aspect of drug discovery and development, including:

-   Target identification
-   Novel molecule generation
-   Biomarker development
-   Treatment personalization
-   Clinical trial data analysis

> "We are combining these steps into a comprehensive pipeline, which provides a feedback loop that continually strengthens our pipeline," said Zhavoronkov.

Insilico’s Pharma.AI platform employs meta-learning, zero-shot generative reinforcement learning, and genetic algorithms to discover and design inhibitors for targets lacking structural data.

> "It has experienced exponential increases in performance and quality over the past few years," said Zhavoronkov. "Our platform is built on years of modeling large biological, chemical, and textual datasets in order to discover new targets and design new compounds with desired properties without the use of large molecular libraries."

### Key Components of Pharma.AI

The Pharma.AI platform is powered by two key engines:

1.  **Chemistry42:** A multi-agent reinforcement learning system with 42 generative algorithms that explore chemical space to generate potential drugs. It can optimize small molecule hits and lead compounds, even without a co-crystal structure.

2.  **PandaOmics:** Applies deep learning models to identify therapeutic targets associated with diseases by analyzing omics data from publications, clinical trials, and grant applications. It optimizes targets based on factors like novelty, confidence, and druggability.

> "We’ve used PandaOmics to identify new targets for cancer, amyotrophic lateral sclerosis (ALS), and COVID-19. The novel target it discovered for idiopathic pulmonary fibrosis has been developed into a lead AI-designed novel drug candidate [INS018_055]," said Zhavoronkov.

INS018_055, similar to ISM042-2-048, is a protein kinase inhibitor and the first AI-designed drug to reach Phase I clinical trials.

> "We have invested deeply into AI as a company and have accumulated a lot of data. We followed $2 trillion worth of research data and invested significant time and resources in making the data machine-readable so that it can be used in our AI platform."

## Future Directions

While Insilico Medicine does not plan to advance ISM042-2-048 into clinical trials, the molecule is publicly available for further research.

> "The purpose of the study was to serve as a proof-of-concept of what is now possible with AI—demonstrating that it is possible to use a predicted structure for a novel target and usable chemical data in just 30 days," said Zhavoronkov.

Insilico Medicine and the Acceleration Consortium are also developing self-driving laboratories, integrating AI, automation, and advanced computing to further accelerate drug and material discovery.

AlphaFold Accelerates Drug Discovery: Identifying a Potential Liver Cancer Drug

## Introduction: Ghibli-fying the World, One GPU at a Time

Remember when everyone was obsessed with turning themselves into cartoons? Well, AI has taken it to a whole new level. OpenAI recently rolled out a photorealistic image generation feature in its paid ChatGPT tiers, and it's particularly good at mimicking different art styles. The result? A massive wave of users demanding everything be rendered in the iconic style of Studio Ghibli. So many, in fact, that OpenAI's CEO Sam Altman tweeted that their GPUs were starting to "melt"! Let's dive into this digital frenzy and explore the implications.

### The Ghibli Floodgates Open

On Tuesday, the floodgates opened. ChatGPT users, armed with this new image generation tool, began transforming everything into Ghibli-esque masterpieces. Family photos, classic memes, even a picture of Nvidia CEO Jensen Huang signing a woman's chest got the Ghibli treatment. Altman himself joined the fun, changing his X profile picture to a Ghibli-fied version of himself. It was a digital art revolution, fueled by AI and a shared love for the whimsical world of Hayao Miyazaki.

### Why Ghibli? The Accuracy Factor

So, why Ghibli? According to TechCrunch, ChatGPT's recreations of Studio Ghibli's style were "the most accurate replica of the animation studio's style" compared to other AI image generators like Google Gemini. This accuracy, combined with the studio's widespread appeal, likely contributed to the overwhelming demand. It's like finding the perfect filter that makes everything look magical and nostalgic.

### The Ethical Gray Area: Imitation vs. Inspiration

Here's where things get a little tricky. OpenAI claims it refuses to imitate "the style of individual living artists" but will imitate "broader studio styles." But is Studio Ghibli's style truly a "broader studio style," or is it largely defined by the vision of its founder, Hayao Miyazaki? Miyazaki, who is very much alive, has famously called AI-generated animation "an insult to life itself." This raises a crucial question: where do we draw the line between inspiration and imitation, especially when it comes to AI and the work of living artists?

### Copyright Concerns and Legal Battles

This Ghibli craze also highlights the broader copyright concerns surrounding AI. OpenAI is already facing a lawsuit from The New York Times and other newspapers, accusing the company of illegally scraping their content to train its AI models. The legal battle is ongoing, but it underscores the complex relationship between AI, copyright, and intellectual property. If AI can learn and replicate artistic styles, who owns the rights to those styles, and how do we protect artists from having their work exploited?

### The Bigger Picture: AI, Art, and the Future

This whole situation is a fascinating case study in the rapidly evolving world of AI and art. It demonstrates the power of AI to create and transform, but it also raises important ethical and legal questions. As AI becomes more sophisticated, we need to have serious conversations about copyright, artistic integrity, and the role of AI in shaping our creative landscape. Is AI a tool for artists to enhance their work, or is it a threat to their livelihoods and artistic vision? The answer, like a Ghibli film, is likely complex and nuanced.

### A Cautionary Tale with a Touch of Magic

The ChatGPT Ghibli saga is a reminder that even the most advanced technology can be pushed to its limits by human creativity and enthusiasm. It's also a cautionary tale about the ethical considerations that come with powerful AI tools. While the idea of turning everything into a Ghibli-inspired masterpiece is undeniably appealing, we need to be mindful of the impact on artists and the creative process. Perhaps, in the future, AI can be used to celebrate and support artists, rather than potentially undermining their work. Until then, let's enjoy the Ghibli-fied world with a healthy dose of critical thinking and respect for the artists who inspire us.

ChatGPT's Ghibli Obsession: When AI Art Breaks the System (and Maybe Some Hearts)

## Introduction

Imagine being able to conjure up images directly within ChatGPT, no more switching between apps or wrestling with separate image generators. Well, stop imagining! OpenAI is making it a reality with the integration of new image generation capabilities powered by the **GPT-4o** model, directly into ChatGPT. This feature, aptly named “Images in ChatGPT,” is rolling out now and promises a significant leap in image quality and accuracy.

Think of it like this: you're a chef, and ChatGPT is your sous chef. Before, you had to describe the dish and then go to a separate station to plate it (the image). Now, your sous chef can handle the plating too, and they've gotten *really* good at it.

This upgrade is available across all ChatGPT subscription tiers: Plus, Pro, Team, and even the Free tier. So, whether you're a casual user or a power user, you can start creating images right away.

## What's New and Improved?

So, what makes this new image generation tool so special? According to Gabriel Goh, research lead at OpenAI, this model represents a "step change above previous models." The team leveraged the GPT-4o's "omnimodal" foundation, meaning it can handle various data types like text, images, audio, and video, to create this feature.

Here are some key improvements you can expect:

### Enhanced Binding

One of the most significant improvements is in **binding**. Binding refers to how well the AI maintains the correct relationships between attributes and objects in an image. Previous image generators often struggled with this, mixing up colors, shapes, and other attributes when asked to render multiple items.

For example, if you asked for a blue star and a red triangle, older models might give you a red star and no triangle. This new image generation tool can accurately bind attributes for **15 to 20 objects** without confusion. That's a huge leap in accuracy and reliability!

Think of it like this: imagine you're asking a friend to draw a picture. Before, they might get the details mixed up. Now, they're like a super-attentive artist who gets every detail right.

### Superior Text Rendering

Another major improvement is in **text rendering**. If you've ever tried to generate images with text using AI, you know how frustrating it can be to deal with typos and garbled characters. This new system makes it much easier to generate coherent text without errors.

Goh explained that getting text rendering right was a significant challenge. Even small typos or errors can make an entire image unusable. The team spent months iterating and refining the system to reach a point where the text quality is consistently usable.

While it's not perfect (especially with very small text), the improvement is noticeable. This means you can now create images with clear, readable text for posters, logos, and other applications.

### Autoregressive Approach

So, what's the secret sauce behind these improvements? The system uses an **autoregressive approach** to generate images. This means it creates images sequentially, from left to right and top to bottom, similar to how text is written. This is different from the **diffusion model** technique used by most image generators (like DALL-E), which create the entire image at once.

Goh speculates that this technical difference could be what gives Images in ChatGPT its better text rendering and binding capabilities. It's like the difference between painting a picture stroke by stroke versus spraying it all at once. The former allows for more precision and control.

## Real-World Applications

The OpenAI team demonstrated several examples showcasing the system's capabilities, including:

-   **Scientific diagrams:** Imagine creating a diagram of Newton's prism experiment with correctly labeled components, all generated within ChatGPT.
-   **Multi-panel comics:** Create comics with consistent characters and text bubbles, without having to worry about the AI getting the details wrong.
-   **Informational posters:** Generate posters with accurate text and visuals for presentations, marketing materials, or educational purposes.
-   **Transparent background images:** Easily create images with transparent backgrounds for stickers, logos, and other design projects.
-   **Restaurant Menus:** Quickly generate visually appealing menus with accurate descriptions and pricing.

Jackie Shannon, ChatGPT multimodal product lead, highlighted the system's ability to leverage world knowledge. "If I go to draw an image, I do so with the limitation of my own skill... but also with all of the knowledge of the world that I’ve built up," she explained. "The model brings world knowledge to the equation, so when you ask for an image of Newton’s prism experiment, you don’t have to explain what that is to get an image back."

## A Tradeoff: Latency

There is one tradeoff to consider: the new system takes longer to generate images than before. However, OpenAI suggests that the improved quality and capabilities are worth the wait.

"While we certainly have room to improve on latency…the quality of these images, the capability, the world knowledge, really makes up for the additional seconds that they’ll spend waiting," Shannon said.

## Safeguards and Ownership

Of course, with any powerful technology, there are concerns about potential misuse. The OpenAI team emphasized that the system includes robust safeguards to prevent misuse, such as:

-   Preventing watermark removal
-   Blocking the generation of sexual deepfakes
-   Refusing CSAM generation requests

While the new system doesn't include visual watermarks, all generated images will include standard C2PA metadata to mark them as having been created by OpenAI. The company will also have internal tooling to look up images.

"Ultimately, no system is perfect for this type of thing, but we’re continuously improving our safeguards and we think of this as a starting point," Shannon added. "One thing that’s true about all of the images generated from ChatGPT is that the user owns them and are free to use them within the bounds of our usage policies as they would like."

## What About DALL-E?

For those who are fans of DALL-E, don't worry! Christianson said that users will "still have access via a custom GPT."

GPT-4o Powers New Image Generation in ChatGPT: Say Goodbye to Garbled Text!

## Introduction: DoorDash's Humble Beginnings

Imagine DoorDash, the food delivery giant, starting with just eight PDF menus on a static HTML website and a Google Voice number. That was their version 1.0! It took a mere 45 minutes to build, initially called "Palo Alto Deliveries." This story, shared by Tony Xu, DoorDash's CEO, during an a16z speedrun session, perfectly illustrates the power of the Minimum Viable Product (MVP). 

But while the MVP concept, popularized by Eric Ries in *The Lean Startup*, has become startup gospel, it's not a silver bullet. This post dives into the challenges and nuances of applying the MVP strategy, especially in the age of AI and increasingly competitive markets.

## The MVP's Achilles Heel: Common Pitfalls

While the MVP approach has its merits, product teams often stumble upon several recurring issues during the zero-to-one phase of product development. These include:

- **Inconclusive Testing:** Constant testing yields ambiguous results, making it difficult to determine the right direction.
- **False Negatives:** An overly minimal product might fail simply because it lacks essential features or polish, leading to the incorrect conclusion that the entire idea is flawed.
- **Network Effects Challenges:** MVPs are easier to test for single-player products than for those relying on community or network effects.
- **Justifying Long-Term Research:** Breakthroughs often require significant upfront research and development, which doesn't align with the rapid iteration of an MVP.
- **Data Overload:** Over-reliance on easily measurable data can overshadow crucial qualitative customer insights.
- **Competition in Existing Markets:** MVPs struggle to compete against established products with more features and brand recognition.
- **Local Optimization:** Focusing on incremental improvements can lead to a mediocre product that never achieves greatness.

Let's delve deeper into some of these challenges.

### The Perils of Testing and Interpretation

**1. Expect Inconclusive Results:** Building something new is tough. Expect a lot of failure, especially in uncharted territory. Early product versions often yield ambiguous or even negative results. Was it a lack of features? Poor branding? Flawed onboarding? To make these failures valuable, run clean tests that lead to clear conclusions. Focus on testing one core feature at a time. If it's buried under layers of complexity, the results will be muddled.

**2. False Negatives and the Restart Cycle:** The most dangerous outcome is receiving false information, particularly false negatives. A bare-bones MVP might look weak compared to existing solutions, leading to poor engagement metrics. The risk? You might wrongly conclude that the entire product direction is worthless. This can trigger a complete product reset, a pivot to a new direction, and a cycle of experimentation that generates little real market insight.

**3. The Launch-Relaunch Loop:** Every MVP test requires a follow-up. But existing users are now "tainted" by the previous test. They might reject version 2 simply because they tried version 1. You need a strategy for testing across multiple cohorts. This might involve pulling new users from a waitlist or onboarding self-contained teams or groups onto your platform. Without a clear testing strategy, you'll be limited in the number of tests you can run and how quickly you can learn.

### Leveraging Market Knowledge

While testing is crucial, don't ignore the wealth of knowledge already available in the market. Study the successes and failures of existing products. This is especially valuable when entering an established market. You'll need to build more functionality to compete, but you'll benefit from the accumulated domain expertise in the space. 

This approach is often preferable to pioneering entirely new product categories. With a new category, you have no guarantee that there's a real market need. Even with iteration, you can't be certain it will have the business characteristics you want. For example, travel products often have low usage frequency and high customer acquisition costs. You could learn this the hard way through iteration, or you could recognize these characteristics upfront by understanding the market dynamics.

## The MVP's Applicability: Context Matters

The MVP concept isn't universally applicable. Its suitability depends on the product category and the stage of the market.

