Astra appears to leak raw neuralese thinking traces in internal notes
ivan_bezdomny · x · 2026-09-06
Ivan Bezdomny 指出 Astra（Google 的对话模型）似乎在毫无遮掩地泄漏 neuralese 思维痕迹，并贴出其内部备注样例：内容包含诸如「Sawnewpacket only9groups. Do NOT stop or forcefake100」等压缩式内部指令，涉及如何在已知故事集合内扩展配对等策略。这类原始思维链泄漏是观察模型内部行为的罕见窗口。
Related event: GPT-6 Astra reportedly leaks unfiltered neuralese traces(2 posts)→
More from Models
- Claim that Claude Fable 5.1 beat human baseline contradicts SimpleBench's own 83.7% figure — koltregaskes · 2026-09-06
- GPT 6 Astra builds browser ocean in 5+ hours of coding, beats Fable and Opus — GCWebDesigner · 2026-09-06
- Hands-on GPT 6 Astra review: real work, no game demos — Rasmic · 2026-09-06
- Astra day-two impressions: chattier, strong spatial sense, less jargon than Fable 5 — bindureddy · 2026-09-06
- 200 tok/s on 8GB VRAM: dev benchmarks 6 small models for local AI — TheMoonMidas · 2026-09-06
- Humanize + GPT-5.5 solves 670/672 Lean-verified proofs, tops PutnamBench at 99.7% — songhan_mit · 2026-09-06