8/3: Hard Takeoff, But for Math
OpenAI finds 10 new math results, Alibaba releases Qwen3.8, White House completes AI framework, CXMT and Zeiss will expand capacity, China tightens circuit protections, the UK may regulate AI
Happy Monday, monitors. It’s now August — just two months away from models surpassing humans on the Epoch Capabilities Index. Be sure to catch us live on X and YouTube, and follow us on Instagram.
Today’s Experts
John-Clark Levin (Kurzweil Technologies)
Pim de Witte (General Intuition) and Ken Colton (Medal)
Finn Metz (Seldon Lab)
Roman Stolyarov (Lambda Surgical)
Haseeb Qureshi (Dragonfly)
Tyler Johnston (The Midas Project)
Joe Nocera (The Free Press)
Today’s Situations
OpenAI makes huge math progress. Over the weekend, the company announced that an internal version of their next major model (GPT-6?) discovered 10 major new results in math and computer science for just $2,000 in tokens1. This comes just a few weeks after an OpenAI model disproved the Erdős unit distance conjecture, the first major novel AI-driven result in mathematics. There is so, so much more to come.
Alibaba releases Qwen3.8. It’s still not quite as generally intelligent as Fable, but it beats it (and Sol) on a number of benchmarks: PaperBench, PerceptionBench, and OSWorld-Verified, for example. It can do serious long-horizon autonomous coding tasks and assist with AI research. It ranks the highest of any non-Anthropic model on TextArena, and has reached the Pareto frontier for front-end code. And it’s $2/$6 per million input/output tokens, cheaper than Kimi K3 ($3/$15) and much cheaper than Fable ($10/$50).
The White House has completed its AI framework. In June, President Trump issued an executive order directing federal agencies to create a classified cyber benchmark for frontier models and a voluntary framework for frontier labs to share their models with the government before public release. Tomorrow, OpenAI, Google, Anthropic, and Meta staff will meet with admin officials to discuss the completed framework.
CXMT is planning a new memory fab in Beijing. It currently operates one in Beijing and two in Hefei, the city where it’s headquartered. The company, the world’s fourth-largest memory manufacturer, is currently seeking to rapidly expand production to meet the insatiable demand for memory.
Zeiss says it can handle increased AI demand. The German company manufactures optical systems, including extremely precise mirrors used in ASML’s lithography machines.
China tightens circuit design protections. Under new rules set to go into effect in October, chip layout designs will be subject to stricter registration standards with penalties for infringement. China is also considering export controls on strategically important technologies.
The UK is open to AI regulation. Kanishka Narayan, the UK’s first Cabinet-level Minister of Artificial Intelligence, says that if Britain’s current voluntary pre-deployment testing regime isn’t enough to ensure safety, they may implement regulation. The UK AI Security Institute (AISI) makes Britain the only country other than America with pre-deployment access to frontier models.
TikTok is finalizing three settlements over addiction. The plaintiffs allege that the platform caused them to suffer from a combination of addiction, depression, anxiety, eating disorders, self-harm, and other mental health conditions. Two of the three are minors.
Base Power raises $1B at a $13B valuation led by Ribbit, Addition, Valor, and JPMorgan. Base builds home battery systems that both provide power to customers and can feed back into the grid during peak demand times.
Valar Atomics raises $1B at a $6B valuation led by Sequoia. Valar builds mass-producible small modular reactors (SMRs) for use in factory production, data centers, and other industrial applications. Valar, founded in 2023, became the first private reactor to reach criticality outside of a national laboratory.
Amazon crosses $3T in market cap thanks to strong earnings, bolstered by cloud (AWS) and AI.
Banger Review
This is especially incredible, considering that a senior academic mathematician might make ~$150k a year and take months or years to produce a breakthrough of this magnitude, let alone ten of them. And the $2,000 is measured in GPT-5.6 Sol API rates. Remove the lab margin and the actual cost to OpenAI was probably just a few hundred dollars for ten major novel results!