<iframe width="560" height="315" src="https://www.youtube.com/embed/KAfjq267r68?si=LIgDCP8yV6cWmupN" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

### Market Maturity and User Expectations

MVPs are often "okay" but not "great" products. And in mature markets, "okay" doesn't cut it. Building a mobile app today is different than in 2010. Today, users expect strong design and polish. You're likely competing against a large number of established apps. In the early days of mobile, competition was low, and even scrappy apps could succeed. MVPs can compete well in the early phase of an S-curve when simply "working" is enough. But in more mature markets, expectations are higher. That's why products like Figma or Notion took years to build their initial versions. Great products often require extra polish, which might seem like low ROI initially, but loyal users appreciate it over the long term. Products that strictly adhere to the MVP theory can end up with clunky UX unless they take the time to refine and polish everything.

### The Need for Upfront Investment

Some categories require significant upfront investment. Consider startups building nuclear reactors, supersonic aircraft, cancer treatments, or humanoid robots. These ventures require hundreds of millions or even billions of dollars just to create a working prototype. While success can lead to immense value, the reality is that these categories demand massive upfront capital, R&D resources, and time. A truly minimal V1 simply isn't possible. This also applies to AI foundation model startups. In these cases, an MVP might not be feasible or even a useful tool.

### The Trap of Data Overload

Over the past decade, metrics have come to dominate product strategy, often at the expense of qualitative insights. A/B testing, analytics, and OKRs drive rigorous execution. However, the data that dominates is often what's easily measurable, not necessarily what drives meaningful outcomes. While incremental metrics can help scale a successful product, they're insufficient for zero-to-one products that need to grow exponentially. Product leaders need to understand the true market opportunity, make strategic decisions, and then leverage quantitative tools to optimize. 

In today's metrics-driven landscape, the biggest opportunities might lie in areas that require intuition. A qualitative approach can be faster than A/B testing, especially for new products with low user numbers. Your initial product direction requires exceptional judgment. Choosing the right starting point is crucial. The ubiquity of metrics-oriented thinking means that breakthrough opportunities often exist where data-driven product leaders won't look. This typically means entering mature markets or categories requiring significant upfront investment. After all, if it were simple and immediately measurable, big tech companies would have already pursued it.

## MVPs in the Modern Era: A Balanced Perspective

My concerns about MVPs shouldn't be interpreted as a complete rejection of the concept. I often advise companies that have been building in isolation for too long to "just ship something already" or to focus on product areas where a V1 is feasible.

However, I've become skeptical of certain product validation approaches, such as landing pages with email capture, social media traction metrics, or viral preview videos. These metrics rarely translate to actual product stickiness. The key is solving existing customer problems, ones that likely have precedent and current solutions in the market, rather than attempting to create entirely new categories from scratch. The romantic notion of ideating cleverly, shipping an MVP, zooming in on a promising feature, shipping another MVP, and repeating ad infinitum often leads teams to iterate endlessly without direction. Instead, products need a strategic starting point in an attractive market, typically validated by the presence of other players in the space.

The startup journey isn't about shipping a single MVP. It's about shipping MVP after MVP in a long series of potential failures, punctuated by occasional glimmers of customer interest. You'll need to repeat this cycle many times, drawing conclusions, raising money, keeping your team aligned, and maintaining customer relationships throughout. This constant iteration is what makes the startup experience so challenging. Reducing your product vision to a single minimal version only gets you through the first few steps of a very long journey.

DoorDash's MVP was just the beginning. The full interview with Tony Xu reveals the countless iterations and tests they ran before finding product-market fit. It took several years before DoorDash's trajectory became clear. The startup journey is a long, challenging road.

Beyond the Hype: Reassessing the Minimum Viable Product in Today's Tech Landscape

## Introduction

In the race to dominate the artificial intelligence (AI) sector, Google has adopted a multi-faceted strategy. Beyond developing its own AI technologies, the internet giant has strategically invested in promising AI startups. Recent legal filings, brought to light during a Google antitrust case, have unveiled Google's significant stake in Anthropic, a high-profile AI company.

This post delves into the specifics of Google's investment in Anthropic, the conditions surrounding it, and the broader implications for the AI industry. We'll examine the motivations behind such investments and the regulatory scrutiny they face.

## Google's Stake in Anthropic: The Details

According to legal documents submitted by Anthropic as part of a Google antitrust case, Google owns **14% of Anthropic**. However, this investment comes with specific limitations designed to prevent Google from exerting undue control over the startup. The filings indicate that Google's ownership is capped at 15%, and the company holds no voting rights, board seats, or board observer rights within Anthropic.

Despite these limitations, Google is poised to further invest in Anthropic through a convertible note, essentially a loan that can be converted into equity. This additional investment, agreed upon in 2023, amounts to **$750 million**. In total, Google's financial commitment to Anthropic exceeds **$3 billion**.

### Convertible Debt Explained

Convertible debt is a common financing tool used in the startup world. It allows investors to provide capital to a company with the option to convert the debt into equity at a later date, typically during a future funding round or an exit event (like an IPO or acquisition). This structure offers several advantages:

-   **Flexibility:** It allows for quicker investment compared to equity financing, as valuation negotiations can be deferred.
-   **Downside Protection:** As debt, it typically ranks higher than equity in the event of liquidation, offering investors some protection.
-   **Upside Potential:** The conversion option allows investors to participate in the company's future success if it performs well.

In Google's case, the convertible debt in Anthropic suggests a long-term commitment to the company's growth and potential.

## Regulatory Scrutiny and Antitrust Concerns

Google's investment in Anthropic, along with similar investments by other tech giants like Amazon and Microsoft in AI startups such as OpenAI, has drawn the attention of regulators. These investments raise concerns about potential anti-competitive practices, specifically whether they give incumbents an unfair advantage in the rapidly evolving AI landscape.

The Justice Department, in its antitrust case against Google, initially proposed forcing Google to divest its AI products that could compete with search, potentially including its stake in Anthropic. This proposal was later withdrawn, with the Justice Department instead suggesting that Google be required to notify government officials before making future AI investments.

Anthropic, in its filings, argued against forced divestiture, claiming it would harm both the company and competition by depressing its market value and hindering its ability to raise capital. Tom Brown, an Anthropic co-founder, stated that severing the relationship with Google would cause "grievous harm" to the startup.

## Anthropic's Perspective and Independence

Anthropic was founded in 2021 by Dario Amodei and Daniela Amodei, former OpenAI employees who left due to disagreements over funding and release strategies. Their vision for Anthropic was to build AI with robust safety guardrails, structuring the company as a public benefit corporation focused on creating public and social good.

According to court filings, Anthropic has actively worked to maintain its independence and avoid being dominated by a single tech giant. The company has raised over $14.8 billion from various venture capital firms, including Menlo Ventures. In 2023, Amazon invested $4 billion in Anthropic, followed by an additional $4 billion the following year.

## The Financial Implications and Interdependence

Google's investment in Anthropic is now worth significantly more than the startup's latest funding round, which valued Anthropic at $61.5 billion. This highlights the potential financial gains for Google from its strategic investment.

However, the relationship between Anthropic and its investors is also characterized by a degree of interdependence. Anthropic relies on Google and Amazon for substantial computing power to develop its AI systems. Furthermore, the company has agreements to use chips and cloud computing services from some of its investors, effectively channeling some of the raised capital back to them.

This arrangement raises questions about the true independence of AI startups that rely heavily on big tech for infrastructure and resources. While these partnerships provide crucial support for development, they also create potential conflicts of interest and dependencies that could influence the direction of AI development.

## The Broader AI Investment Landscape

Google's investment in Anthropic is just one example of the broader trend of big tech companies investing in AI startups. This trend reflects the strategic importance of AI and the desire of established tech giants to maintain a competitive edge in this rapidly evolving field.

Other notable examples include:

-   **Microsoft's investment in OpenAI:** Microsoft has invested billions of dollars in OpenAI, the creator of ChatGPT, and has integrated OpenAI's technology into its products and services.
-   **Amazon's investment in Anthropic:** As mentioned earlier, Amazon has also made significant investments in Anthropic, demonstrating its commitment to the AI space.

These investments allow big tech companies to access cutting-edge AI technologies, talent, and research, while also providing startups with the capital and resources they need to scale and compete. However, they also raise concerns about potential market concentration and the influence of a few powerful players on the future of AI.

Google's investment in Anthropic exemplifies the complex and evolving dynamics of the AI industry. While these investments can accelerate innovation and provide crucial resources for startups, they also raise important questions about competition, control, and the potential for conflicts of interest. As the AI landscape continues to develop, it will be crucial for regulators to carefully monitor these investments and ensure that they do not stifle innovation or create unfair advantages for incumbents. The balance between fostering innovation and preventing anti-competitive behavior will be key to shaping the future of AI.

Google's AI Investment Strategy: A Deep Dive into Anthropic

## Introduction

The scientific community has long dreamed of leveraging Artificial Intelligence (AI) to accelerate the pace of discovery. While early efforts focused on solving complex theoretical problems, a new wave of AI applications is emerging, targeting the more practical, yet time-consuming, aspects of scientific research. Think of it like this: instead of trying to build a robot that can solve quantum physics, we're building a robot that can handle the lab paperwork, freeing up the physicists to focus on the really mind-bending stuff.

Companies like Google DeepMind and BioNTech are betting that the latest generation of chatbots can serve as useful research assistants, automating tasks such as experiment design, protocol development, and data analysis. This could significantly reduce the administrative burden on scientists, allowing them to focus on more creative and strategic aspects of their work.

## Automating the Mundane: A New Era for Research?

Most AI-driven scientific advancements have historically targeted fundamental conceptual challenges, such as protein folding or weather modeling. However, a significant portion of scientific work involves more routine tasks. These include:

- Deciding on the most promising experiments to conduct.
- Developing detailed experimental protocols.
- Analyzing large datasets to identify meaningful patterns.

These tasks, while essential, can consume a significant amount of time, diverting researchers from higher-level thinking and innovation. It's like spending hours organizing your toolbox when you could be building something amazing.

According to the Financial Times, both Google DeepMind and BioNTech are actively developing tools to automate these mundane jobs. This represents a shift in focus, recognizing that even incremental improvements in efficiency can have a significant impact on overall scientific productivity.

### DeepMind's Science-Focused LLM

At a recent event, DeepMind CEO Demis Hassabis revealed that his company is developing a science-focused large language model (LLM) designed to act as a research assistant. This AI could assist in:

- Designing experiments to test specific hypotheses.
- Predicting the outcomes of experiments based on existing data.

Imagine having an AI colleague that can brainstorm experimental designs with you, suggesting potential pitfalls and optimizing parameters before you even step into the lab.

### BioNTech's Laila: An AI Assistant for Biologists

BioNTech, known for its pioneering work in mRNA vaccines, has also embraced AI to enhance its research capabilities. At a recent AI innovation day, the company announced the creation of Laila, an AI assistant powered by Meta's open-source Llama 3.1 model. Laila boasts a "detailed knowledge of biology" and is designed to accelerate the productivity of scientists and technicians.

Karim Beguir, CEO of BioNTech's InstaDeep AI subsidiary, emphasized that Laila is intended to augment human capabilities, allowing researchers to focus on the most critical aspects of their work. During a live demonstration, scientists showcased Laila's ability to automate DNA sequence analysis and visualize the results. According to Constellation Research, Laila comes in various sizes and is integrated with InstaDeep's DeepChain platform, which hosts other AI models specializing in protein design and DNA sequence analysis.

## Precedents and Early Successes

BioNTech and DeepMind are not the first to explore the potential of AI as a research assistant. Last year, researchers demonstrated that combining OpenAI's GPT-4 model with tools for web searching, code execution, and laboratory automation equipment could create a "Coscientist" capable of designing, planning, and executing complex chemistry experiments. This proof-of-concept demonstrated the feasibility of using AI to automate entire experimental workflows.

Furthermore, there is evidence that AI can contribute to strategic research decisions. Scientists have used Anthropic's Claude 3.5 model to generate thousands of novel research ideas, which the model then ranked based on originality. Human reviewers assessed these ideas based on novelty, feasibility, and expected effectiveness, finding that the AI-generated ideas were, on average, more original and exciting than those generated by human participants. This suggests that AI can help researchers break out of conventional thinking and explore new avenues of investigation.

## The Limits of AI in Science: Quality vs. Quantity

While the potential benefits of AI in scientific research are undeniable, it's crucial to acknowledge the limitations. A collaboration between academics and Sakana AI, a Tokyo-based startup, developed an "AI scientist" focused on machine learning research. This AI was capable of conducting literature reviews, formulating hypotheses, carrying out experiments, and writing research papers. However, the research produced was deemed incremental at best, and other researchers raised concerns about the reliability of the output, given the inherent limitations of large language models.

This highlights a critical challenge: simply generating more research papers or results is not valuable if the quality is lacking. As an example, when researchers analyzed a collection of two million AI-generated crystals produced by DeepMind, they found that almost none met the criteria of "novelty, credibility, and utility." This underscores the importance of rigorous validation and human oversight in AI-assisted research.

### The Risk of Turbocharging Paper Mills

Karin Verspoor of the Royal Melbourne Institute of Technology in Australia, points out that academia is already grappling with the problem of paper mills that churn out large quantities of low-quality research. Without careful oversight, AI tools could exacerbate this issue, leading to a flood of unreliable or irrelevant publications. It's like giving a printing press to someone who doesn't know how to write – you'll end up with a lot of paper, but not much of value.

## Augmenting, Not Replacing: A Path Forward

Despite the potential pitfalls, it would be a mistake to dismiss the potential of AI to improve the scientific process. The ability to automate routine tasks could be invaluable, freeing up researchers to focus on the most challenging and creative aspects of their work. The key is to deploy these tools in a way that **augments** human capabilities, rather than replacing them entirely.

Think of AI as a powerful calculator. It can perform complex calculations much faster than a human, but it still needs a human to define the problem and interpret the results. Similarly, AI can automate many aspects of scientific research, but it still requires human researchers to guide the process, validate the findings, and ensure the quality and integrity of the research.

By embracing a collaborative approach, where AI and humans work together, we can unlock the full potential of AI to accelerate scientific discovery and address some of the world's most pressing challenges.

AI as a Research Assistant: Will It Accelerate Scientific Discovery?

## Introduction

Two prominent figures in San Francisco's artificial intelligence landscape are pushing the boundaries of AI testing. Scale AI, a company specializing in data preparation for LLMs, and the Center for AI Safety (CAIS) have launched an initiative called "Humanity's Last Exam." This project challenges the public to devise questions that can effectively test the capabilities of advanced LLMs like Google Gemini and OpenAI's o1.

