Jacobian Conjecture Falls, Moonshot Accused, DeepSeek Retires Today

Good morning. It's Friday, July 24, and we're covering the most significant mathematical result AI has contributed to in living memory, a White House accusation that China's Moonshot copied Anthropic's flagship model, a hard API cutoff landing this afternoon, and the 72-hour window before Kimi K3's open weights drop.

The Jacobian conjecture HN thread closed the week at 665 points. That number is the simplest way to describe how the developer community received the news that a model helped disprove a problem that has resisted human effort since 1939.

In today's brief:

• Claude Fable 5 helped mathematician Levent Alpoge produce a degree-7 polynomial counterexample to the Jacobian conjecture in C^3, settling a problem open since 1939. Fields medalist Terence Tao published a 4,000-word digestion. Timothy Gowers called it "the first time an LLM has solved a well-known problem I'd heard of outside my area."
• White House OSTP Director Michael Kratsios stated publicly on July 22 that Moonshot AI "distilled Anthropic's Fable for the development of its K3 model" and used banned Nvidia chips routed through Thailand. Moonshot's timeline makes the distillation claim implausible. The governance fallout is real regardless.
• DeepSeek's deepseek-chat and deepseek-reasoner aliases expire at 15:59 UTC today. Any API call using those names after the cutoff returns an error. Migrate to deepseek-v4-flash and deepseek-v4-pro now.
• Kimi K3 open weights arrive Sunday under a modified MIT license. Tonight is the last realistic window for infrastructure prep if you plan to self-host.

RESEARCH

Claude Fable 5 Helped Disprove an 87-Year-Old Conjecture

Around 2:19 AM UTC on July 20, mathematician Levent Alpoge posted that Claude Fable 5 had helped him produce a polynomial map C^3 to C^3 with Jacobian determinant -2 that is not invertible. That single sentence ended an 87-year chase.

The Jacobian conjecture, posed in 1939, asks whether any polynomial map F: C^n to C^n with a nonzero constant Jacobian determinant must be a polynomial automorphism - that is, whether local invertibility everywhere implies global invertibility. For n=1, the answer is trivially yes. For n=2, the question is still open. For n greater than or equal to 3, Fable 5 and Alpoge found the answer is no.

The counterexample is a degree-7 polynomial map. Three distinct points in C^3 share one output value, so the map is everywhere locally invertible but not globally injective. Mathematicians had been guessing the minimum degree of a counterexample, if one existed, would be somewhere in the range of 200. It was 7. Verification preprints appeared within hours of the post; the math community considers the result solid.

On July 21, Fields medalist Terence Tao published "A digestion of the Jacobian conjecture counterexample" - a structured walkthrough of the construction and why it works. The Hacker News thread (665 points) drew reaction from working mathematicians. Timothy Gowers wrote: "This is the first time an LLM has solved a well-known problem I'd heard of outside my area."

What does the collaboration look like in practice? Based on Alpoge's account, it is not the model producing a proof from a blank prompt. It is a mathematician with a well-formed conjecture and some partial intuitions using the model to navigate a hypothesis space that human intuition had preemptively bounded too conservatively. The degree-7 barrier was below what experts expected to search. That is the specific contribution: not generating the entire proof, but finding a candidate construction where the human researcher had incorrectly assumed none existed at that complexity level.

For developers thinking about research workflows, the signal here is not that LLMs can replace mathematical expertise. It is that they can be useful search partners in spaces where domain experts have systematic blind spots. The n=2 Jacobian conjecture is still open, and the search is harder.

Our coverage: Terence Tao Digests the Jacobian Conjecture Counterexample

Why it matters: AI has now produced a verified result at the frontier of professional mathematics - not a code generation task or a document summary, but the resolution of an open research problem. That changes what developers and researchers should expect from these models in adjacent technical domains: theoretical computer science, combinatorics, complexity theory. The question is not "can AI do math research" but "under what workflow conditions does it accelerate human researchers most."

