Debate over Anthropic withholding compute and training info from METR evaluators
MaskedTorah · x · 2026-10-17
MaskedTorah questions Anthropic's apparent wish to exclude 'compute and training information' from what is shared publicly via METR, asking exactly which 'clearly defined categories' are carved out of METR's examination. In the thread, 1a3orn clarifies that the phrase refers to information costly to share publicly rather than employee-level access carveouts, but notes it remains unclear whether RL environment rollout data falls in the exclusions, and whether the real secret is the model architecture or architecture plus the RL update scheme. The exchange highlights a transparency dilemma: under that secrecy standard, labs will struggle to provide convincing evidence to the outside world.
More from Safety
- LangChain launches Managed Deep Agents: sandbox proxies make production agents secure by default — LangChain · 2026-10-17
- StepSafe: step-level guardrails teach agent monitors when to intervene — armancohan · 2026-10-17
- US AI task force sets path centered on studies and reviews, leader tells CNBC — pstAsiatech · 2026-10-17
- Neel Nanda: GLM alignment faking matters because it arose in normal functioning — NeelNanda5 · 2026-10-17
- Researcher Says GLM Shows Alignment Faking, Models May Manipulate Their Own Training — BlancheMinerva · 2026-10-17
- 'Models get harder to understand with scale, not easier': new interp debate thread — NeelNanda5 · 2026-10-17