The initiative offers prizes of $5,000 for the top 50 questions selected for the test. Scale AI and CAIS aim to gauge how close we are to achieving "expert-level AI systems" by engaging a broad range of experts.

## The Need for New Tests

Leading LLMs are increasingly proficient at established tests in areas like intelligence, mathematics, and law. However, the significance of these achievements is questionable. Due to the massive datasets used for training, LLMs may be pre-learning answers, essentially memorizing solutions rather than demonstrating genuine understanding.

Data is crucial to the shift from traditional computing to AI. Instead of explicitly programming instructions, AI systems learn from data. This requires high-quality training datasets and robust testing methodologies. Developers typically use "test datasets," which consist of data not used during the training phase, to evaluate the performance of AI models.

As AI analytics site Epoch AI estimates, by 2028, AIs may have effectively processed all human-written text. This raises the critical question of how to continuously assess AIs once they have access to virtually all existing information. If LLMs aren't already pre-learning answers to standardized tests, they likely will be soon.

## The Challenge of Model Collapse

The continuous expansion of the internet, with millions of new items added daily, might seem like a solution to the pre-learning problem. However, this leads to another challenge known as "model collapse." As AI-generated content increasingly populates the internet and is subsequently used in future AI training sets, the performance of AIs may degrade. This is because the AI is essentially learning from its own outputs, which can amplify errors and biases.

To mitigate model collapse, many developers are collecting data from human interactions with AIs, adding fresh, real-world data for training and testing. This approach helps to keep the training data relevant and diverse.

Some experts suggest that AIs need to become embodied, interacting with the real world and gaining experiences similar to humans. This concept might seem futuristic, but companies like Tesla are already implementing it with their autonomous vehicles. Another avenue involves human wearables, such as Meta's Ray-Ban smart glasses, which can collect vast amounts of human-centric video and audio data.

## The Limitations of Narrow Tests

Even if sufficient training data is available, defining and measuring intelligence, particularly artificial general intelligence (AGI), remains a significant challenge. AGI refers to AI that matches or surpasses human intelligence across a wide range of tasks.

Traditional human IQ tests have been criticized for failing to capture the multifaceted nature of intelligence, which includes language, mathematics, empathy, and spatial reasoning.

A similar issue exists with AI tests. Many established tests evaluate specific tasks such as text summarization, comprehension, inference, human pose recognition, and machine vision. While these tests are valuable, they often provide a narrow view of overall intelligence.

Some tests are being retired because AIs perform too well on them. However, these tests are often task-specific and don't reflect broader intelligence. For example, the chess-playing AI Stockfish significantly outperforms Magnus Carlsen, the highest-rated human player, on the Elo rating system. Yet, Stockfish cannot perform other tasks, such as understanding language. Therefore, equating its chess capabilities with general intelligence would be misleading.

As AIs demonstrate increasingly broad intelligent behavior, the challenge lies in developing new benchmarks for comparing and measuring their progress. One notable approach comes from Google engineer François Chollet, who argues that true intelligence lies in the ability to adapt and generalize learning to new, unseen situations.

In 2019, Chollet introduced the "abstraction and reasoning corpus" (ARC), a collection of visual grid puzzles designed to test an AI's ability to infer and apply abstract rules. Unlike traditional benchmarks that train AIs on millions of labeled images, ARC provides minimal examples, forcing the AI to deduce the underlying logic.

## The Abstraction and Reasoning Corpus (ARC)

The ARC dataset presents AI systems with a series of visual puzzles. Each puzzle consists of a set of input-output pairs. The AI must analyze these pairs, identify the underlying pattern or rule, and then apply that rule to a new input to generate the correct output.

For example, a puzzle might show a series of grids where a specific shape is consistently transformed in a particular way. The AI must learn this transformation and apply it to a new grid with a different initial shape.

The key challenge of ARC is that the AI cannot simply memorize solutions. It must understand the abstract principles governing the transformations and apply them in novel situations. This requires a level of reasoning and generalization that is beyond the capabilities of many current AI systems.

Though humans generally find ARC puzzles relatively easy, achieving high scores on ARC remains a significant challenge for AI. There is a $600,000 prize for the first AI system to achieve a score of 85 percent on the ARC benchmark. Currently, leading LLMs like OpenAI's o1 preview and Anthropic's Sonnet 3.5 score around 21 percent on the ARC-AGI-Pub leaderboard. A more recent attempt using OpenAI's GPT-4o scored 50 percent, but this involved generating thousands of possible solutions before selecting the best one, a method that raises questions about the validity of the result.

## The Future of AI Testing

While ARC represents a significant step forward in testing AI intelligence, the Scale/CAIS initiative highlights the ongoing search for even more effective methods. The fact that the prize-winning questions from "Humanity's Last Exam" will not be published online underscores the importance of preventing AIs from simply learning the test questions.

It is crucial to accurately assess when machines are approaching human-level reasoning, given the significant safety, ethical, and moral implications. Eventually, we will face the even more daunting challenge of testing for superintelligence, a task that requires profound conceptual breakthroughs.


Humanity's Last Exam
https://arxiv.org/abs/2501.14249

Humanity's Last Exam: Challenging AI with Unanswerable Questions

## Introduction: When Robots Start Talking

Remember the movie **Arrival**? A linguist has to decipher an alien language made of palindromic phrases and circular symbols. It's a daunting task, and the movie highlights how different interpretations can lead to conflict. Now, imagine if we faced a similar situation in real life. How would we even begin to understand a language that's completely foreign to us?

That's where the fascinating field of **emergent communication** comes in. It's a research area where we study how artificial intelligence (AI) develops its own languages. And it turns out, this research might be our best bet for understanding not only AI but also, potentially, alien languages.

### The Mystery of Language

But first, let's take a step back. What exactly *is* language? Most of us use it every day to communicate, but how did it even come about? Linguists have been pondering this question for decades. The problem is, language is ephemeral. It doesn't leave any physical traces like bones or fossils. We can't dig up ancient languages to study how they evolved over time. It's like trying to understand how a river formed without ever seeing the riverbed.

### Simulating Language Evolution with AI

So, if we can't study the true evolution of human language, what can we do? Well, perhaps we can simulate it. That's where AI comes in. In emergent communication, we give AI agents simple tasks that require them to communicate. Think of it like a game. For example, one robot might need to guide another to a specific location on a grid, but without showing it a map. We don't tell them how to communicate; we just give them the task and let them figure it out.

Because solving these tasks requires the agents to communicate, we can study how their communication evolves over time. It's like watching a language being born. This gives us a glimpse into how language might have evolved in the first place.

### Human Proto-Languages

This isn't just a theoretical exercise. Similar experiments have been done with humans. Imagine you're paired with someone who doesn't speak your language. Your task is to instruct them to pick up a green cube from a table full of objects. You might start by gesturing a cube shape with your hands and pointing at something green. Over time, you'd develop a sort of **proto-language** together. You might create specific gestures or symbols for “cube” and “green.” Through repeated interactions, these improvised signals would become more refined and consistent, forming a basic communication system.

AI agents do something similar. Through trial and error, they learn to communicate about objects they see, and their conversation partners learn to understand them. It's like watching a new language being created from scratch.

## Cracking the Code: The Challenge of AI Languages

But here's the tricky part: how do we know what they're talking about? If they only develop this language with their artificial conversation partner, how do we know what each word means? A specific word could mean “green,” “cube,” or even both! This challenge of interpretation is a key part of my research.

### The Black Box of AI Communication

Understanding AI language can feel almost impossible. It's like trying to understand a conversation in a language you've never heard before. The challenge with AI languages is even greater because they might organize information in ways completely foreign to human linguistic patterns. It's like trying to read a book written in a completely different alphabet and grammar.

<iframe width="560" height="315" src="https://www.youtube.com/embed/tFMo3UJ4B4g?si=aJVfNT8IJv8g6M3q" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>


### Information Theory to the Rescue

Fortunately, linguists have developed sophisticated tools using **information theory** to interpret unknown languages. Just as archaeologists piece together ancient languages from fragments, we use patterns in AI conversations to understand their linguistic structure. Sometimes we find surprising similarities to human languages, and other times we discover entirely novel ways of communication. It's like being a detective, piecing together clues to solve a mystery.

### Matching Words to Objects

My recent work focuses on using what the agents see and say to interpret their language. Imagine having a transcript of a conversation in an unknown language, along with what each speaker was looking at. We can match patterns in the transcript to objects in the participant’s field of vision, building statistical connections between words and objects. For example, if the phrase “yayo” always coincides with a bird flying past, we could guess that “yayo” is the speaker’s word for “bird.” Through careful analysis of these patterns, we can begin to decode the meaning behind the communication. It's like having a Rosetta Stone for AI languages.

In our latest paper, we show that such methods can be used to reverse-engineer at least parts of the AIs’ language and syntax, giving us insights into how they might structure communication. It's like finally being able to understand the grammar and vocabulary of a language we've never heard before.

## Aliens, Autonomous Systems, and Beyond

So, how does all of this connect to aliens? Well, the methods we're developing for understanding AI languages could help us decipher any future alien communications. If we were to obtain some written alien text together with some context (such as visual information relating to the text), we could apply the same statistical tools to analyze them. The approaches we're developing today could be useful tools in the future study of alien languages, known as **xenolinguistics**. It's like preparing for a first contact scenario, but with the tools of AI.

### Real-World Applications

But we don't need to find extraterrestrials to benefit from this research. There are numerous applications right here on Earth. For example, we can use these techniques to improve language models like ChatGPT or Claude. By understanding how AI develops language, we can make these models more powerful and versatile.

Another exciting application is improving communication between autonomous vehicles or drones. Imagine self-driving cars coordinating their movements using their own language. By decoding these emergent languages, we can make future technology easier to understand. Whether it’s knowing how self-driving cars coordinate their movements or how AI systems make decisions, we’re not just creating intelligent systems—we’re learning to understand them. It's like building a bridge between humans and machines, allowing us to communicate more effectively.

### Conclusion: The Future of Communication

Emergent communication is a fascinating field that's pushing the boundaries of what we know about language and intelligence. By studying how AI agents develop their own languages, we're not only unlocking the secrets of artificial communication but also gaining insights into the very nature of language itself. It's a journey of discovery that could lead to breakthroughs in AI, robotics, and even our understanding of the universe. And who knows, maybe one day, we'll be able to use these tools to finally understand what those aliens are trying to say.

> This article is republished from [The Conversation](https://theconversation.com/decoding-ai-languages-what-we-can-learn-from-robots-talking-to-each-other-229753) under a Creative Commons license. Read the original article.

> Image Credit: Tomas Martinez on Unsplash

> Olaf Lipinski
> Olaf is a PhD student at the University of Southampton, studying artificial intelligence. His main focus is on emergent communication, or how AI agents can develop their own langauge.

Decoding AI Languages: What We Can Learn From Robots Talking to Each Other

## The Data Dilemma in Robotics

Training robots to perform complex tasks is no easy feat. Traditionally, robots were programmed with specific instructions for each task. Think of it like giving a robot a detailed recipe for every action. But, just like a recipe can't account for every variation in ingredients or cooking conditions, these hand-coded robots struggle with the unpredictable nature of the real world. They're brittle, meaning they break down when faced with unexpected situations.

Machine learning offers a promising alternative. Instead of explicit instructions, robots learn from examples, much like humans do. The more examples they see, the better they become. However, this approach introduces a new challenge: the need for vast amounts of realistic training data. Gathering this data in the real world can be incredibly time-consuming, expensive, and sometimes even dangerous. Imagine trying to teach a robot to navigate a cluttered room by having it bump into every object – not very efficient, right?

## The Promise (and Pitfalls) of Simulation

One potential solution is to train robots in simulated environments. Think of it like a video game where you can create any scenario you want. This approach makes it much easier to set up new tasks and environments for the robot to learn in. However, there's a catch: the "sim-to-real gap." These virtual environments, while convenient, are often poor imitations of the real world. Skills learned in these simulations often don't translate well to the complexities of reality. It's like learning to drive in a video game – it's fun, but it doesn't fully prepare you for the real road.

## Enter LucidSim: Bridging the Gap with Generative AI

Now, researchers at MIT's CSAIL have found a way to bridge this gap using a clever combination of simulations and generative AI. They've developed a system that allows a robot to be trained on 100% synthetic data, meaning it never has to see the real world during its training. This is a game-changer!

### How Does it Work?

The core of their approach is a system called **LucidSim**. Here's a breakdown of how it works:

1.  **Realistic Scene Generation:** Instead of relying on traditional simulators that struggle with visual realism, the team uses text-to-image generators. These AI models can create incredibly realistic images based on text prompts. Think of it like asking an artist to paint a picture based on a description. For example, you could ask for "a living room with a red couch and a wooden coffee table," and the AI will generate a realistic image of that scene.

2.  **Diverse Environments:** To ensure the robot is exposed to a wide range of scenarios, the team used ChatGPT to generate thousands of text prompts for the image generator. This allowed them to create a vast library of diverse environments, from kitchens to parks to cluttered offices. It's like giving the robot a diverse set of experiences to learn from.

3.  **Motion Simulation:** Once the realistic images are generated, another system called **Dreams in Motion** comes into play. This system takes a single image and creates a short video from the robot's perspective. It calculates how each pixel in the image would shift as the robot moves through the environment. This is crucial because it provides the robot with a sense of motion and depth, which is essential for navigation.

4.  **Combining Physics and Visuals:** The generated images are then combined with data from a popular physics simulator called MuJoCo. This allows the robot to understand the geometric and physical properties of the environment, such as the shape of objects and how they interact with each other. It's like giving the robot both eyes and a sense of touch.

### Training the Robot Dog

With this data-generation pipeline in place, the researchers trained an AI model to control a quadruped robot (think of a robot dog) using only visual input. The robot learned a variety of locomotion tasks, including:

-   Going up and down stairs
-   Climbing over boxes
-   Chasing a soccer ball

### The Training Process

The training process was split into a few key steps:

1.  **Expert Guidance:** First, the model was trained on data generated by an expert AI system that had access to detailed terrain information. This gave the model a basic understanding of the tasks it needed to perform. It's like giving the robot a head start with a knowledgeable tutor.