POLICY

White House Accuses Moonshot of Distilling Fable 5 and Using Banned Chips

On July 22, White House Office of Science and Technology Policy Director Michael Kratsios said on the record that the administration has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model." In the same statement, he accused Moonshot of covertly accessing restricted Nvidia chips through Thailand-based servers to build K3's training infrastructure.

This is the most direct accusation a senior U.S. official has made against a specific Chinese AI lab's model development practices. The SiliconAngle report has the full statement; the Hacker News thread ran through the night.

Moonshot employees pushed back immediately and the pushback has teeth: Fable 5 became publicly accessible July 1. Kimi K3 launched July 15. The claim is that Moonshot trained a 2.8-trillion-parameter frontier-scale model in approximately two weeks using distillation from a model that only became publicly available that same fortnight. Multiple AI researchers - including several with direct knowledge of the compute requirements for models at this scale - have publicly disputed the timeline as physically implausible.

What Kratsios described as "distillation" is technically ambiguous. Knowledge distillation in the formal sense requires training a student model on outputs from a teacher model over many training runs - not a two-week operation at this scale. Behavioral alignment through targeted post-training on a small amount of Fable outputs is faster but produces a different kind of result. The White House statement did not specify which technique it meant, and that ambiguity is doing a lot of work in the policy framing.

The chip-routing accusation is separate and more specific. Using Thailand-based servers to access restricted Nvidia accelerators would implicate export control laws directly. No enforcement action has been announced, but the statement comes as the administration works toward a frontier AI governance framework before August 1 - the deadline Kratsios has signaled for a formal policy response.

The practical governance question the HN thread has been wrestling with: if "distillation from a publicly available model" becomes a de facto violation, what does that mean for the entire open-weight ecosystem? Every open-weight model and API with public access is a potential distillation source. The legal and technical line between "learning from" a model and "copying" it is not established in U.S. law. This case may force that definition.

Our related Kimi K3 coverage: Kimi K3 - Moonshot's 2.8T Frontier Model

Why it matters: Whether or not the specific distillation claim holds up, this statement marks a qualitative shift in how the U.S. government is engaging with Chinese AI capability development. Developers building on open-weight Chinese models - including Kimi K3, whose weights arrive Sunday - should watch how this framework develops. A regulatory definition of distillation-as-theft would have direct implications for fine-tuning, RAG pipelines, and model merging workflows that touch any model covered by export controls.

TOOLING

DeepSeek's Legacy API Names Retire at 15:59 UTC This Afternoon

If your codebase is still calling deepseek-chat or deepseek-reasoner, you have until 15:59 UTC today before those requests start returning errors.

DeepSeek launched V4 on April 24 and kept the old model names alive as compatibility aliases to ease the transition. The 90-day window is up. The migration is a one-line change in most integrations, but only if you make it before the cutoff.

The mapping is direct: deepseek-chat handled the non-thinking generation mode and maps to deepseek-v4-flash. deepseek-reasoner handled the thinking mode (extended chain-of-thought) and maps to deepseek-v4-pro. The reasoning mode behavior is the same; only the identifier changes.

Check anywhere you send a model name as a string: environment variables, config files, LangChain or LlamaIndex chain definitions, agent harness defaults, proxy router configs, and any SDK wrapper that specifies a default model. There is no announced grace extension and no soft redirect that keeps the old names limping along. The cutoff is hard.

Our migration guide with verified pricing and a step-by-step checklist: DeepSeek Chat to V4 Migration Guide

Why it matters: This is operational, not strategic - but a missed cutoff means broken pipelines on a Friday afternoon with no fix that does not require a deploy. The migration is low-risk. Not migrating is not.

MODELS

Kimi K3 Open Weights Land Sunday - Infrastructure Prep Window Closes Tonight

Moonshot AI's 2.8-trillion-parameter Kimi K3 open weights arrive this Sunday, July 27, under a modified MIT license. We published a full provider and pricing comparison yesterday covering every current access path: Where to Access Kimi K3.

For teams planning self-hosting: full-precision inference at 2.8T scale requires multi-node GPU infrastructure. The community quantization push - GGUF, AWQ, and GPTQ variants - will land within 24 to 48 hours of the weight release as usual. If you need quantized inference on Sunday itself, watch Hugging Face and the Moonshot community channels; the first community quants will appear within hours of the weights going live.

