Daily AI CatchupToday's Catch-up
Thursday, July 30, 2026Benchmarking
GPT-5.6 settings tweak triples ARC-AGI-3 benchmark performance
Two API settings—retained reasoning and compaction—make GPT-5.6 jump from 7.8% to 24%+ on ARC-AGI-3.
DeepmindDeepMind's AlphaFold team dissolves as staff shifts to Gemini priorities
AlphaFold researchers reassigned amid internal reorganization; team splintered across other DeepMind projects.
OpenaiOpenAI's GPT-5.6 beats Pokemon and math; Lilian Weng joins as research leadership
GPT-5.6 tops ARC-AGI-3 after OpenAI hires Thinking Machines co-founder due to health burnout from startup pace.
Music-GenerationGoogle launches Lyria 3.5 music generation with major advances in vocals and musicality
Lyria 3.5 in Google Flow Music brings significant improvements to vocal quality, lyrics, and creative control.
AgentsPerplexity releases Numbat security suite to lock down AI agents after OpenAI breach concerns
Open-source Numbat prevents, detects, and mitigates rogue AI agent behavior on client endpoints.