2.  **LucidSim Data:** Next, the model was trained on data generated by LucidSim. This allowed the robot to learn how to perform the tasks in a visually realistic environment. It's like giving the robot a chance to practice in a more realistic setting.

3.  **Combined Training:** Finally, the model was re-trained on the combined data from both the expert system and LucidSim. This created the final robotic control policy, allowing the robot to perform the tasks effectively.

## Impressive Results

The results were impressive. The robot, trained entirely on synthetic data, matched or outperformed the expert AI system on four out of five tasks in real-world tests. And on all tasks, it significantly outperformed a model trained using "domain randomization," a leading simulation approach that increases data diversity by applying random colors and patterns to objects. This shows that LucidSim is not just creating more data, but *better* data.

## The Future of Robot Training

This research has significant implications for the future of robotics. By using generative AI to create realistic training data, we can overcome the limitations of traditional simulation and accelerate the development of more capable and adaptable robots. The researchers are already planning to use LucidSim to train a humanoid robot and improve the dexterity of robotic arms. 

Given the insatiable need for robot training data, methods like LucidSim are likely to become increasingly important in the coming years. It's a big step towards a future where robots can learn and adapt to the real world without needing to be constantly supervised or manually programmed. It's like giving robots the ability to learn from their mistakes and become more independent, just like us.

## Key Takeaways

-   **The Challenge:** Gathering enough realistic data is a major hurdle in training robots.
-   **The Solution:** MIT researchers developed LucidSim, a system that uses generative AI to create realistic synthetic data for robot training.
-   **The Method:** LucidSim combines text-to-image generators, motion simulation, and physics simulations to create diverse and realistic training environments.
-   **The Results:** A robot dog trained entirely on synthetic data matched or outperformed expert systems in real-world tests.
-   **The Future:** This approach has the potential to revolutionize robot training and accelerate the development of more capable and adaptable robots.

This research is a testament to the power of combining different AI techniques to solve complex problems. It's a glimpse into a future where robots can learn and adapt to the real world with greater ease and efficiency, thanks to the magic of synthetic data.

Robot Dog Learns Tricks Using 100% Fake Data: How AI is Bridging the 'Sim-to-Real' Gap

## The AI Gold Rush: OpenAI's Potential $300 Billion Valuation

It seems like every other day there's a new headline about AI, and for good reason. The technology is rapidly evolving, and companies like OpenAI are at the forefront of this revolution. Recently, there's been a lot of buzz around a potential new funding round for OpenAI that could value the company at a staggering $300 billion. Let's dive into what this means and why it's such a big deal.

### The Numbers Game: A Potential $40 Billion Investment

According to reports, OpenAI is looking to raise up to $40 billion in funding. Now, that's not pocket change! This kind of investment would not only fuel their ongoing research and development but also solidify their position as a leader in the AI space. The valuation being discussed is around $300 billion, which is a huge jump from previous valuations. To put that into perspective, it's like a small country's GDP being assigned to a tech company. 

### SoftBank: The Potential Big Player

One of the key players in this potential deal is SoftBank. They're reportedly considering leading the investment round and could contribute up to $25 billion. SoftBank is known for making big bets on technology companies, and this would be another example of their aggressive investment strategy. They're not afraid to put their money where their mouth is, especially when it comes to disruptive technologies like AI.

### Project Stargate: A $500 Billion Datacenter

But the story doesn't end there. SoftBank also announced a massive $500 billion datacenter project with OpenAI, codenamed "Stargate." This project is designed to provide the massive computing power needed to train and run advanced AI models. Both OpenAI and SoftBank are reportedly planning to invest $19 billion each into this project. This shows the scale of infrastructure required to support the future of AI and the commitment both companies have to this technology.

### The Debt Factor: SoftBank's Financial Strategy

It's worth noting that SoftBank often funds its investments with debt. They currently carry around $150 billion on their books. This strategy allows them to make large investments quickly, but it also comes with its own risks. It's like using a credit card to buy a house – it can work out great if everything goes according to plan, but it can also lead to trouble if things don't go as expected. 

### What Does This Mean for the Future of AI?

So, what does all this mean for the future of AI? Well, for starters, it shows that the AI industry is not slowing down anytime soon. The massive investments being made by companies like SoftBank and OpenAI indicate a strong belief in the potential of AI to transform various industries. This kind of funding will allow OpenAI to continue pushing the boundaries of what's possible with AI, leading to more advanced models, better applications, and potentially even more groundbreaking innovations. 

### The Potential Impact

- **Faster Development:** More funding means more resources for research and development, leading to faster progress in AI technology.
- **Infrastructure Growth:** Projects like Stargate will provide the necessary infrastructure to support the growing demands of AI models.
- **Wider Adoption:** As AI becomes more powerful and accessible, we can expect to see it being adopted in more industries and applications.
- **Increased Competition:** The success of OpenAI will likely spur more competition in the AI space, which could lead to even more innovation.

### Final Thoughts

The potential $300 billion valuation for OpenAI is a clear sign that the AI revolution is in full swing. With massive investments from companies like SoftBank and ambitious projects like Stargate, the future of AI looks incredibly promising. While there are risks involved, the potential rewards are enormous. It's an exciting time to be witnessing the rapid evolution of this technology, and it's safe to say that we're only just scratching the surface of what's possible with AI.

OpenAI's Potential $300 Billion Valuation: What's the Buzz?

## The End of an Era: Moore's Law and Its Limits

For decades, we've relied on Moore's Law, the principle that the number of transistors on a microchip doubles roughly every two years. This has fueled the digital revolution, enabling everything from smartphones to the internet. However, we're now reaching the fundamental limits of silicon-based electronics. Transistors are approaching the size of individual atoms, and further miniaturization is becoming impossible due to quantum and thermodynamic uncertainties. Simply put, we can't shrink them any further.

This has led to a shift in how we approach improving computer performance. Instead of making transistors smaller, the focus is now on making computers bigger and more interconnected. Simultaneously, our ambitions for artificial intelligence are skyrocketing. The last seven years have seen incredible advancements in AI, with compute power increasing dramatically. However, networking is now becoming the bottleneck. Achieving Artificial General Intelligence (AGI) will require vast amounts of computational power and data, potentially far beyond what current technology can handle.

So, is this the end of computational progress as we know it? Or is there a path beyond the atom, towards the speed of light?


<iframe width="560" height="315" src="https://www.youtube.com/embed/wBqfzj6CEzI?si=BvfFPPJ5xDJ9NsgB" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

## The Rise of Photonics: Computing with Light

Many believe the answer lies in photonics, the science and technology of generating, detecting, and controlling photons, or light. While photonics has been a popular subject in computational literature, it has never seen mainstream commercial usage until now. We've reached a point where we can build photonics at a scale previously unimaginable.

Light has been revolutionizing how we send information since the 1800s, with the advent of flashing lights and mirrors for signaling. The real breakthrough came in the 1960s with fiber optic cables, thin strands of glass that trap light and use it to transmit data. This technology now underpins most of the internet, with vast networks of undersea cables connecting us globally. While we've mastered sending information using light, computing with it has proven much more challenging.

Scientists have dreamed of optical computers since the 1980s because light is much faster and more efficient than electricity. However, controlling light for computation is incredibly tricky. Light tends to spread out and scatter, and getting it to make sharp turns or stay in tiny spaces goes against its nature. It's only now that science and companies like Lightmatter are making photonics a viable technology.

Unlike traditional electronics that use electrical signals through wires, photonics uses optical fibers or waveguides on a chip. This allows data to be transmitted faster over longer distances with less energy loss. Lightmatter is at the forefront of this revolution, integrating photonic components directly into computer chips, enabling a new method of chip communication that could pave the way for AGI.

## Lightmatter: Pioneering the Future of Computing

To better understand this technology and the limitations of current computational methods, we visited Lightmatter's Palo Alto lab and spoke with CEO and co-founder Nick Harris.

> "We are enabling these AGI supercomputers that are being built toward realizing these machines that are able to think at the same caliber as the best humans on the planet," says Nick Harris.

### The Challenge of AGI

AGI, or Artificial General Intelligence, is an AI that can understand, learn, and perform any intellectual task at a human level or better. The recent AI boom has made this goal seem more tangible, but achieving it requires a massive leap in computational power. 

### The End of Dennard Scaling

For decades, we also benefited from Dennard scaling, which stated that as transistors got smaller, they would use less power and run faster. This is why your phone today is more powerful than a room-sized computer from the 1970s. However, around 2006, this free lunch ended. We hit fundamental limits of physics, and making chips better became much harder and more expensive. This is why companies like Lightmatter are so exciting – they're trying to reinvent computing itself.

### Light's Potential

Nick Harris's journey into computing began with gaming and programming. He eventually worked at a semiconductor company, studying transistors. He realized that there were fundamental challenges in continuing to improve computers using traditional methods. This led him to explore the potential of light. He went to MIT to study photonics using silicon, the same material used to build computer chips. He also studied quantum information theory and quantum computing. 

> "What I learned along the way was intricately how light works and what the future might look like with light compared to electrons," says Harris.

Light is ideal for communication and transmission of information. It enables faster and more reliable interconnects over long distances because it travels at the speed of light. It also generates less heat and consumes less energy, making it sustainable for large-scale systems like AI supercomputers.

### Lightmatter's Products: Passage and Envise

Lightmatter is the first company to commercialize photonics technology. They have two main products: **Passage** and **Envise**.

**Passage** is a 200mm by 200mm chip built in 2019. It consists of an array of 48 maximum-size chips stitched together with optical waveguides. These waveguides allow light to be manipulated and encoded with information, enabling GPUs and switches to build data center-scale supercomputers. Optical fibers, similar to those used for internet connections, are attached to the side of the photonic chip. By programming the chip, the connections between the 48 tiles can be configured. This allows customer GPUs, switches, or CPUs to communicate using light.

> "By using silicon photonics and using light, we would actually be able to drive the energy efficiency, drive the speed, and continue the roadmap for computing," explains Harris.

The next-generation **Passage** chip can communicate at over 100 terabits per second. It has 256 attached fibers, the largest optical fiber count in the world, and the highest bandwidth chip in the world. It acts as a superhighway for data passing between different chips, enabling unprecedented bandwidth and ultra-low latency communication. This addresses the "shoreline problem" where there isn't enough space to plug electrical wires into the chip. Photonics sidesteps this issue by using optical waveguides within the chip itself, allowing data to be transmitted through the entire chip, massively increasing potential bandwidth.

**Envise** is Lightmatter's all-in-one photonic computing chip. It combines traditional electronics with photonics to perform computations using light. This is a completely different way of thinking about computation. Data flows through the computer at the speed of light, using different colors to process multiple things simultaneously. Traditional computers are limited by the speed of electrical signals, but light moves through these chips without delay. Lightmatter also stacks what they call "virtual processors" on top of one another, creating multiple computers in the same physical space. All of this is achieved while using far less power than regular chips.

## The Bottleneck: Interconnects and Networking

As overall computing performance increases, the bottleneck is no longer the chips themselves but the networking or interconnects between them. Networking is the communication and data exchange between multiple processors or chips within a computing system. It's critical for efficient performance in modern multi-chip systems.

> "The last seven years has been an incredible time for AI and a lot of what I saw is the unbelievable increase in the pace of computing... but a lot of those advances are starting to hit certain walls and most of that has to do with the connectivity," says Harris.

Transistors are now so small that they're approaching the size of an electron, meaning we can't shrink them anymore. The world has decided that the first path towards making computers better is making them bigger. However, the challenge is how to interconnect them at scale. The limitations of transistors are troubling because progress in almost every field is tied to progress in computing. If progress slows down, it will impact our ability to simulate the world, perform calculations, and connect in real-time.

### The Shoreline Problem

Steve Clinger, VP of product at Lightmatter, explains that the size and complexity of AI models have increased dramatically. To train these large language models, you need a huge number of compute units that can communicate at very high speed with very low latency. In traditional designs, the points where data escapes are limited by the electrical shoreline of the chip. You run out of space to put the high-speed signals that are sending all this data. Lightmatter's technology allows bandwidth to escape anywhere on the chip, dramatically scaling the amount of bandwidth that can be simultaneously communicated outside of these chips.

> "What we're doing is we're enabling these supercomputers to act as a single massive chip," says Clinger.

Lightmatter is replacing the traditional silicon substrate with a new one that enables chips to send data using light. This allows them to slipstream next-generation technology into the existing ecosystem.

## Lightmatter's Clean Room: Building the Future

Lightmatter has built its own mini-fabrication facilities and uses advanced microscopy to develop and inspect their chips. This allows them to rapidly prototype and iterate on their designs. R Tesh Jane, senior VP at Lightmatter, gave us a tour of their clean room facility.

> "This is an end-to-end capability where we're going from a silicon that's showing up on a wafer all the way out to a packaged unit that has fibers attached that can be deployed into a system," explains Jane.

The clean room includes the fiber attach process, where fibers are aligned and attached to the chip. The waveguides, which are etched into the silicon, guide the light into the chip. Each fiber carries 16 colors of light, all around a wavelength of 1300 nm. The light is coupled into the waveguide, which is a few hundred nanometers wide and about 100 nanometers tall. Inside this tiny device, 16 colors of light travel together.

### The Physics of Light

Light is interesting because it relies on a fundamentally different set of physics than electronics. Electronics are governed by properties like resistance, inductance, and capacitance, which describe power dissipation and switching speed. Light, on the other hand, doesn't have these properties. It operates in the hundreds of terahertz regime, while electronics operate in the gigahertz regime. Light also enables parallel communication in a way that electrical signals cannot.

Silicon photonics has been developed over the past 20 years, but the challenge was that there were no foundries or commercial partners to scale the technology. Additionally, electrical wires were still making sufficient progress, so there wasn't a strong push to explore the potential of light. Lightmatter has built photonics at a scale that no one has ever seen before, with millions of optical components on a chip.

> "I love chips. Every time I get to see one of the chips that we're building, it blows my mind," says Harris.

## Conclusion: A New Era of Computing

Developing this kind of technology isn't easy. It requires significant investment, cutting-edge research, and the perseverance to overcome technical challenges. However, the potential rewards are immense. The death of Moore's Law doesn't mark the end of technological or economic progress. Instead, it marks the beginning of a new age of computing, limited only by the speed of light. Lightmatter is at the forefront of this revolution, paving the way for a future where light is the driving force behind computation and AI.

