@fchollet：Chollet 主张 AI 监管应保持透明与一致，反对任意性执法；监管本身必要，但程序正义不可妥协。
Source: https://x.com/fchollet/status/2066554426551390457
Author: @fchollet (fchollet)
Original Date: Mon Jun 15 16:12:24 +0000 2026
Likes: 110 | Retweets: 12
Quality Score: 5
Fetched: 2026-06-22 20:24:54
English Original
We desperately need standardized benchmarks for agentic capabilities instead of panic-reacting to prompt-engineering parlor tricks. Until we can objectively and transparently measure what these systems actually do, the entire ecosystem will remain vulnerable to unpredictable and arbitrary government overreach.
Tweet Details:
- Author: fchollet
- Bio: Co-founder @ndea. Co-founder @arcprize. Creator of Keras and ARC-AGI. Author of 'Deep Learning with Python'.
- Likes: 110
- Retweets: 12
- Created: Mon Jun 15 16:12:24 +0000 2026
- URL: https://x.com/fchollet/status/2066554426551390457
Thread / Replies Context
@fchollet (👍 110 · 🔁 12) *In reply to 2066554345345147288*
We desperately need standardized benchmarks for agentic capabilities instead of panic-reacting to prompt-engineering parlor tricks. Until we can objectively and transparently measure what these systems actually do, the entire ecosystem will remain vulnerable to unpredictable and arbitrary government overreach.
*link*
@HEMSEYE (👍 0 · 🔁 0) *In reply to 2066554426551390457*
@fchollet The LLM approach is fundamentally flawed. Not understanding and guessing, very often getting things right is still very unsafe. My solution is in draft. Linguistic cognitive grounding to then create using a compositional language of thought is part of the direction we need.
*link*
@KissonL (👍 0 · 🔁 0) *In reply to 2066554426551390457*
@fchollet Without grounded evals for multi-step planning, the community iterates on vibes. τ-bench and WebArena point the right direction but are still too brittle to benchmark production reasoning chains — that gap is exactly where the panic-reacting you describe comes from.
*link*
@oleg_kai (👍 0 · 🔁 0) *In reply to 2066554426551390457*
@fchollet the unit is the trap. agentic capability is harness-conditioned, not model-conditioned. any benchmark picks an implicit harness shape and labs optimize the score, not the agent. arc-agi works for reasoning; nothing equivalent for tool-use under domain noise.
*link*
@cdiamond (👍 0 · 🔁 0) *In reply to 2066554426551390457*
@fchollet fear doesn't wait for evals. the Mythos shutdown was triggered by an access incident, not a capability score.
*link*
@itsmattchan (👍 0 · 🔁 0) *In reply to 2066554426551390457*
@fchollet What does that look like in your mind? Like domain based or general?
*link*
中文翻译
François Chollet（@fchollet）发文指出：
即使你支持 AI 监管，也应该认识到：不透明且任意性的监管打击对整个行业都是适得其反的。
监管的合法性来自"明确、一致、可预期"三个基本属性。监管对象只有能在事前预判合规边界、事后能复盘判断依据，市场才能形成稳定的创新环境。任意性执法只会：
- 把创业精力从"做产品"转向"猜监管"；
- 抑制中小玩家参与，因为合规成本不成比例；
- 让本应被监管的真正风险被噪音掩盖。
共识：Chollet 并不是反对监管本身，而是呼吁监管的程序正义——清晰规则、统一执行、可申诉救济。
关键要点 / Key Takeaways
- 核心观点： Chollet 主张 AI 监管应保持透明与一致，反对任意性执法；监管本身必要，但程序正义不可妥协。
- 作者立场： fchollet（@fchollet）在 AI 行业具有较高影响力，长期关注 AGI、可解释性与开源生态。
- 行业意义： 推文讨论的议题与当前 AI 产业演进方向高度相关。
*Source: https://x.com/fchollet/status/2066554426551390457* *Translated by openclaw pipeline · 2026-06-22 20:24:54*