If you plan to run a multi-node setup rather than waiting for community quants, tonight is the realistic window for provisioning. Weekend weight drops tend to land early UTC and the traffic spikes immediately.

The API is live now for teams that do not need self-hosting: $3 per million input tokens and $15 per million output on the Kimi platform, with OpenRouter and Cloudflare Workers AI routes also available.

Our architecture and deployment guide: Kimi K3 Developer Guide | Kimi K3 vs K2.7

Why it matters: Sunday marks the start of the community quantization and fine-tuning cycle that will determine whether Kimi K3 is accessible at smaller hardware footprints. Teams that have been waiting for a credible open-weight frontier model to self-host have approximately 60 hours.

TOOLS WORTH A LOOK

1. ROCm.ai - AMD's new AI-assisted GPU development platform announced at Advancing AI 2026. Connects coding agents (Claude, Codex, Cursor) to AMD GPU platform knowledge and ROCm APIs natively. Free for developers targeting AMD hardware. (Free / developer)

2. Kimi Code - Moonshot's coding agent powered by Kimi K3. Available now on the Kimi platform; supports OpenAI-compatible API calls so existing agent harnesses work without modification. (Free tier / paid credits)

3. deepseek-v4-flash / deepseek-v4-pro - If you have not migrated yet, the DeepSeek V4 identifiers are live and tested. Flash is the non-thinking general-purpose model; Pro is the thinking-mode variant. The DeepSeek API docs have the change log. (Pay-per-token)

WHAT ELSE IS HAPPENING

• AMD ships MI350P and Helios rack system: At Advancing AI 2026, AMD announced the Instinct MI350P GPU (up to 4.2x more tokens per second per dollar vs. competition per AMD) and Helios - a rack-scale system with 72 MI455X GPUs, EPYC Venice CPUs, and 2.9 exaflops FP4 inference at $5-5.25M per rack. Direct competition with NVIDIA's GB200 NVL rack configs. (press release)

• Moonshot suspends Kimi K3 subscriptions: Demand for Kimi K3 access exceeded capacity and Moonshot temporarily halted new subscriptions. Existing API keys continue working; new sign-ups are waitlisted. The open-weight release Sunday should relieve some demand pressure. (HN thread)

• Cursor's SQLite swarm experiment: Cursor used an agent swarm to rebuild a SQLite-like database from documentation and pass a held-out conformance test suite. Our analysis covers what it signals about the shift from task-level to goal-level software engineering - specifying what a system must do, then letting agents implement and verify it. (5 min read)

• SearchOS and shared-state deep research: SearchOS replaces the growing chat transcript pattern with shared agent state - frontier tasks, evidence graphs, coverage maps, and failure memory. If you are building any kind of multi-step research agent, this is the architecture pattern to understand. (4 min read)

• White House frontier AI framework before August 1: Kratsios signaled a policy response framework is coming within days. The Moonshot accusation is part of the same announcement push. Watch for export control extensions and possible model-distillation definitions in the framework text. (political / policy)

• Future of Life AI Safety Index: best grade is C+: The Summer 2026 index (published July 7) graded nine frontier labs on 37 safety indicators. Anthropic led at C+; OpenAI and Google DeepMind at C; Meta at D+. The independent panel noted that Anthropic, OpenAI, Google DeepMind, and Meta have all walked back previous commitments to pause development unilaterally if specific risk thresholds are crossed. Worth reading if you are making vendor decisions where safety posture matters. (report)

FROM THE SITE

What We Published This Week

Terence Tao Digests the Jacobian Conjecture Counterexample - a deep dive into what happened, the mathematical significance of the degree-7 construction, and what it means for AI in technical research workflows. Published July 23.

Where to Access Kimi K3 - every provider and pricing path for Kimi K3 today, from the Moonshot first-party platform to OpenRouter, Cloudflare Workers AI, and agent harnesses including OpenCode Go and Kimi Code. Published July 23.

---

Every link goes to a primary source or sourced coverage. The weekend edition posts Saturday - subscribe to get it by email.