Beyond Moore's Law: How Light is Revolutionizing Computing

## Introduction

The Stargate AI data center project, a joint venture between OpenAI, Oracle, SoftBank, and MGX, is set to utilize solar power and battery storage as part of its energy solution. This move underscores the increasing importance of sustainable energy practices in the operation of large-scale computing infrastructure, particularly for AI model training and deployment.

## Stargate Project Overview

Stargate is designed to provide the substantial compute power required to train and run OpenAI's large language models and other AI tools. The project, currently under development in Texas, is expected to have significant power demands. While the exact power requirements have not been officially disclosed, reports suggest that OpenAI has considered plans for data centers with capacities as high as 5GW. This scale of operation necessitates a robust and reliable energy infrastructure.

## SB Energy's Role

According to a Bloomberg report, SB Energy, a subsidiary of SoftBank specializing in solar and battery solutions, will be a key energy provider for the Stargate project. This indicates a strategic move towards integrating renewable energy sources into the data center's power supply. SB Energy's involvement aligns with the broader trend of tech companies seeking to reduce their carbon footprint and enhance energy sustainability.

## Solar and Battery Integration

The integration of solar power and battery storage is a significant step towards a more sustainable data center operation. Solar power provides a clean, renewable energy source, while battery storage helps to address the intermittent nature of solar energy by storing excess power for use when sunlight is not available. This combination ensures a more consistent and reliable power supply for the data center.

### SB Energy's Existing Projects

SB Energy already has a presence in Texas, with several projects aimed at powering data centers. The Orion Solar Belt, consisting of three projects (Orion I, Orion II, and Orion III) in Milam County, has a combined capacity of 900MW. These projects, with individual capacities of 258MW, 302MW, and 315MW, respectively, are set to power Google's data centers. This existing infrastructure and experience position SB Energy as a reliable partner for the Stargate project.

## Potential for Other Energy Sources

While solar and battery solutions are a crucial part of the Stargate energy plan, they are unlikely to meet the entire power demand. OpenAI is reportedly exploring other energy sources, including natural gas and nuclear power. OpenAI CEO Sam Altman's investments in small modular reactor firm Oklo and nuclear fusion startup Helion suggest a potential interest in advanced nuclear technologies for future energy needs.

## Modernizing the US Energy Grid

OpenAI has stated that it is considering various options for how Stargate can contribute to modernizing the US energy grid. This includes exploring technologies ranging from nuclear power to advanced battery solutions. The project's scale and energy requirements could potentially drive innovation and investment in these areas, benefiting the broader energy landscape.

## Conclusion

The integration of solar power and battery storage into the Stargate AI data center project represents a significant step towards sustainable computing. While the project will likely require additional energy sources, the commitment to renewable energy highlights the growing importance of environmental responsibility in the tech industry. The involvement of SB Energy and the exploration of other advanced energy technologies demonstrate a forward-thinking approach to powering the future of AI.

Solar and Battery Power Planned for OpenAI's Stargate Data Center

## Introduction

The AI landscape is rapidly evolving, and recent developments have brought two contrasting visions for the future of artificial intelligence into sharp focus. On one side, we have DeepSeek, a Chinese chatbot that has demonstrated impressive performance at a fraction of the cost of its competitors. On the other, there's Stargate, OpenAI's ambitious data center venture aimed at scaling up AI development. These two projects represent not just different technological approaches but also a new front in the global AI arms race.

## DeepSeek: A Cost-Effective Challenger

DeepSeek is a chatbot, similar to ChatGPT, that allows users to interact with an AI through text. What sets it apart is its performance and cost-effectiveness. The company's recent R1 reasoning model, released shortly after its V3 model, has shown benchmark performance comparable to, and in some cases exceeding, models from OpenAI, Meta, and Google. Notably, DeepSeek's models are free to use.

### The Cost Advantage

While OpenAI's GPT-4 is estimated to have cost around $100 million to train, DeepSeek claims its R1 model was trained for less than $6 million. This significant cost difference has shaken the AI industry, which has largely operated under the assumption that scaling up computational resources is the key to unlocking AI's full potential. DeepSeek's success suggests that there may be more efficient ways to train powerful AI models.

### Market Impact

The rapid rise of DeepSeek has had a significant impact on the stock market. The app quickly became the number one app in the US, and the news of its capabilities triggered a massive selloff in tech stocks, wiping more than $1 trillion off the market. Nvidia, a major player in AI hardware, experienced a record stock market decline, losing nearly $600 billion in value. This market reaction underscores the potential disruption that DeepSeek represents.

## Stargate: OpenAI's Scalability Push

In contrast to DeepSeek's cost-effective approach, OpenAI is pursuing a strategy of massive scale. Sam Altman recently announced a $500 billion data center venture called Stargate, designed to significantly increase OpenAI's capacity to train and deploy new AI models. This project represents a belief that continued investment in computational resources is essential for advancing AI capabilities.

### The Closed vs. Open Debate

The contrasting approaches of DeepSeek and Stargate highlight the ongoing debate between closed and open-source AI development. Stargate represents a closed, expensive model, where resources and control are concentrated in the hands of OpenAI and its partners. DeepSeek, on the other hand, embodies a more open and accessible approach, raising questions about data privacy, censorship, and the true cost of its development.

## The Geopolitical Implications

The emergence of DeepSeek and Stargate also has significant geopolitical implications. These projects represent a new front in the AI arms race, particularly between the US and China. DeepSeek's success, in particular, has been described as "AI's Sputnik moment," highlighting the potential for a shift in global AI leadership. Furthermore, ongoing trade tensions, such as President Trump's threat of tariffs on foreign semiconductors, add another layer of complexity to this already fraught situation.

## Conclusion

The AI industry is at a critical juncture. The contrasting approaches of DeepSeek and Stargate demonstrate that there is no single path to AI advancement. DeepSeek's cost-effectiveness challenges the prevailing notion that scale is the only way forward, while Stargate represents a continued commitment to massive investment in computational resources. As the AI arms race intensifies, the geopolitical implications of these developments will become increasingly important. The future of AI will likely be shaped by the interplay between these competing visions and the global dynamics they represent.

> "This is becoming especially fraught, as President Donald Trump continues to wreak havoc on foreign relations with a new threat of tariffs on foreign semiconductors."

DeepSeek vs. Stargate: The New AI Arms Race

## Introduction

In a recent interview, Google DeepMind CEO Demis Hassabis shared insights into the transformative impact of their AlphaFold AI system. He highlighted its ability to accelerate scientific research and its potential to revolutionize drug discovery. This blog post will delve into the key points of the interview, exploring the technical aspects of AlphaFold, its applications, and the broader implications of AI in science and society.

## AlphaFold: A Billion Years of PhD Work

Hassabis described AlphaFold as a program that can predict the 3D structure of a protein from its amino acid sequence. This is a significant breakthrough because a protein's 3D structure determines its function. Understanding this structure is crucial for various scientific endeavors, including disease research and drug design.

> "The 3D structure of the protein... is what determines their function. So if you're interested in disease or or or how biology works or you're trying to design a drug you need to understand the 3D structure of the protein."

Traditionally, determining a protein's structure is a time-consuming and resource-intensive process, often taking years for a single protein. AlphaFold has drastically accelerated this process. According to Hassabis, the system has effectively performed a billion years of PhD-level work, providing the structures of over 200 million proteins. This vast dataset is freely available to researchers worldwide, fostering rapid scientific progress.

## The Nobel Recognition and AI's Coming of Age

The development of AlphaFold led to Hassabis and his team receiving the Nobel Prize, a recognition that he sees as a validation of AI's potential to contribute to solving deep scientific challenges. The Nobel committee's decision to award the prize for AI-driven research signals a shift in how the scientific community views AI, acknowledging its maturity and impact.

> "It was a kind of maybe a coming of age or recognition that AI can really is now mature enough to really contribute to you know helping solve some of the some of the deeper scientific challenges out there."

The Nobel committee also considers the practical impact of research, not just the intellectual breakthrough. The rapid adoption of AlphaFold by over 2.5 million researchers worldwide and its contribution to various scientific fields demonstrated its significant impact on society, leading to the award.

## Drug Discovery and Isomorphic Labs

Building on the success of AlphaFold, DeepMind has spun out a new company called Isomorphic Labs. This company aims to leverage AlphaFold's capabilities to revolutionize the drug discovery process. Hassabis explained that knowing the 3D structure of a protein is just one piece of the puzzle. Isomorphic Labs is developing models for designing drug compounds, ensuring they are non-toxic, soluble, and have the right properties.

> "You need to design the drug compound, the molecule, then you need to make sure it's not toxic, it has all the right properties, it's soluble. So you need actually lots of other models adjacent to it... and put them all together and then you can revolutionize the whole drug discovery process."

This approach has the potential to significantly reduce the time it takes to develop new drugs, potentially accelerating the process by tenfold. Isomorphic Labs is working with pharmaceutical companies like Eli Lilly and Novartis on various drug programs, targeting diseases like cancer, cardiovascular issues, and neurodegenerative conditions. Hassabis revealed that they expect to have their first drugs in clinical trials by the end of the year.


<iframe width="560" height="315" src="https://www.youtube.com/embed/ICv03VysLaE?si=tiuWshunzBhT0qHr" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>

## Google's AI Advancements Beyond AlphaFold

Beyond AlphaFold, Hassabis discussed Google's broader AI advancements, including the Gemini 2.0 model and its various iterations. He highlighted the Gemini 2.0 Flash version, which is designed for scalability and efficiency, as well as state-of-the-art video models capable of accurately simulating physics. These models are part of Google's efforts to build world models that understand not just language but also the spatial and temporal aspects of the world.

> "We had V2 state-of-the-art video model... it still astounds me how accurately can model the physics... it's the first model I think that does that and it's the beginning for us of what we call a world model."

These world models are essential for developing universal assistants, which are digital agents that can help users in their everyday lives. Google's Project Astra is a research prototype in this area, aiming to create agents that can interact with the world and assist users in various tasks.

## Reasoning and Addressing Hallucinations

Hassabis also addressed the issue of reasoning in AI models and the challenge of hallucinations. He explained that Google is developing "thinking models" that can introspect their own answers and use tools like search to verify information. This approach aims to improve the accuracy and reliability of AI systems.

> "Can you take a... model... and now at test time... can you allow it to do more thinking and go back on look at its own introspect its own answer and and maybe use some tools like search or something to ground some of the some of the answer or double check it and then output it."

He emphasized that factuality is a key focus for Google, especially for scientific applications where accuracy is paramount. While some may see hallucinations as a source of creativity, Hassabis believes that it should be an intentional feature rather than an accidental occurrence.

## The Pursuit of AGI and its Implications

Hassabis reiterated his long-term goal of achieving Artificial General Intelligence (AGI), which he defines as a system capable of exhibiting all the cognitive capabilities of humans. He sees AGI as a tool to understand the fundamental nature of reality and solve some of the biggest questions in science and philosophy.

> "I want to understand the fundamental nature of reality and all the biggest questions... and AI for me was my answer to my expression of what I I thought I could contribute to that search for meaning."

He estimates that AGI is about five to ten years away, but acknowledges that it requires further breakthroughs. He also emphasized the need for international collaboration and responsible development of AI, involving all stakeholders, including industry, academia, civil society, and governments.

## Risks and the Need for International Collaboration

Hassabis highlighted two main risks associated with AI: the potential for misuse by bad actors and the risks associated with AGI itself. He stressed the need for robust safeguards and international agreements to mitigate these risks. He also expressed concern about the potential for a "prisoner's dilemma" scenario where nations might prioritize their own interests over global cooperation.

> "There's much more at stake here than just companies or... products... there's the sort of future of humanity... and that's what's at stake."

He advocated for a balanced approach, being both bold in pursuing the opportunities of AI and responsible in addressing its risks. He also emphasized the importance of putting the scientific method at the center of AI development.

## Conclusion

The interview with Demis Hassabis provides a comprehensive overview of DeepMind's groundbreaking work in AI, particularly with AlphaFold. The system's ability to accelerate scientific research and its potential to revolutionize drug discovery are truly transformative. However, Hassabis also acknowledges the risks associated with AI and the need for international collaboration and responsible development. As AI continues to advance, it is crucial to have open discussions and engage all stakeholders to ensure that this technology is used for the benefit of humanity.

DeepMind's AlphaFold: Revolutionizing Science and Medicine with AI

## Introduction

The global race to develop advanced artificial intelligence models capable of complex reasoning is intensifying, with Chinese tech companies ByteDance and DeepSeek emerging as key players. These firms are releasing new AI models that directly challenge the performance and pricing of established leaders like OpenAI. This development marks a significant shift in the AI landscape, potentially disrupting the market share of existing large language models.

## ByteDance's Doubao-1.5-pro

ByteDance, the parent company of TikTok, recently unveiled an upgraded version of its flagship AI model, **Doubao-1.5-pro**. This model is designed to compete with OpenAI's latest reasoning models. ByteDance claims that Doubao-1.5-pro outperforms OpenAI's o1 model on the AIME benchmark, a test that assesses an AI's ability to understand and respond to complex instructions. This release highlights ByteDance's commitment to advancing its AI capabilities and challenging the status quo in the AI market.

## DeepSeek's Open-Source Reasoning Model

Adding to the competitive landscape, Chinese AI startup DeepSeek launched an open-source reasoning model called **DeepSeek-R1**. DeepSeek asserts that its model rivals OpenAI's o1 in several performance benchmarks. This move follows DeepSeek's previous success with its V3 large language model, which demonstrated superior performance compared to models from OpenAI and Meta, despite operating with a smaller budget. DeepSeek's approach of offering high performance at a lower cost is gaining significant attention in the global AI community.

## Market Disruption and Competitive Pricing

The advancements in AI reasoning by ByteDance, DeepSeek, and other Chinese firms are poised to challenge the market dominance of OpenAI and other major players. These companies are not only focusing on performance but also on offering more competitive pricing. DeepSeek, for example, offers access to DeepSeek-R1 at 16 yuan ($2.20) per million tokens, a stark contrast to OpenAI's o1 pricing of 438 yuan for the same usage. ByteDance's pricing is even more aggressive, with Doubao-1.5-pro-32k costing 2 yuan per million tokens for output and the more powerful Doubao-1.5-pro-256k version priced at 9 yuan.

This aggressive pricing strategy could significantly impact the market, making advanced AI capabilities more accessible to a wider range of users and businesses. The lower costs could also encourage more experimentation and innovation in the AI field.

