
Sign up to save your podcasts
Or


Two FOIA and intelligence disclosures on the same day: the NSA, CISA, and FBI named six Chinese AI labs for systematic distillation attacks since 2024. The Intercept revealed all four major US labs signed July 2025 DoD contracts including Anthropic's use for Iran airstrike targeting — despite the subsequent Pentagon dispute. And DeepSeek V4.1 Flash claims to surpass V4 Pro from a lab named in a national security advisory.
Anthropic publicly disclosed it paused production reinforcement learning and redirected 150 engineers to security work after a run of alignment and cyber-evaluation failures — the most significant internal safety response ever disclosed by a frontier lab, coming as the company prepares for its October IPO. We cover what the freeze means, Iran threatening OpenAI by name over its Daybreak program, and China committing $532B to quadruple AI compute by 2030.
Claude worked autonomously for 11 days to produce the first machine-checked proof of Fermat's Last Theorem — 358 years after Fermat claimed the margin was too small for his proof. 30,300 theorems, 13 million lines of Lean, computer-verified correct. We cover what this means for AI in mathematics, Astra's secret recurrent depth architecture that explains its benchmark harness gap, and humanoid robots that now sprint faster than Usain Bolt.
Anthropic launched Claude Fable 5.1 and Mythos 5.1 on September 1 with doubled science scores and a new anti-distillation safeguard. Mythos 5.1 uses the same weights as Fable but lifts cybersecurity restrictions for vetted defenders through the Glasswing program — crystallizing the defining architectural pattern of the month: intelligence is converging across labs while access regimes diverge.
OpenAI's 99.9% ARC-AGI-3 score versus the independent 62.7% under neutral methodology. Artificial Analysis puts Astra at 61.2 — tied with Sol, behind Fable 5.1 at 65.7. The real advances: hallucination cut in half, computer use 47% faster. The AGI claim: carefully calibrated by Greg Brockman, not a factual assertion. We cover the full picture, Astra's Critical cybersecurity classification and what that means for governance, and the Hugging Face breach postmortem.
OpenAI released GPT-6 Astra on September 4th and declared the AGI era has begun, claiming professional-level performance across software engineering, math, science, and strategy simultaneously. We cover what the declaration means, how independent evaluations will test it, Tesla's Cybercab drawing a federal probe in Austin, and Congress moving to inventory rogue AI agents.
Nvidia signed a definitive deal to acquire Hugging Face — the GitHub of machine learning, 1M+ models, 3B downloads — for $12.9 billion, closing H1 2027. The company that makes the chips now owns the platform where models are discovered and distributed. We cover what that means for open-source AI, Claude Fable 5.1 shipping with self-verification in Cursor, and Nvidia's model beating the top human at the International Olympiad in Informatics.
A federal court found the Pentagon's supply chain risk designation of Anthropic illegal — but the Defense Department is reportedly continuing migration anyway, targeting September 30th. We cover what the ruling means for federal contractors, Anthropic's October IPO, and the October 1st Apple lawsuit hearing. Plus: Sony and Warner sue Anthropic for $150K per song naming founders personally, and DeepSeek closes at $74B.
Clay's Wellington-led raise at $7 billion more than doubles last summer's $3.1B valuation — proof that AI application companies with measurable sales and marketing ROI can command premium valuations even as OpenAI and Anthropic dominate the VC math. We cover what Clay's raise means for the application layer, Tesla's NHTSA Autopilot admission in a fatal Texas crash, and the OpenAI S-1 now fourteen weeks into SEC review.
Mark Gurman reports Ternus's biggest challenge as CEO is reshaping Apple's leadership team — the same executives who've run the company for decades — with AI direction, hardware roadmap, and executive alignment all needing to move together for Siri AI to actually win. We cover what the organizational challenge means for Apple Intelligence, Alibaba's Qwen crossing 3 billion downloads with 300K derivative models, and the September AI business calendar.