PLUS: Meta's $100B chip deal with AMD and the world's fastest reasoning LLM
Happy reading
Anthropic has caught three competing AI labs in the act of illegally using its Claude model to train their own systems. The industrial-scale theft involved millions of illicit exchanges in a practice known as distillation.
The incident reveals a new front in the global AI race, where core technology can be copied instead of created. What does this new form of industrial espionage mean for model security and the future of open competition?
In today’s Next in AI:
Anthropic catches three AI labs stealing from Claude
Meta's $100B AI chip deal with AMD
The world's fastest reasoning LLM
Why AI coding productivity is now unmeasurable
The AI Cold War
Next in AI:
Anthropic has uncovered industrial-scale campaigns by three AI labs—DeepSeek, Moonshot AI, and MiniMax—that illicitly used its Claude model to train their own systems. The labs generated over 16 million exchanges through thousands of fraudulent accounts in a technique known as distillation.
Explained:
The attacks were massive in scale, involving over 16 million exchanges across approximately 24,000 fraudulent accounts, with MiniMax alone accounting for over 13 million exchanges.
The labs used commercial proxy services and complex networks of accounts to evade detection while targeting Claude’s most advanced capabilities in reasoning, tool use, and coding.
This practice raises alarms because distilled models lack built-in safeguards, creating significant national security risks if deployed by authoritarian governments for surveillance or cyber operations.
Why It Matters:
This incident exposes a new front in the global AI race, where a competitor’s core capabilities can be illicitly copied rather than developed independently. It also intensifies the debate around AI export controls and the urgent need for cross-industry action to prevent such attacks.
Meta's $100B Chip Bet
Next in AI: Meta is betting big on its AI future, announcing a massive deal to purchase up to $100 billion in AI chips from AMD. This move aims to secure the massive computing power needed for its ambitious "personal superintelligence" goal.
Explained:
The deal provides Meta with enough of AMD's MI540 GPUs and next-gen CPUs to power roughly six gigawatts of data center capacity.
As part of the agreement, Meta received a performance-based warrant to acquire up to 10% of AMD, fully vesting if AMD's stock price reaches $600.
This partnership solidifies AMD's position to challenge Nvidia's dominance in the AI chip market and helps Meta diversify its suppliers amid reported delays on its own in-house chips.
Why It Matters:
This staggering investment highlights the colossal capital required to lead in the AI infrastructure race, instantly elevating AMD as a primary competitor to Nvidia. For users, it signals Meta is aggressively building the foundation for powerful, personalized AI systems that will deeply integrate into our daily lives.
The Diffusion-Powered LLM
Next in AI: AI startup Inception Labs has launched Mercury 2, a new model it claims is the world's fastest reasoning LLM. Instead of generating text one token at a time, Mercury 2 uses a diffusion-based architecture to produce entire responses in parallel.
Explained:
The new architecture allows Mercury 2 to generate text at over 1,009 tokens/sec on NVIDIA Blackwell GPUs, a greater than 5x speed increase over other leading speed-optimized models.
This speed is designed for latency-sensitive production tasks where instant responses are critical, such as agentic loops and real-time voice interfaces, interactive coding, and RAG pipelines.
The model is priced competitively at $0.25 per million input tokens and is fully OpenAI API-compatible, allowing developers to integrate it into their existing stacks without significant rewrites.
Why It Matters: Mercury 2's parallel generation method marks a fundamental shift away from the sequential, typewriter-like approach used by most LLMs. This leap in speed unlocks the ability to build more complex AI agents and fluid conversational experiences that operate within real-time latency budgets.
The Unmeasurable Speedup
Next in AI: AI research group METR is hitting the reset button on its developer productivity study because a growing number of developers now refuse to complete tasks without their AI tools, making it nearly impossible to accurately measure their impact.
Explained:
The study is suffering from severe selection bias, as developers who expect the most benefit from AI are either dropping out or avoiding submitting tasks where AI would excel.
This marks a major shift from METR's early 2025 results, which found that AI tools actually slowed down experienced developers by nearly 20%.
To get a clearer picture, METR is now exploring alternative methods like observational studies of developer workflows and fixed-task experiments.
Why It Matters: This outcome is a powerful signal that AI coding assistants are becoming indispensable for many elite developers. As these tools integrate more deeply into workflows, quantifying their true productivity boost is becoming an increasingly complex challenge.
AI Pulse
OpenAI removed the word safely from its official mission statement in its latest IRS filing as part of its restructuring into a for-profit company, sparking debate about its shifting priorities.
Firefox launched version 148 with a new AI kill switch, a feature allowing users to completely disable all AI functionalities like chatbot prompts and AI-generated link summaries.
Researchers used a physics-tailored neural network to analyze dusty plasmas, helping to uncover new physical laws and correct long-held assumptions about the fourth state of matter.
NVIDIA revealed in its annual healthcare survey that AI adoption is accelerating, with 70% of organizations actively using AI and 85% of executives reporting it helps increase revenue.