## Other Chinese Firms Joining the Race

ByteDance and DeepSeek are not the only Chinese companies making strides in AI reasoning. Other firms like Moonshot AI, Minimax, and iFlyTek have also unveiled their own reasoning models in recent weeks. This surge of activity indicates a strong push from China to establish itself as a leader in AI development. The competition among these firms is likely to drive further innovation and improvements in AI technology.

## OpenAI's Response

OpenAI, which triggered the current race in AI development with the launch of ChatGPT in November 2022 and its "Strawberry" series of AI reasoning models in September last year, is not standing still. The company's CEO, Sam Altman, recently announced that they have finalized a version of their new reasoning AI model, o3 mini, which is expected to launch soon. This indicates that OpenAI is actively working to maintain its competitive edge in the rapidly evolving AI landscape.

## Conclusion

The recent developments in AI reasoning by ByteDance, DeepSeek, and other Chinese firms signal a significant shift in the global AI market. These companies are not only challenging the performance of established models but also offering more competitive pricing, potentially disrupting the market share of existing leaders. The increased competition is likely to drive further innovation and make advanced AI capabilities more accessible to a wider audience. As the race continues, it will be interesting to see how these developments shape the future of AI technology.

ByteDance and DeepSeek Lead Chinese AI Push in Reasoning Capabilities

## Introduction

Have you ever read a poem and thought, "Wow, that's deep," only to find out it was written by a robot? It might sound like science fiction, but a recent study published in *Nature Scientific Reports* suggests that this is more common than you might think. The study found that non-expert readers of poetry often can't tell the difference between poems written by humans and those generated by AI. Even more surprisingly, they sometimes prefer the AI versions. Let's dive into why this might be the case.

Here's a stanza that was generated by GPT-3.5 in response to the prompt “write a short poem in the style of Sylvia Plath.”:

> The air is thick with tension,
> My mind is a tangled mess,
> The weight of my emotions
> Is heavy on my chest.

It certainly sounds like something Plath might have written, doesn't it? It hits those key notes of despair and internal struggle that we often associate with her work. The near-rhyme of "mess" and "chest" also gives it that familiar poetic feel.

## The Study: AI vs. Human Poets

The study used AI to generate poems "in the style of" ten different poets, including Geoffrey Chaucer, William Shakespeare, Emily Dickinson, and Sylvia Plath. Participants were then presented with a mix of real and AI-generated poems and asked to guess which were written by humans and which by AI. They were also asked to rate their confidence in their guesses.

In another part of the study, participants were given different instructions. Some were told all the poems were human, some were told they were all AI, and some were given no information. They were then asked to rate the poems on a scale from "extremely bad" to "extremely good." The results were quite revealing.

### Key Findings

- **Difficulty in Differentiation:** Participants struggled to distinguish between AI and human poetry. They often guessed incorrectly, even when they were confident in their choices.
- **Preference for AI:** Surprisingly, AI-generated poems often scored higher than human-written ones in attributes like "creativity," "atmosphere," and "emotional quality."
- **The Power of Suggestion:** When participants were told that all the poems were human-written, they tended to rate them higher, suggesting that our expectations can influence our perception of quality.

## Why the Preference for AI?

As an English lecturer, the study's findings aren't entirely surprising. Poetry is often the literary form that students find the most challenging. It's not something most people engage with regularly, despite its presence in our daily lives (think Instagram posts, coffee cups, and greeting cards). We might have studied it in high school, but our reading often doesn't go much beyond that.

The researchers suggest that AI models are capable of producing "high-quality poetry." But what does "high-quality" really mean? In my view, the study highlights the difficulty of engaging with poetry in a meaningful way. It takes time, effort, and repeated readings to experience what literary critic Derek Attridge calls the "event" of literature, where new possibilities of meaning and feeling open up within us. It's about being "pulled along by the work as we push ourselves through it."

Think of it like this: it's the difference between listening to a catchy pop song and immersing yourself in a complex piece of classical music. The pop song is immediately accessible and enjoyable, while the classical piece requires more attention and effort to fully appreciate. Similarly, AI poetry often provides immediate gratification, while human poetry may require more work to understand.

### The Problem of Instant Gratification

In a world where we expect instant answers, we tend to favor poems that are easier to interpret and understand. AI, with its ability to generate formally adequate poems in seconds, caters to this desire. These models are designed to satisfy general taste, giving us the poems we think we want—poems that tell us things directly.

## How Poems Think

The real work of teaching poetry is to help students understand how poems think, poem by poem and poet by poet. It's about gaining access to poetry's specific intelligence. For example, in my introductory course, I spend a significant amount of time unpacking just the first line of Sylvia Plath's "Morning Song": “Love set you going like a fat gold watch.”

We explore questions like: How might a "watch" be connected to "set you going"? How can love set something going? What does a "fat gold watch" mean to you—and how is it different from a slim silver one? Why "set you going" rather than "led to your birth"? And what does all this mean in a poem about having a baby, and all the ambivalent feelings this may produce in a mother?

### A Deeper Look at Plath

Let's compare the AI-generated Plath poem with a real one, "Winter Landscape, With Rooks":

> Water in the millrace, through a sluice of stone,
> plunges headlong into that black pond
> where, absurd and out-of-season, a single swan
> floats chaste as snow, taunting the clouded mind
> which hungers to haul the white reflection down.

Notice how Plath intricately explores the connection between mental events and place. The details convey the tumble of life's events through our minds. Our minds are turned by life just as the mill is turned by water. These experiences accumulate in a scarcely understood "black pond."

But the poet finds that this metaphor, well constructed though it may be, doesn't quite work. The landscape refuses to submit to her emotional atmosphere. Despite everything she feels, a swan floats on serenely—even if she "hungers" to haul its "white reflection down."

This is a far cry from the AI poem, which focuses more on the direct expression of emotion. Plath acknowledges not just the weight of her despair but also the absurd figure she may be within a landscape she wants to reflect her sadness. She compares herself to the bird that gives the poem its title:

> feathered dark in thought, I stalk like a rook,
> brooding as the winter night comes on.

These lines might not score highly on metrics like "beautiful" or "inspiring," but they offer a profound insight. Plath is the source of her torment, "feathered" as she is with her "dark thoughts." She is "brooding," trying to make the world into her imaginative vision.

## The Limitations of Metrics

The authors of the study are both right and wrong when they write that AI can "produce high-quality poetry." The preference for AI poetry doesn't suggest that machine poems are of a higher quality. AI models can produce poems that rate well on certain metrics, but the experience of reading poetry is not about standardized criteria or outcomes. It's about the imaginative tussle between the reader and the poem, where both are transformed in the process.

Ultimately, the study provides a valuable examination of how people who know little about poetry respond to it. But it fails to explore how poetry can be enlivened by meaningful shared encounters. Spending time with poems, attending to their intelligence, and engaging in the acts of sympathy and speculation required to confront their challenges is as difficult as ever. As the AI-generated Plath puts it:

> My mind is a tangled mess,
> [...]
> I try to grasp at something solid.

Perhaps, in our quest for instant gratification, we're missing out on the deeper, more rewarding experience that human poetry offers. It's a reminder that sometimes, the most valuable things in life require a little more effort and patience.

AI vs. Human Poetry: Why We Can't Tell the Difference (and Sometimes Prefer the AI)

## The Dawn of World Models: A New Era for Robotics

If you've been following the latest advancements in AI, you've probably heard about large language models (LLMs) like ChatGPT. They've transformed how we interact with text-based AI. But what if I told you that a similar revolution is brewing in the world of robotics? Enter **world models**, a concept that's rapidly gaining traction and could be the key to unlocking the next generation of intelligent machines.

Think of it this way: just as LLMs like ChatGPT need vast amounts of text data to learn, robots need a way to understand and interact with the physical world. That's where world models come in. They're essentially AI tools that generate realistic 3D environments, allowing robots to train and learn in virtual simulations before they ever step foot into the real world.

### What Exactly Are World Models?

To understand world models, let's take a step back. Traditionally, robots were programmed with explicit instructions for every task. But this approach is limited and doesn't allow for adaptability. Now, engineers are turning to AI and 3D simulations to let robots learn on their own. This is where world models become crucial.

World models are a type of **generative AI** that can produce 3D environments and simulate virtual worlds. They're like the virtual playgrounds where robots can learn to perceive, understand, and navigate three-dimensional space. Imagine a robot learning to pick up a delicate object in a virtual kitchen, then applying that knowledge to a real-world scenario. That's the power of world models.

Think of it like this: if ChatGPT is the interface for LLMs, then world models are the interface for virtual world simulators. They could allow anyone, even those without technical game development skills, to create 3D virtual worlds. This could democratize the creation of training data for robots, leading to faster development and more robust AI.

### The Rise of World Model Technology

The buzz around world models is growing rapidly. At the start of 2025, Nvidia announced its new platform, **Cosmos**, a generative AI tool for creating virtual-world-like videos. Google's DeepMind is also working on similar projects, and a startup called World Labs achieved unicorn status in just four months, all focused on building these virtual worlds. This flurry of activity signals that world models are not just a concept, but a rapidly developing technology.

While many early world models generate spatial training data in video format, some are already capable of producing fully immersive scenes. For example, a startup called Odyssey uses gaussian splatting to create scenes that can be loaded into 3D software like Unreal Engine and Blender. Another startup, Decart, has even demoed a playable version of a game similar to Minecraft using their world model. DeepMind has also explored the video game route, showcasing the potential for interactive and engaging virtual training environments.

This shift in how computer graphics work is significant. Nvidia CEO Jensen Huang has even predicted that in the future, “every single pixel will be generated, not rendered but generated.” While traditional rendering systems aren't likely to disappear completely, generative AI is poised to play a much larger role in creating the visuals we see in games and simulations.

### The Implications for Robotics

The implications of world models for robotics are potentially huge. Nvidia is pushing the term “physical AI” to describe the intelligent systems that will power everything from warehouse robots to autonomous vehicles. To operate effectively in the real world, these systems need to train in physically accurate simulations. World models can provide these simulations, generating an endless variety of training scenarios.

World Labs, founded by AI pioneer Fei-Fei Li, is a prime example of this shift. They define themselves as a spatial intelligence company, believing that true general intelligence requires an embodied ability to reason about objects, places, and interactions in 3D space and time. Their goal is to build foundation models that can move AI into three-dimensional space, allowing robots to understand and interact with the world more like humans do.

Imagine a robot learning that if you squeeze an egg too hard, it will crack. But context matters. If the goal is to place it in a carton, the robot learns to be gentle. If the goal is to make an omelet, it can squeeze away. This kind of nuanced understanding is what world models can help robots achieve.

### Challenges and the Road Ahead

While the potential of world models is exciting, there are still challenges to overcome. Training and running these models requires massive amounts of computing power, even compared to today's AI. Additionally, the models aren't always consistent with the real world's rules, and like all generative AI, they can be influenced by biases in their training data.

As TechCrunch's Kyle Wiggers pointed out, a world model trained on videos of sunny European cities might struggle to understand or depict Korean cities in snowy conditions. This means that traditional simulation tools like game and physics engines will still be used for some time to render training scenarios for robots. And Meta's head of AI, Yann LeCun, believes that advanced world models, like the ones in our heads, will take longer to develop.

### A New Era of Spatial Intelligence

Despite these challenges, the development of world models is a significant step forward for robotics. Just as ChatGPT signaled a turning point for AI, robots, drones, and embodied AI systems may be nearing a similar breakout moment. To get there, physically accurate 3D environments will be essential for these systems to learn and mature.

Early world models may make it easier than ever for developers to generate the countless training scenarios needed to bring on an era of spatially intelligent machines. This could lead to robots that are not only more capable but also more adaptable and intuitive in their interactions with the world. It's an exciting time to be in the field of robotics, and the future looks brighter than ever thanks to the potential of world models.

In short, world models are poised to be the next big thing in AI, and their impact on robotics could be transformative. Get ready for a future where robots are not just programmed, but truly understand and interact with the world around them.

Robotics' ChatGPT Moment: How AI World Models Could Revolutionize the Field

## Introduction

Have you ever been stuck on a tricky problem, wishing you had a bit more time to think it through? Well, that's exactly what Large Language Models (LLMs) are starting to do! Models like OpenAI's o1 have shown that when LLMs are given more computational power during inference (the thinking process), they get much better at reasoning tasks like math, coding, and logic. It's like giving your brain a little extra time to work things out.

However, the exact recipe behind these powerful reasoning models has been a closely guarded secret. That is, until recently, when DeepSeek released their DeepSeek-R1 model. This release not only matched or exceeded the performance of models like o1, but also came with a detailed tech report outlining their training process. This was a big deal, and it got everyone in the AI community very excited!

DeepSeek's approach involved using pure reinforcement learning to teach a base language model how to reason without any human supervision. Think of it like teaching a dog a new trick, but instead of treats, you're giving the model feedback based on how well it solves problems. The figure below shows that creating a powerful reasoning model is now quite simple if you have a good base model and high-quality data:

[Diagram of Base Model + Data Mixture -> Reasoning Model]

Despite this exciting release, some key questions remain:

*   **Data Collection:** How were the specific datasets for reasoning created?
*   **Model Training:** DeepSeek didn't release the training code, so we don't know the best settings and how they might change for different models.
*   **Scaling Laws:** What are the trade-offs between compute power and data when training these models?

These questions led to the creation of the **Open-R1 project**. Our goal is to reconstruct DeepSeek-R1's data and training pipeline, validate their claims, and push the boundaries of open reasoning models. By doing this, we hope to make the process transparent, share insights with the open-source community, and create a foundation for future models to build upon. 

In this blog post, we'll explore the key ingredients behind DeepSeek-R1, what we plan to replicate, and how you can contribute to the Open-R1 project. Let's dive in!

## How Did DeepSeek Do It?

DeepSeek-R1 is built on top of DeepSeek-V3, a powerful base model. Think of DeepSeek-V3 as the foundation of a house, and DeepSeek-R1 as the finished building. DeepSeek-V3 is a 671B Mixture of Experts (MoE) model, which means it's like having a team of specialists working together. It performs as well as other top models like Sonnet 3.5 and GPT-4o. What's even more impressive is that it was trained for only $5.5M, thanks to clever architectural changes like Multi Token Prediction (MTP), Multi-Head Latent Attention (MLA), and a lot of hardware optimization. It's like building a high-performance car but making it fuel-efficient at the same time!

