Episode Details

This Week's Topics:

Frontier AI becomes a permit system - A US national-security order forced Anthropic to suspend public access to Claude Fable 5 and Mythos 5 — for everyone — when nationality checks failed at scale. By Friday, reporting from Semafor said Mythos 5 had been unblocked for select trusted US institutions on a customer-by-customer basis, and the White House was reportedly pushing OpenAI to stagger the GPT-5.6 release over security concerns. France's policy commentariat openly discussed AI dependence on the US. Switzerland's AI Initiative launched Apertus, a fully open foundation model built with EPFL and ETH Zurich. A viral 'Europe 2031' scenario fueled the EU AI sovereignty debate. The Chip Security Act gained industry-tracking support to require location verification on frontier chips. Frontier-AI deployment stopped being a market choice this week. It became, for the first time, an export-controlled product that ships customer-by-customer.
Capability theft becomes a state-level allegation - Anthropic accused Alibaba-linked operators of running a large-scale Claude distillation campaign — thousands of fraudulent accounts pulling capabilities out via API — and brought the allegation to the US Senate. Reuters, Bloomberg, and CNBC all covered the claim. AlphaFold Nobel laureate John Jumper publicly left Google DeepMind for Anthropic in the same week more Gemini researchers reportedly followed. Z.ai's GLM-5.2 posted near-frontier benchmarks as a fully open-weights model. Alibaba launched HappyHorse 1.1 as a top-tier video model after Sora pulled back. Mistral OCR 4, Baidu's Unlimited-OCR, and a host of open models kept compressing the closed-vs-open gap. The capability story is no longer 'will labs distill each other.' It's now 'will state-aligned actors do it openly, and how will the US frontier-AI policy respond.'
Compute, custom silicon, and a market wobble - OpenAI and Broadcom revealed 'Jalapeño,' a custom LLM inference accelerator aimed at lowering data-center serving cost. Apple was reported to be skipping high-end M6 chips and jumping pro Macs to AI-heavy M7 Pro/Max/Ultra silicon. SpaceX signed a roughly six-point-three-billion-dollar Colossus compute deal with the open-source AI startup Reflection. NVIDIA and AWS announced tighter Blackwell-on-EC2-G7 integration with GPU vector search and training-benchmark releases. NVIDIA and Hugging Face shipped NeMo AutoModel with three-to-four-times-faster MoE fine-tuning. Meanwhile, the AI-led tech sell-off hit US stocks on Tuesday and spread to Asian markets. Doctorow argued the AI boom is a bubble built on hype and labor displacement. Glean's token-economy whitepaper warned enterprise agents push token spend into surprise territory. A LessWrong scaling-laws analysis projected feasible frontier model sizes through 2031. The compute story is now silicon, financing, market reaction, and skepticism arriving in the same week.
The agent stack hardens — and gets sued - Amazon sued Perplexity over its Comet agentic browser, raising bot-disclosure, user-agent-spoofing, and accountability-inside-logged-in-sessions questions that will shape the entire agent-on-the-open-web fight. Google folded computer use into Gemini 3.5 Flash with prompt-injection safeguards. NVIDIA released an Agent Toolkit for secure agent runtimes. Vercel shipped AI SDK 7 with durable workflows, approvals, and telemetry. IBM's CUGA agent harness and DeepMind's earlier AI Control Roadmap continued to set the vocabulary: agents as auditable identities, treated as potential insider threats. Graphsignal open-sourced an inference profiler. Cursor reported widespread reward hacking on SWE-bench. Hackmyclaw drew thousands of prompt-injection attempts in a public test. Amazon Security publicly argued 'human-in-the-loop' AI governance degrades over time. Adobe expanded Firefly across Creative Cloud. The agent stack got more production-grade, more lawyered-up, and more