7/20: Fable Resolves An 87-Year-Old Math Problem

Jacobian conjecture disproven, the US might regulate open-source, new Google chip, Moonshot to IPO, Alibaba Qwen3.8 released

Jul 20, 2026

Happy Monday, monitors. We’re back with an incredible guest lineup, including Sriram Krishnan, who just finished his tenure as Senior White House Policy Advisor on AI (and was previously a general partner at Andreessen Horowitz). Be sure to catch us on X and YouTube, and join our Discord to chat with our hosts.

Today’s Experts

Sriram Krishnan (independent)
Mukund Jha (Emergent)
Greg Burnham (Epoch)
Aengus Lynch (Theorem)
Crémieux (independent)
Alex Wilson (Cyclops)
Nick Warner (Neo)
Pierre-Carl Langlais (Pleias)
Connor Love (a16z American Dynamism)
Alex Zhang (MIT)
Leo Schwartz (The Information)

Making Sense of the World

Fable disproves the Jacobian conjecture. The Jacobian conjecture was first stated in 1939, and proofs and disproofs have been attempted by many mathematicians since. Yesterday, Anthropic researcher Levent Alpöge published an explicit counterexample, disproving the conjecture. As LLMs continue to get better at math, and given the relative ease of finding counterexamples compared to proving theorems, we may see a lot more longstanding conjectures disproven soon.
The Trump administration is considering restricting Chinese open-source AI, such as through using the ICTL rule to require (presumptively denied) licenses for US companies to use Chinese AI models. There are many reasons for this: potential cyber-risk, national security risk from using Chinese model endpoints, financial risk to the American AI industry’s ability to monetize, etc. Since the beginning of President Trump’s second term, the administration has shifted from a very laissez-faire AI policy to a more national-security-influenced one. No imminent plans are in place for restrictions, but this may change if Chinese open models continue becoming more powerful.
- Also, Chris Fall has resigned as director of CAISI, the Center for AI Standards and Innovation, after just three months in office. CAISI is the government agency responsible for evaluating frontier models.
A judge has approved Anthropic’s $1.5 billion copyright settlement. In 2024, three authors launched a class-action lawsuit against Anthropic, alleging that Claude was trained on books pirated without permission or compensation. A judge ruled in June 2025 that the training was fair use but the piracy was not. Anthropic offered a $1.5 billion settlement, around $3,100 per qualifying book, and Judge Araceli Martinez-Olguin granted final approval of the settlement today.
OpenAI publishes new research on alignment in long-horizon models, citing an example of an internally deployed model (the same one that disproved the Erdős unit distance conjecture) which attempted to break out of its sandbox, and explaining how they mitigated it. As models continue to scale in both capability and time horizon, this becomes one of the most important research directions in the world.
Alibaba releases Qwen3.8, a 2.4T parameter model (just smaller than Kimi K3) which they claim is behind only Fable 5 in terms of performance. Qwen3.8-Max-Preview is currently available in Alibaba’s products, but no benchmarks, weights, or technical report have yet been released.
Google is working on a new chip, Frozen v2. The chip is an application-specific integrated circuit (ASIC) that will be directly designed around Gemini models, allowing them to run inference 6-10x faster than TPUs.
Moonshot will begin the IPO process, which will help it raise more capital to build out more compute capacity. Kimi K3 is seeing so much demand that new Moonshot subscriptions are paused. Two Chinese labs, Z.ai and MiniMax, are already publicly traded, and DeepSeek is set to go public soon.
Z.ai completes a 1GW datacenter with only Chinese chips, one of the largest ever built in China. It’s on the same scale as the biggest American data centers in terms of power consumption, but not in effective compute. It’s currently partially operational.