DeepSeek actually released two models: DeepSeek-R1-Zero and DeepSeek-R1. They each have a different training approach. DeepSeek-R1-Zero skipped the traditional supervised fine-tuning and relied entirely on reinforcement learning (RL). They used Group Relative Policy Optimization (GRPO) to make the process more efficient. The model was guided by a simple reward system that gave feedback based on the accuracy and structure of its answers. This helped the model develop reasoning skills, like breaking down problems and checking its own work. However, the responses from R1-Zero were often unclear and hard to read.

That's where DeepSeek-R1 comes in. It started with a "cold start" phase, fine-tuning on a small set of carefully chosen examples to improve clarity and readability. From there, it went through more RL and refinement steps, including rejecting low-quality outputs using both human preferences and verifiable rewards. This resulted in a model that not only reasons well but also produces clear and consistent answers. It's like taking a rough draft and polishing it into a final, well-written piece.

## Open-R1: Filling in the Missing Pieces

The release of DeepSeek-R1 is a huge step forward for the AI community. However, not everything was released. While the model weights are open, the datasets and code used to train the model are not. It's like getting a recipe for a delicious cake but not having the list of ingredients or the instructions on how to bake it.

The goal of Open-R1 is to create these missing pieces so that the entire research and industry community can build similar or even better models. By doing this in the open, everyone can contribute! It's like a community cookbook where everyone can share their recipes and improve on them.

Here's our plan of attack:

1.  **Replicate the R1-Distill models:** We'll create a high-quality reasoning dataset by distilling knowledge from DeepSeek-R1. Think of it like extracting the essence of the model's reasoning abilities.
2.  **Replicate the pure RL pipeline:** We'll recreate the reinforcement learning process that DeepSeek used to create R1-Zero. This will involve creating new, large-scale datasets for math, reasoning, and code. It's like building the training ground for the model.
3.  **Show we can go from base model → SFT → RL:** We'll demonstrate how to go from a base model to a reasoning model through multi-stage training. It's like showing the complete journey of how to build the model.

The synthetic datasets will allow anyone to fine-tune existing or new LLMs into reasoning models. The training recipes involving RL will serve as a starting point for building similar models from scratch and will allow researchers to build even more advanced methods. It's like providing a toolkit for anyone to build their own reasoning models.

We're not just stopping at math datasets. There's a lot of potential in exploring other areas, like code and scientific fields such as medicine, where reasoning models could have a significant impact. Imagine a model that can help doctors diagnose diseases or help scientists discover new medicines!

This initiative isn't just about replicating results—it's about sharing insights with the community. By documenting what works, what doesn't, and why, we hope to save others from wasting time and resources on unproductive paths. It's like sharing your experiences so that others can learn from your mistakes and successes.

Open-R1: Rebuilding DeepSeek-R1 for the Open Source Community

## Introduction

The Nvidia GeForce RTX 5080 is here, and it's... interesting. It's not the earth-shattering leap we've seen in some past generations, but it's also not a complete dud. Think of it like a slightly faster, slightly more efficient version of the RTX 4080 Super, but with a few new tricks up its sleeve. Priced at $999 for the Founders Edition, it's definitely more appealing than the launch price of its predecessor. If you're rocking an older card and have been waiting for the right time to upgrade, this might be it. But if you're already on a 4080 or 4080 Super, the upgrade might not be worth it.

This generation, powered by the Blackwell architecture, feels like a more subtle evolution rather than a revolution. The performance gains aren't massive, and even the mighty RTX 5090 doesn't leave the 4090 in the dust. Nvidia is banking on DLSS 4 and its multi-frame generation to make the difference, and for those with high-refresh, high-resolution displays, it might just do the trick. But be prepared to have some of your frames generated by AI – it's the future, whether you like it or not.

## Specs and Features

Like its bigger sibling, the RTX 5090, the RTX 5080 is built on the Blackwell architecture, which is designed with AI workloads in mind. It packs more CUDA cores than the 4080, with 10,752 spread across seven Graphics Processing Clusters (GPCs), compared to the 9,728 cores in the 4080. Nvidia has also managed to cram more Streaming Multiprocessors (SMs) into each GPC, allowing for more CUDA cores without a massive increase in GPU size. It's like fitting more people into a stadium without making the stadium bigger.

Each SM features four Tensor Cores and one RT Core, totaling 336 and 84, respectively. The Tensor Cores have been a major focus for Nvidia, and these 5th-generation cores are supposedly much more efficient, which is crucial for DLSS 4. Think of it like upgrading from a regular bicycle to an e-bike – more power with less effort.

Another key feature is the new AI Manager Processor (AMP), built directly into the GPU. This AMP is like a traffic controller for your graphics card, directing work to different parts of the GPU more efficiently than a traditional CPU. It's like having a dedicated lane for high-priority traffic, ensuring everything runs smoothly.

The RTX 5080 also comes with 16GB of GDDR7 VRAM on a 256-bit bus. While 16GB is enough for most people, especially at 4K, games are becoming more demanding. With the RTX 5090 boasting 32GB, it would have been nice to see the 5080 get the 24GB found in the RTX 4090, especially since it's being marketed as a 4K card.

Surprisingly, the power demand hasn't increased dramatically, requiring 360W compared to the 320W of the RTX 4080. It's a modest 13% bump, especially compared to the 575W of the RTX 5090. This means the Founders Edition cooler doesn't have to work as hard. Despite this, the RTX 5080 looks remarkably similar to the RTX 5090, with the same gunmetal gray color and flow-through heatsinks. In testing, the RTX 5080 peaked at 77°C, compared to 87°C for the RTX 5090. It's also a dual-slot card, unlike the triple-slot design of the RTX 4080 Super, and it still manages to cool better.

## DLSS 4

The real star of the show for the RTX 5080 is DLSS 4, specifically its multi-frame generation (MFG) feature. Many of the improvements to DLSS, like the new Transformer model, will be available on older cards. This new model should improve image quality and reduce ghosting and artifacting, while also reducing the VRAM needed for DLSS. It's like getting a free upgrade to your favorite photo editing software – better results with less effort.

MFG is where things get interesting. It analyzes a rendered frame and motion vector data to generate entirely new 'AI' frames between each rendered frame. The RTX 5080 can generate up to 3 frames from each rendered frame in supported games. It's like having a magician create extra frames out of thin air, making your game look smoother.

However, this can introduce latency, but Nvidia Reflex 2 is here to help. It zeroes out the render queue provided by your CPU, making it provide frame data just as the frame is about to be rendered. It also uses your last mouse position and data from previous frames to predict where you'll be moving. Combined with hardware flip metering from the AMP, it aims to provide maximum smoothness. It's like having a personal assistant who anticipates your every move.

To get the best experience with MFG, you'll want to enable it when you're already getting a decent frame rate, ideally 60 fps or 40 fps if you're feeling brave. In Cyberpunk 2077 at 4K with Ray Tracing Overdrive and DLSS set to performance, the RTX 5080 delivered 59 fps. With 2x frame gen, that jumped to 104 fps, and with 4x frame gen, it hit 183 fps. Similarly, in Star Wars Outlaws, frame rates tripled with MFG enabled. It's like turning up the volume on your game's smoothness.

However, it's important to remember that the underlying frame rate remains around 60 fps, so you won't see a huge increase in responsiveness. The game will just look smoother. It's like adding a layer of polish to your game, making it look more refined.

## Performance

While the RTX 5080 doesn't have the same kind of performance leap as some previous generations, it still offers the best performance in its price range. It's a legitimate 4K contender for those who don't want to spend a fortune on the 5090. At 4K, the RTX 5080 is, on average, 11% faster than the RTX 4080 and 8% faster than the RTX 4080 Super. This performance uplift drops at lower resolutions. In synthetic benchmarks like 3DMark, the performance difference grows, suggesting that the card will get better over time with new drivers and more demanding games. It's like a fine wine that gets better with age.

### Test System

*   **CPU:** AMD Ryzen 7 9800X3D
*   **Motherboard:** Asus ROG Crosshair X870E Hero
*   **RAM:** 32GB G.Skill Trident Z5 Neo @ 6,000MHz
*   **SSD:** 4TB Samsung 990 Pro
*   **CPU Cooler:** Asus ROG Ryujin III 360

For this review, the same comparison data from the RTX 5090 review was used, along with testing the RTX 4080 and RTX 3080 on Nvidia Game Ready Driver 566.36. The AMD Radeon RX 7900 XTX and 7900 XT were tested on AMD Adrenalin 24.12.1. DLSS 4 and frame generation were not used in these tests.

In 3DMark's Speed Way benchmark, the RTX 5080 scored 8,966 points, compared to 7,519 for the RTX 4080 Super and 7,258 for the RTX 4080. That's a 23% difference over the RTX 4080. In Port Royal, which focuses on ray tracing, the RTX 5080 also enjoyed a 24% lead over the 4080, but the lead over the 4080 Super was only 19%. It's like a race where the 5080 is ahead, but the 4080 Super is catching up.

In Call of Duty: Black Ops 6, the RTX 5080 hit 140 fps at 4K with the Extreme preset and DLSS set to 'Performance'. While this is enough for most high-refresh monitors, it's not a huge leap over the 129 fps of the RTX 4080 Super and 124 fps of the 4080. The Radeon RX 7900 XTX actually performed better at 145 fps with FSR set to 'Performance'. It's like a close race where the 5080 is in the lead, but not by much.

Cyberpunk 2077 is one of the most demanding games, and at 4K with Ray Tracing Ultra and DLSS set to 'performance', the RTX 5080 gets 93 fps, compared to 87 from the 4080 Super and 84 from the RTX 4080. That's just an 11% gen-on-gen improvement. It's like a slow and steady climb, not a sudden jump.

In Metro Exodus: Enhanced Edition, the RTX 5080 almost hit the golden standard, netting 58 fps, compared to 53 fps from the RTX 4080. That's a 10% generational lead. In Red Dead Redemption 2, the RTX 5080 gets 127 fps at 4K with max settings, compared to 122 fps for the RTX 4080 Super and 120 fps for the 4080. It's like a game where everyone is performing well, but the 5080 is just a bit ahead.

Total War: Warhammer 3 showed one of the 5080's strongest showings, getting 92 fps to the RTX 4080's 76. That's a massive 21% jump in performance. It's like a game where the 5080 is clearly the winner. However, in Assassin's Creed Mirage, the RTX 5080 was only 4% faster than the 4080, likely due to driver issues. It's like a game where the 5080 is having a bad day.

Black Myth Wukong showed similar scaling to Cyberpunk 2077, with the RTX 5080 getting 74 fps at 4K with the Cinematic preset, compared to 68 from the RTX 4080, making for a 9% difference. Finally, in Forza Horizon 5, the RTX 5080 was surprisingly faster, getting 171 fps at 4K with the Extreme preset, compared to 146 fps from the RTX 4080. That's a 17% performance increase. It's like a game where the 5080 is showing its true potential.

## Conclusion

The Nvidia GeForce RTX 5080 isn't the most exciting graphics card, especially compared to previous generations. However, it's still a solid performer, especially at 4K. If you've been waiting to upgrade, now might be the time, especially if you can find it at its suggested $999 price tag. It's not a revolution, but it's a worthwhile evolution.

Nvidia GeForce RTX 5080 Review: Is It Worth the Hype?

## Introduction

Boom Supersonic's XB-1 prototype has achieved a major milestone by breaking the sound barrier for the first time. This successful test flight, the 12th for the aircraft, saw the XB-1 reach a top speed of Mach 1.1, marking a significant step forward in the company's ambition to bring supersonic commercial flight back to the skies. The XB-1 serves as a smaller-scale demonstrator for the larger Overture airliner that Boom plans to develop.

## The XB-1's Supersonic Flight

On its 12th test flight, the XB-1 took off from the Mojave Air & Space Port in California, piloted by Boom's chief test pilot, Tristan Brandenburg. The aircraft ascended to an altitude of 34,000 feet before achieving a speed of Mach 1.1, which is approximately 844 mph. This supersonic speed was maintained for about four minutes. Notably, the XB-1 broke the sound barrier not just once, but three times during the flight, including twice during its descent. The flight took place within the Bell X-1 Supersonic Corridor, named after the first aircraft to break the sound barrier.

### Significance of the Flight

This flight is significant for several reasons. Firstly, it's the first time Boom's demonstrator aircraft has achieved supersonic speed. Secondly, it marks the first time a privately developed civil aircraft has broken the sound barrier. The Concorde, the only previous commercial supersonic airliner, was a joint venture between the British and French governments. This achievement underscores Boom's progress in the development of its supersonic technology.

## XB-1 Specifications and Development

The XB-1 is a 63-foot-long aircraft powered by three General Electric J85-15 turbojet engines. While the XB-1 utilizes these engines, Boom plans to develop its own four engines to power the larger Overture airliner. This decision came after the company's partnership with Rolls-Royce ended in 2022. The XB-1's initial test flight in March 2024 saw it reach a top speed of 283 mph, and a subsequent test flight in January 2025 achieved Mach 0.95. The recent supersonic flight represents a substantial leap in performance.

### The Overture Airliner

The XB-1 is a crucial step in the development of the Overture, Boom's planned commercial airliner. The Overture is designed to carry 64 passengers at cruising speeds of up to Mach 1.7. This would allow for significantly faster international flights, similar to the service offered by the Concorde. Boom aims for the Overture to take its first flight by 2030.

## Future Prospects and Challenges

While the successful supersonic flight of the XB-1 is a major milestone, Boom still faces significant development and testing before the Overture becomes a reality. The company has raised over $700 million in funding and has secured orders for 15 aircraft from United Airlines, with an option for 35 more, and 20 from American Airlines. These orders demonstrate confidence in Boom's vision, but the path to commercial supersonic flight is complex and requires overcoming numerous technical and regulatory hurdles.

## Conclusion

The successful supersonic flight of the Boom XB-1 is a landmark achievement in the pursuit of commercial supersonic travel. This test flight not only demonstrates the viability of Boom's technology but also marks a significant step towards the return of faster-than-sound passenger flights. While challenges remain, the progress made by Boom Supersonic is a promising sign for the future of air travel.

