AI Frontier Daily Briefing — 2026-07-21

Note: Grok did not return an extractable answer this run (the browser was running, but automated submission did not trigger generation). The following is based on WebSearch, official vendor pages, and Ollama cross-verification.

1. Today's Headlines

• Claude Fable 5 credited with a counterexample to the Jacobian Conjecture. On July 20–21, Levent Alpöge (Anthropic / Harvard) posted on X a three-variable polynomial map whose Jacobian determinant is constantly −2 yet sends three distinct points to the same output — falsifying the 1939 conjecture (open for 87 years) in dimension ≥3. He attributed the discovery to Anthropic's Claude Fable 5. The counterexample has been independently verified (Wolfram MathWorld, SymPy); Google DeepMind's FormalConjectures Lean formalization PR is under review. (Sources: Sina, NetEase, QQ, New Scientist — vendor-official pending.)
• China's open-weight models fire a three-model volley. Moonshot AI's Kimi K3 (2.8T params, released Jul 16–17), Alibaba's Qwen 3.8 (2.4T, open-weight announced Jul 19), and DeepSeek V4 (formal build in grey-test) all landed within days. (Sources: Caixin, Huxiu, Toutiao, QQ.)

2. Model Release Roundup

• Kimi K3 (Moonshot AI, Jul 16–17): 2.8 trillion parameters, 1M-token context, native vision; Kimi Delta Attention (KDA) + Attention Residuals; MoE with 896 experts / 16 active. Topped the Frontend Code Arena leaderboard at 1679, ahead of Claude Fable 5. Full open weights drop by Jul 27. (Sources: Caixin, Huxiu, Toutiao, QQ.)
• Qwen 3.8 (Alibaba, announced Jul 19): 2.4 trillion parameters, slated for open weights; preview build Qwen3.8-Max-Preview already live on Token Plan / Qoder / QoderWork. Alibaba calls it "possibly the strongest model besides Fable 5." (Sources: Huxiu, Toutiao, AI Insiders — vendor-official pending.)
• Microsoft MAI family: at Build 2026 (Jun 2) Microsoft shipped 7 in-house models; flagship MAI-Thinking-1 (35B active params, 256K context) now handles tens of thousands of weekly Excel/Outlook prompts, displacing OpenAI and Anthropic models. (Sources: Bloomberg, Huanqiu, Cailianhe.)
• DeepSeek V4 API migration — hard cutoff Jul 24, 15:59 UTC: deepseek-chat / deepseek-reasoner retire and map to V4 (note: reasoner defaults to Flash, not Pro). (Source: aggregator — vendor-official pending.)

3. Industry & Capital

• Kling AI (Kuaishou) closed a ~$3B round (announced Jul 2) at an $18B post-money valuation — a record single raise for a multimodal model company — with Tencent, Alibaba Cloud, and Baidu among backers. (Sources: Xinhua, Pandaily.)
• Current AI secured $400M for public-interest AI infrastructure (incl. $100M from France). (Source: TechCrunch — aggregator, vendor-official pending.)
• Neko Health raised $700M to scale AI body scans; Bunkerhill raised $55M for agentic AI across health systems; Agility Robotics opened a Digit humanoid training center in Fremont, CA. (Sources: aggregators — vendor-official pending.)

4. China Power

• Kimi K3 became the world's largest open-weight model (2.8T); a compute crunch forced Moonshot to pause new consumer sign-ups on Jul 19. (Sources: Caixin, Huxiu.)
• DeepSeek V4 formal build in grey-test; inference compute reportedly cut to ~27% of its predecessor at 1M context. (Sources: Toutiao / aggregator — vendor-official pending.)
• Zhipu GLM-5.2, MiniMax M3 and others continue to advance (debuted at WAIC). (Sources: Toutiao / aggregator — vendor-official pending.)
• Kling AI spun off with independent funding, signaling Chinese multimodal going global. (Source: Xinhua.)

5. Today's Observation

• AI is shifting from verifying human proofs to constructing verifiable mathematical objects: Fable 5 is credited with the Jacobian counterexample, yet Alpöge has not disclosed the prompt or process — the human–model division of labor remains undefined.
• Open-weight models have entered the "3T era"; China's labs (Kimi / DeepSeek / Qwen) are closing the gap on scale and price-performance, though composite intelligence still trails Fable 5 / GPT-5.6 Sol.
• "De-third-party-ization" is accelerating among giants: Microsoft's MAI lands in Office, DeepSeek tightens its API, Kling raises independently — compute cost and supply-chain autonomy are the defining themes of H2 2026.
• Policy watch: the EU Parliament's EPGenAI Hub is set to reach MEPs by September (Meta / OpenAI / Anthropic / Mistral); US public-health agencies will pilot OpenAI and Anthropic models.