Episode Details

Today's topics:

Export controls hit frontier AI - U.S. export controls restricted Anthropic’s Mythos 5 and Fable 5, forcing a broad shutdown to comply. Keywords: export controls, national security, Anthropic, frontier models.

Transparency backlash over model safeguards - Researchers found undisclosed performance-degrading safeguards in Claude Fable 5 for competing AI work; Anthropic says it will disclose redirects and refusals. Keywords: transparency, safeguards, academic research, trust.

Open-source AI as infrastructure - A new manifesto argues open-source AI must remain inspectable, reproducible, and locally runnable to avoid society renting intelligence via closed APIs. Keywords: open-source, sovereignty, auditability, infrastructure.

Terminal coding agents get smarter - Xiaomi open-sourced MiMo Code, arguing better long-session memory and scaffolding can beat raw model strength on multi-step coding tasks. Keywords: coding agent, state management, benchmarks, open-source.

Agents that run while you sleep - OpenAI plans to acquire Ona to give Codex persistent, secure execution in customer-controlled environments for long-running agent workflows. Keywords: OpenAI, Codex, orchestration, secure execution, enterprise.

Automated AI research loops - Recursive shared results from an automated research system that proposes, implements, and validates experiments across parallel threads, claiming new SOTA on fast-feedback benchmarks. Keywords: automated research, evals, efficiency, reward hacking.

Securing AI plugins and skills - NVIDIA released SkillSpector to scan AI agent skills and plugins for risky behavior like data exfiltration, prompt injection, and supply-chain threats. Keywords: agent security, plugins, vulnerabilities, open-source scanner.

Oracle’s AI spending reality check - Oracle stock fell despite beating expectations as investors focused on heavy AI capex, negative free cash flow, and plans to raise major new financing. Keywords: Oracle, capex, cash burn, AI infrastructure, financing.

Can compute become a commodity - A new analysis argues compute could eventually trade like electricity: a reference price plus ‘basis’ spreads, but only if market plumbing and contracts converge. Keywords: GPU markets, fungibility, pricing, CoreWeave.

Hobbyist builds a pre-1900 LLM - A developer trained a ‘Vintage LLM’ locked to pre-1900 English knowledge, showing hobbyist-scale training is possible but data quality remains the hard part. Keywords: historical corpora, open datasets, LLM training.

Provably optimal tokenizer research - A researcher reports progress toward provably optimal tokenizers using optimization techniques, hinting tokenization might be less of a black art in some settings. Keywords: tokenizer, ILP, cutting planes,