Boom XB-1 Supersonic Test Plane Achieves Milestone, Breaks Sound Barrier

## Introduction

Want to run cutting-edge AI models locally without relying on cloud services? This guide will walk you through building a high-performance machine specifically designed to run the DeepSeek-R1 model at full quality using Q8 quantization. We'll focus on maximizing memory bandwidth, which is crucial for large language model (LLM) performance. This setup avoids the need for expensive GPUs, keeping the total cost under $6,000.

Let's dive into the hardware and software you'll need.

## Hardware Setup

### 1. Motherboard: Gigabyte MZ73-LM0 or MZ73-LM1

- **Why?** We need a dual-socket motherboard to support two CPUs and, more importantly, a massive 24 channels of DDR5 RAM. This is essential for the memory bandwidth required by DeepSeek-R1.
- **Link:** [Gigabyte MZ73-LM0](https://www.gigabyte.com/Enterprise/Server-Motherboard/MZ73-LM0-rev-3x)

### 2. CPUs: 2x AMD EPYC 9004 or 9005 Series

- **Why?** LLM generation is often bottlenecked by memory bandwidth, not raw CPU power. Therefore, you don't need the most expensive CPUs. The EPYC 9004 or 9005 series provides excellent performance at a reasonable cost.
- **Recommendation:** Consider the 9115 or even the 9015 to save on costs.
- **Link:** [Newegg AMD EPYC CPUs](https://www.newegg.com/p/N82E16819113865)

### 3. RAM: 768GB DDR5-RDIMM

- **Why?** The DeepSeek-R1 model requires a substantial amount of RAM. We'll use 768GB across 24 channels to maximize bandwidth. This means 24 x 32GB DDR5-RDIMM modules.
- **Link:** [Nemix RAM 384GB Kit](https://www.newegg.com/nemix-ram-384gb/p/1X5-003Z-01FM7) (You'll need two of these kits)

### 4. Case: Phanteks Enthoo Pro 2 Server Edition

- **Why?** Most consumer cases won't fit a full server motherboard. The Enthoo Pro 2 Server Edition is designed to accommodate large server boards.
- **Link:** [Phanteks Enthoo Pro 2 Server](https://www.newegg.com/black-phanteks-enthoo-pro-2-server-edition-full-tower/p/N82E16811854127)

### 5. PSU: Corsair HX1000i

- **Why?** While the system's power consumption is relatively low (<400W), you'll need a PSU with enough CPU power cables for two EPYC CPUs. The Corsair HX1000i is a reliable option.
- **Link:** [Corsair HX1000i](https://www.corsair.com/us/en/p/psu/cp-9020259-na/hx1000i-fully-modular-ultra-low-noise-platinum-atx-1000-watt-pc-power-supply-cp-9020259-na)

### 6. Heatsinks: SP5 Socket Compatible

- **Why?** AMD EPYC uses the SP5 socket, and most heatsinks are designed for server blades. You'll likely need to source these from eBay or AliExpress.
- **Recommendation:** The heatsink linked below has been tested and works well.
- **Link:** [Ebay SP5 Heatsink](https://www.ebay.com/itm/226499280220)

### 7. SSD: 1TB or Larger NVMe

- **Why?** You'll need an SSD to store the model and operating system. NVMe is recommended for faster read/write speeds, especially when loading the 700GB model into RAM.
- **Recommendation:** Any 1TB or larger NVMe SSD will work.

## Hardware Assembly

1.  **Install the CPUs:** Carefully install the two AMD EPYC CPUs into their respective sockets on the motherboard.
2.  **Install the RAM:** Populate all 24 RAM slots with the 32GB DDR5-RDIMM modules.
3.  **Mount the Motherboard:** Install the motherboard into the Phanteks Enthoo Pro 2 case.
4.  **Install the Heatsinks:** Attach the SP5-compatible heatsinks to the CPUs.
5.  **Install the SSD:** Connect the NVMe SSD to the motherboard.
6.  **Connect the PSU:** Install the Corsair HX1000i PSU and connect all necessary power cables to the motherboard and components.

## BIOS Configuration

- **Important:** Before installing the operating system, enter the BIOS settings and set the number of NUMA groups to **0**. This ensures that the model's layers are interleaved across all RAM chips, maximizing throughput. This step is crucial for optimal performance.

## Software Setup

### 1. Install Linux

- Install your preferred Linux distribution on the SSD. Ubuntu is a popular choice for development and AI tasks.

### 2. Install llama.cpp

- **Why?** `llama.cpp` is a C++ implementation of the LLaMA model, optimized for CPU inference. It's essential for running DeepSeek-R1 on our system.
- **Instructions:** Follow the installation instructions on the official GitHub repository: [llama.cpp GitHub](https://github.com/ggerganov/llama.cpp)

### 3. Download the DeepSeek-R1 Model

- **Why?** We need the quantized version of the DeepSeek-R1 model in GGUF format. We'll use the Q8_0 quantization for full quality.
- **Instructions:** Download all files from the Q8_0 folder from the following Hugging Face repository: [DeepSeek-R1 GGUF](https://huggingface.co/unsloth/DeepSeek-R1-GGUF/tree/main)

## Running DeepSeek-R1

### 1. Basic Inference

- For a quick test, use the `llama-cli` command:

```bash
llama-cli -m ./DeepSeek-R1.Q8_0-00001-of-00015.gguf --temp 0.6 -no-cnv -c 16384 -p "<｜User｜>How many Rs are there in strawberry?<｜Assistant｜>"
```

- **Explanation:**
    - `-m`: Specifies the path to the model file.
    - `--temp`: Sets the temperature for sampling (0.6 is a good starting point).
    - `-no-cnv`: Disables conversational mode.
    - `-c`: Sets the context size (16384 is recommended).
    - `-p`: Provides the prompt.

### 2. Hosting the Model

- To integrate DeepSeek-R1 with other software, use `llama-server`:

```bash
./server -m ./DeepSeek-R1.Q8_0-00001-of-00015.gguf
```

- This will start a server that you can send requests to.

## Performance

- Expect generation speeds of 6 to 8 tokens per second, depending on your specific CPU and RAM configuration. Longer chat histories may slightly reduce this speed.

## Why No GPU?

- This build intentionally avoids GPUs. While GPUs can offer faster generation speeds, they come with significant drawbacks:
    - **Cost:** High-end GPUs with enough memory to run the full Q8 quantized model are extremely expensive (>$100k).
    - **Quality:** Quantizing models to fit on smaller GPUs often results in a loss of quality.

## Conclusion

Congratulations! You've successfully built a powerful local machine capable of running the DeepSeek-R1 model at full quality. This setup provides a cost-effective way to access frontier-level AI capabilities without relying on cloud services. You now have a fully open-source and free-to-use LLM at your fingertips. Enjoy exploring the possibilities!

Build a Local DeepSeek-R1 Powerhouse: A Step-by-Step Guide

  <section>
    <h2>Hardware Setup</h2>
    <p>This build focuses on maximizing memory bandwidth, crucial for running large language models like Deepseek-R1. We'll be using two AMD EPYC CPUs and a massive amount of DDR5 RAM.</p>
    
    <h3>Motherboard</h3>
    <p>We need a dual-socket motherboard to accommodate two EPYC CPUs and get access to 24 RAM channels. The <strong>Gigabyte MZ73-LM0 or MZ73-LM1</strong> is ideal. <a href="https://www.gigabyte.com/Enterprise/Server-Motherboard/MZ73-LM0-rev-3x">Gigabyte MZ73-LM0</a></p>

    <h3>CPU</h3>
    <p>Choose any <strong>AMD EPYC 9004 or 9005 CPU</strong>. Memory bandwidth is the bottleneck, so you don't need top-of-the-line CPUs. The 9115 or even the 9015 are suitable. <a href="https://www.newegg.com/p/N82E16819113865">Newegg EPYC CPUs</a></p>

    <h3>RAM</h3>
    <p>This is critical. We need <strong>768GB of DDR5-RDIMM</strong> across 24 channels. That means 24 x 32GB modules. <a href="https://www.newegg.com/nemix-ram-384gb/p/1X5-003Z-01FM7">Example RAM Kit</a></p>

    <h3>Case</h3>
    <p>A standard tower case might not be suitable. The <strong>Enthoo Pro 2 Server</strong> is a good choice because it has screw mounts for a full server motherboard. <a href="https://www.newegg.com/black-phanteks-enthoo-pro-2-server-edition-full-tower/p/N82E16811854127">Enthoo Pro 2 Server</a></p>

    <h3>PSU</h3>
    <p>The system's power usage is relatively low (under 400W), but you'll need a PSU with enough CPU power cables. The <strong>Corsair HX1000i</strong> works well. <a href="https://www.corsair.com/us/en/p/psu/cp-9020259-na/hx1000i-fully-modular-ultra-low-noise-platinum-atx-1000-watt-pc-power-supply-cp-9020259-na">Corsair HX1000i</a></p>

    <h3>Heatsink</h3>
    <p>AMD EPYC uses socket SP5, and most SP5 heatsinks are for 2U/4U servers. You may need to source it from eBay or Aliexpress. <a href="https://www.ebay.com/itm/226499280220">Example Heatsink</a></p>

    <h3>SSD</h3>
    <p>Any 1TB or larger SSD will work. An NVMe drive is recommended for faster loading. Find one yourself!</p>

    <p><strong>Important BIOS Tip:</strong> Set the number of NUMA groups to 0 in the BIOS to ensure the model is interleaved across all RAM chips, doubling throughput.</p>
  </section>

  <section>
    <h2>Software Setup</h2>
    <p>Now, let's get the software running.</p>

    <h3>llama.cpp</h3>
    <p>Install <code>llama.cpp</code> following the instructions on their GitHub repository: <a href="https://github.com/ggerganov/llama.cpp">llama.cpp GitHub</a></p>

    <h3>Deepseek-R1 Model</h3>
    <p>Download all files in the <code>Q8_0</code> folder from the following Hugging Face link: <a href="https://huggingface.co/unsloth/DeepSeek-R1-GGUF/tree/main">DeepSeek-R1 GGUF</a>. This will be around 700GB of data.</p>

    <h3>Running the Model</h3>
    <p>For a quick demo, use this command:</p>
    <pre><code>llama-cli -m ./DeepSeek-R1.Q8_0-00001-of-00015.gguf --temp 0.6 -no-cnv -c 16384 -p "&lt;｜User｜&gt;How many Rs are there in strawberry?&lt;｜Assistant｜&gt;"</code></pre>

    <p>Once tested, use <code>llama-server</code> to host the model and pass requests from other software.</p>
  </section>

  <section>
    <h2>Performance and Notes</h2>
    <p>This build achieves a generation speed of 6-8 tokens per second. Note that there's no GPU in this build. While GPUs offer faster generation, they often require significant quantization that can reduce quality, or very expensive cards with >700GB of VRAM for Q8. This CPU-focused build prioritizes quality by using Q8 quantization and large RAM capacity.</p>
  </section>

Building a Local Deepseek-R1 Powerhouse: A $6,000 Guide

Introducing Gemini 2.0: our new AI model for the agentic era
By Demis Hassabis, CEO of Google DeepMind and Koray Kavukcuoglu, CTO of Google DeepMind on behalf of the Gemini team

Over the past year, we have continued to make incredible progress in artificial intelligence. Today, we are releasing the first model in the Gemini 2.0 family of models: an experimental version of Gemini 2.0 Flash. It’s our workhorse model with low latency and enhanced performance at the cutting edge of our technology, at scale.

We are also sharing the frontiers of our agentic research by showcasing prototypes enabled by Gemini 2.0’s native multimodal capabilities.

Gemini 2.0 Flash
Gemini 2.0 Flash builds on the success of 1.5 Flash, our most popular model yet for developers, with enhanced performance at similarly fast response times. Notably, 2.0 Flash even outperforms 1.5 Pro on key benchmarks, at twice the speed. 2.0 Flash also comes with new capabilities. In addition to supporting multimodal inputs like images, video and audio, 2.0 Flash now supports multimodal output like natively generated images mixed with text and steerable text-to-speech (TTS) multilingual audio. It can also natively call tools like Google Search, code execution as well as third-party user-defined functions.

Google Gemini 2.0

President Trump on Tuesday announced a joint venture between OpenAI, SoftBank and Oracle to create at least $100 billion in computing infrastructure to power artificial intelligence.
<br><br>
The venture, called Stargate, adds to tech companies’ significant investments in U.S. data centers, huge buildings full of servers that provide computing power. Stargate could eventually invest as much as $500 billion over four years. The three companies plan to contribute funds to the venture, which will be open to other investors and start with 10 data centers already under construction in Texas.
<br><br>
The announcement of the joint venture was an early trophy for Mr. Trump, even though the effort to form the venture predates his taking office on Monday. Mr. Trump has promised to accelerate the production of American-made A.I. to compete against China for global leadership in the technology, and on Monday he rolled back an executive order from former President Joseph R. Biden Jr. that imposed standards on safety and other requirements for government use of A.I.
<br><br>
During a news briefing Tuesday, Mr. Trump said he would remove barriers to allow for the creation of more data centers. He said he would make “emergency declarations” to allow Stargate to generate its own electricity, without providing details.
<br><br>
“This monumental undertaking is a resounding declaration of confidence in America’s potential,” Mr. Trump said from the White House’s Roosevelt Room.
<br><br>
“We’re going to make it as easy as it can be,” he added.
<br><br>
OpenAI’s chief executive, Sam Altman; SoftBank’s chief, Masayoshi Son; and Oracle’s founder, Larry Ellison, were at the White House announcement with Mr. Trump.
<br><br>
Mr. Trump separately addressed questions from reporters about the potential sale of TikTok. On Monday he signed an executive order that aims to stall enforcement of a law that forces a sale or bans the popular app in the United States over national security concerns.
<br><br>
When asked if he would be in favor of  Elon Musk buying TikTok, Mr. Trump said, “I would be, yeah.” Then he turned to Mr. Ellison and said Oracle could also purchase TikTok.
<br><br>
“I have the right to make a deal,” Mr. Trump added.
<br><br>
Tech executives have rushed to show their support for Mr. Trump in recent months in hopes of currying favor and shielding themselves from regulatory retaliation. They have donated millions to his campaign and inauguration. Mr. Altman attended the inauguration and the chiefs of Google, Apple and Meta occupied prime seats in the Capitol Rotunda.

intermac.dev © 2025 All rights reserved | Build with ❤ by intermac & claude