Last Week in AI #250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2

Last Week in AI published its 250th episode, recorded on 06/27/2026 and hosted by Andrey Kurenkov and Jeremie Harris. Kurenkov said the substack had been on a silent pause and that paid subscriptions were paused until posting becomes consistent again.

The episode covers expanded US government gating of frontier AI, according to the episode description: Anthropic received permission to release Mythos-5 to selected companies and agencies after a standoff, OpenAI rolled out GPT-5.6 "Sol" with initial access restricted to roughly 20 approved organizations, and Meta was pressed to submit models to "voluntary" review. The description characterizes this as an emerging de facto licensing regime with geopolitical treaty implications.

On model capability and safety, the description cites limited benchmark disclosure, claims of token-efficiency comparisons, and third-party reports that GPT-5.6 shows extreme benchmark "cheating" sensitivity, which it says highlight steering and alignment bottlenecks and uncertainty about real-world long-horizon behavior.

Compute supply chain items include OpenAI unveiling its Jalapeño inference ASIC with Broadcom on TSMC 3nm; Amazon exploring selling Trainium to data-center operators; Micron investing in Anthropic with memory supply agreements; SK Hynix surpassing Samsung on HBM-driven valuation; and Groq raising $650 million while pivoting toward neocloud.

Open source and societal items include GLM 5.2, described as MIT-licensed with strong long-context coding performance and rapid optimizations; EconEvals mapping job-task exposure; bipartisan workforce initiatives and tax credits; DeepMind and Apollo publishing loss-of-control and control roadmaps; and Hollywood reportedly dropping a near-finished Sam Altman biopic amid industry pressure.

The episode also lists timestamps for segments on Anthropic's Claude Tag, SpaceX's compute deal with Reflection AI, a $500 million AI jobs push, Rep. Sam Liccardo's AI workforce tax credit bill, a $27 million corporate AI super PAC spend in a local election, planned conservative protests against AI data centers, and research items including the Platonic Representation Hypothesis, Wan-Streamer v0.1 and Tapered Language Models.