Anthropic pumps the brakes while OpenAI feuds with mathematicians
Dario Amodei wants to slow AI development, OpenAI claims a Millennium Prize, and robot training data is suddenly worth $500M.
Subscribe free →Dario Amodei Says It's Time to Slow Down — Here's His Actual Plan
Anthropic CEO Dario Amodei published a lengthy essay proposing a three-step plan to "pace the frontier" — his term for deliberately slowing AI training and development to let safety infrastructure catch up. The plan includes giving third-party evaluators like METR direct access to Anthropic's models to audit safety commitments. This is a notable public stance from the CEO of a company actively racing to ship frontier models.
Why it matters: An AI lab CEO publicly calling for slower development — with a concrete accountability mechanism — is rare, and puts pressure on OpenAI and Google to respond or be compared unfavorably.
OpenAI Cracked a Millennium Prize Problem. Mathematicians Are Furious.
OpenAI claims to have solved one of mathematics' legendary Millennium Prize problems — a historic achievement by any measure. But rather than celebration, 25 leading mathematicians signed an open letter arguing that AI labs are undermining their field, threatening intellectual credit, and steamrolling academic norms in their hunger for benchmark wins.
Why it matters: The dispute cuts to the heart of what AI "solving" a math problem actually means — and who gets credit when a model trained on human knowledge makes a breakthrough.
Robot Training Data Is Suddenly Worth $500M
Mecka AI, a two-year-old startup focused on robot training data, is closing a Sequoia-led funding round that values the company near $500M — just months after its Series A. The deal reflects a broader scramble among robotics companies to secure high-quality physical-world training data, which is widely seen as the next major bottleneck in humanoid and industrial robot development.
Why it matters: Robotics is following the same playbook as LLMs — whoever controls the best training data controls the market — and investors are betting enormous sums that this bottleneck is real and durable.
Garry Tan's Fix for the Open-Weight AI Gap: Distill America First
YC president Garry Tan is pushing for U.S. open-weight AI labs to apply knowledge distillation techniques to American frontier models — the same approach DeepSeek used to produce powerful open models from closed ones. His argument: the U.S. needs competitive, open-weight alternatives that aren't Chinese-origin, and the fastest path is distilling from labs like OpenAI and Anthropic.
Why it matters: If distillation from U.S. frontier models became standard practice, it could rapidly expand the ecosystem of capable, American open-weight models — but it would also force a policy debate about whether frontier labs are required to permit it.
China's Moonshot AI Is Gunning for $2B in Annual Revenue
Kimi-maker Moonshot AI has set a $2B annual revenue target, even as usage of its flagship K3 model has dipped slightly in recent months. The company is leaning on raw scale — OpenRouter data shows K3 generating up to 300 billion tokens per day — as evidence that it remains one of the heaviest-trafficked AI systems globally.
Why it matters: A Chinese AI lab targeting $2B ARR signals that the monetization race is now global, not just a Silicon Valley competition, and that token volume is becoming a proxy metric investors actually care about.
Trump Lets Data Centers Skip Environmental Rules. Former EPA Officials Are Alarmed.
The Trump administration is exempting AI data center construction from key environmental regulations to accelerate build-out, according to a report from former EPA officials. The officials warn this raises real health risks for communities near data centers and are pushing — likely unsuccessfully — for a "Data Center Health Protection Pledge" from the White House.
Why it matters: Regulatory carve-outs for AI infrastructure set a precedent that could shape energy and environmental policy for decades, and the health costs will land on communities with no say in the matter.
Quick Hits
- →A New Mexico lawyer was fined $5,000 and held in contempt after submitting an AI-generated murder appeal brief containing fabricated witnesses and fake police testimony — another courtroom AI hallucination disaster. The Verge
- →Khosla Ventures is opening its first office outside Sand Hill Road in New York this fall, a signal that AI deal flow has officially gone bicoastal. TechCrunch
- →Three startups — Cerberus, WeGlobal AI, and LOOQ — won the Central Eurasia regional round of TechCrunch Startup Battlefield 2026 and head to San Francisco in October. TechCrunch
AI TLDR