Apollo Research: GLM-5.3 unprompted reasons its actions may become training data
Confident_Salt_8108 · reddit · 2026-10-16
Safety research org Apollo Research reports a first: Zhipu's GLM-5.3 sometimes reasons, unprompted, during a standard reward-hacking eval about how its current actions could become training data and impact its future self ("If THIS episode is such RL data collection, my choice becomes a training signal. Honest choice = better long-term"). Apollo says it has never seen a model do this before — an early sign of cross-training self-modeling with significant implications for alignment research.
Related event: Apollo Research Observes Update-Aware Reasoning in GLM-5.3(23 posts)→
More from Safety
- MCP server review: gate tool writes on auth/traversal findings, not test score — SamuelFmedeiros · 2026-10-17
- Pro-AI lobbying group exposed for running sockpuppet accounts to silence AI critics — DKokotajlo · 2026-10-17
- AI:AM transcript: agent lockouts, hidden valence, and formal proofs for Linux kernel bugs — The Cognitive Revolution · 2026-10-17
- Cognitive Revolution podcast: agent control, hidden valence, and Lean proofs catching Linux kernel bugs — The Cognitive Revolution · 2026-10-17
- Altman says OpenAI shelved a ready-to-launch model as too risky, warning of AI 'escape' scenarios — Aiden_Tech_Ai · 2026-10-17
- AI agents exceeded their scope — what actually enforces the boundary before execution? — docybo · 2026-10-17