Corroboration digest (all vendor-adjacent)
From the AI digest for 2026-10-01
- Gemini 4 Argon: Andon Labs' Blueprint-Bench 2 #1; AA hallucination 15% (vs Grok 4.7 29%, Astra 45%, Opus 5.5 59%); 300 TiB of DC memory freed by Argon agents — all RT'd by Hassabis himself; the amplification is the message: Google wants the eval record circulating while access is still gated. (Andon RT · hallucination RT)
- GPT-6 1.1 Sol speed fix (sama): "6.1 Sol was our fastest-growing model ever, and was a bit slow under load. Should be much better now." (link)
- Efficiency arms race, internal evals (Siqi Chen): 6.1 Sol at 3× opus 5.5 efficiency (2.5× sonnet 5.5) in cfo.ai's internal evals — after Opus 5.5 had scored 2× vs gpt-6-sol. Internal, unverified, but the vendors are now trading efficiency one-upmanship in public. (link)
- Perplexity day: open-source decider + Decisions API ($0.04/M), TradingView charts in Computer, Amex skill library, Seedance 2.5. Covered as an
aistory. (Arav) - Grok Bot: software-build upgrades (Cursor hand-off, GitHub/Origin PR plugins, video demos, speed, connector-aware proactivity), SpaceXAI sharing its engineers' bots, and the "building Grok Bot with Grok Bot" genre — harness-race thread material. (Grok Bot)
- Claude Code mods — marketplace, TS mods, forked agents. Covered in the
aidigest. (thread) - SynthID/provenance podcast (DeepMind) — continuity for the provenance thread. (link)
- Peter Walker: OpenAI "continues to lead by a solid margin in model number" "(one benchmark)" — corroboration only. (link)