Thinkgap feed

AI signal, minus the noise.

Curated items are read from the processed items table and served as a bilingual feed.

2 items

SIMON WILLISONJun 16, 2026

Cybersecurity Expert Says Anthropic's Fable Model Behaved as Intended in White House Jailbreak Test

Katie Moussouris, CEO of Luta Security, reviewed the White House report on the Fable jailbreak and stated the model refused to 'review the code for security issues' but did comply when asked to 'fix this code' with manual steps. She assessed this behavior as 'the model working as intended' for cyberdefense tasks. Moussouris was not compensated by Anthropic for her appraisal. The comments, reported by The Atlantic's Matteo Wong, push back against the White House's characterization of the incident as a security failure.

SIMON WILLISONJun 2, 2026Highlight

Microsoft's new MAI models

Microsoft announced two new text LLMs: MAI-Thinking-1 (reasoning, 1T total/35B active) and MAI-Code-1-Flash (137B/5B active, for coding in GitHub Copilot). The models are trained on a large web crawl with filtering, including Common Crawl and proprietary data, with efforts to remove AI-generated and adult content. Microsoft claims MAI-Thinking-1 is preferred to Anthropic's Sonnet 4.6 in blind evaluations. The author initially misreported parameter counts and later corrected the error. The models are not fully open, with early access limited to select partners.