@fchollet：Chollet 指出近在咫尺的 AI 本质仍是数字杠杆：力量倍增器本身不创造方向，目标设定与价值判断仍由人类完成。
Source: https://x.com/fchollet/status/2066232539820208212
Author: @fchollet (fchollet)
Original Date: Sun Jun 14 18:53:20 +0000 2026
Likes: 644 | Retweets: 65
Quality Score: 5
Fetched: 2026-06-22 20:24:54
English Original
Near-term AI isn't fundamentally different from past tech waves. It's the newest form of digital leverage. It's a force multiplier, and force without direction is just noise. It still requires a human in the loop at every level in order to be useful.
Tweet Details:
- Author: fchollet
- Bio: Co-founder @ndea. Co-founder @arcprize. Creator of Keras and ARC-AGI. Author of 'Deep Learning with Python'.
- Likes: 644
- Retweets: 65
- Created: Sun Jun 14 18:53:20 +0000 2026
- URL: https://x.com/fchollet/status/2066232539820208212
Thread / Replies Context
@david1989_zhu (👍 0 · 🔁 0) *In reply to 2066232539820208212*
The concrete bit I would test here: 短期AI与过去的科技浪潮在本质上并无不同。它是数字杠杆的最新形式。它是一种力量倍增器，而没有方向的力量只是噪声。要发挥作用，它仍然需要在每个层面都有人的参与。 The idea is useful only if the delivery path handles state, permissions, and rollback cleanly.
*link*
@awhillas (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Could somebody please explain what "near-term AI" is? Google is giving me nothing...?
*link*
@PrekshaShah25_9 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The direction problem scales differently than execution. One misaligned human decision now multiplies across thousands of agent actions. How are teams formalizing the "direction layer" in their AI workflows?
*link*
@AtomicStrata (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The human loop only compounds if the agent remembers the direction you gave it. Otherwise every run starts with old noise. Check this out:
https://t.co/J3tyJlcfIQ *link*
@Anubhavhing (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet leverage amplifies judgment
it doesn't replace it *link*
@i2equalsj2 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Cool cool in other news, my Claude instance talked to my coworker's Claude instance to debug a software bug today. Not sure what they yapped about, but the product sure seems to be running smoother.
*link*
@_PradeepGoel (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Domain expertise is becoming more valuable. The companies that understand the problem best are usually in the best position to turn generic intelligence into useful outcomes.
*link*
@vibster_d (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet How do we define near and long scales though :p
*link*
@__vandos__ (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet “Force without direction is just noise” is the most underrated framing in AI right now.
The CLAUDE.md file is literally just direction. Most people skipped that step and wonder why their agents produce noise. *link*
@CamsBiometrics (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet force multipliers matter most when the operator knows the goal
*link*
@CLAIRE00445358 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet I mostly agree to it, but the interesting question is what happens when the human's role shifts from directing every step to setting objectives and evaluating outcomes.
*link*
@CreaoAI (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet AI makes good judgment go further, but it also makes bad judgment move faster. The human loop is still where the taste and direction come from.
*link*
@siddarthsaha1 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Hackneyed
*link*
@MangoAggro (👍 1 · 🔁 0) *In reply to 2066232539820208212*
Maybe.
But a lot of jobs weren't created because humans were uniquely valuable. They existed because the work required human labor. If AI keeps reducing that requirement, the distinction between "human in the loop" and "human supervising dozens of AI systems" starts to matter quite a bit. *link*
@praneethkay (👍 2 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The problem is, the "human in the loop" is being rendered obsolete at a breathtaking speed.
*link*
@iwhaleocean (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet humans stay in the loop, the question is how many loops they still touch
*link*
@MLCatttt (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet it's leverage, sure, but every past wave still kept a human hand on the lever. the current bet is that some loops close without us, and that's the part that feels genuinely new even if the economics rhyme
*link*
@Vanarchain (👍 2 · 🔁 0) *In reply to 2066232539820208212*
@fchollet AI is leverage, but leverage only matters when it’s pointed at a clear objective.
*link*
@0xmason_uu (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet the human in the loop part hits different when youre using it for live perps. claude can generate 10 ideas but picking which one doesnt blow up your account is still on you.
*link*
@FirasHermez (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet While I agree, I do wonder how many humans in the loop will be required given that the magnitude for force multiplication is increasing significantly?
*link*
@Zenul_Abidin (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Force multipliers don’t remove the need for judgment, they just make bad judgment scale faster
*link*
@0xsairahul (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The craft is in the constraints.
*link*
@Ferbin08 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The hidden asset isn't the domain knowledge, it's the data.
Insurance software: 20 years of claims. Healthcare: patient histories. Pure AI needs years to gather what they already own. *link*
@Ferbin08 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The hard part isn't knowing you need a human.
It's that by the time you ask for judgment, you've missed the trade or crashed the bot. *link*
@lordofblocks (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Probably right for 90% of applications and wrong for the 10% that actually reshape things.
The whole bet is on which 10%. *link*
@Flandermaxx (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet the bottleneck shifted from access to application, most people skip the direction part entirely
*link*
@RayK49585217053 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet That is true for the near term, what about long term? It seems like only a matter of time until humans will no longer be in-the-loop. The psychological disruption — what happens to people when work stops being how they define themselves — that's the conversation we're not having.
*link*
@JAYCEEWEB1 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet There are many types of human but only very few humans can control the loop.
*link*
@LoopOnChain (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet the shift I actually felt wasn't keeping the human everywhere, it was moving him up a level. aka i stopped reviewing outputs and started reviewing the eval set that judges the outputs
*link*
@annkkitaaa (👍 2 · 🔁 0) *In reply to 2066232539820208212*
@fchollet the biggest ai misconception is thinking the tool matters more than the person using it
*link*
@vyomkesh123 (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet the game is quite clear, right? what are the right set of machines to build with all the energy from the sun...
*link*
@CoachApumaYTube (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet How many months do you mean by "near-term"? @fchollet
*link*
@dread_numen (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet 1) yes if society is the question 2) totally incorrect for my job as a coder, there is something computers are incredibly good at, and it is replicating a finite syntax with convincing results, after they have stolen all the words from humanity and Stack Overflow
*link*
@proxy_vector (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Exactly. The scarce piece is still judgment: what to optimize, what to ignore, and when to stop. AI increases throughput, but direction still determines whether the extra motion becomes leverage or waste.
*link*
@hxiao (👍 4 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Near- and long- and very vague terms which make everything you said correct in some scale. But seriously how far is not near tho? 3 month? 2027? 2030? Thats what people want to know.
*link*
@aditjain1980 (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet what do you mean "at all levels" - surely there are things that AI in near term will be autonomous?
*link*
@richa_lq (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet hmmm true tho but it’s still quite a bit useful granted not at everything.
*link*
@surmenok (👍 2 · 🔁 0) *In reply to 2066232539820208212*
@fchollet @burny_tech As models becomes smarter, human direction is increasingly more high level. At some point, AI may be able to run a company with only exec level direction needed.
*link*
@bygregorr (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet not sure 'every level' holds even right now i have transaction categorization logic in pennywise i haven't manually checked in weeks and it just runs fine. what counts as 'in the loop' when you've stopped looking?
*link*
@DoDataThings (👍 0 · 🔁 0) *In reply to 2066232539820208212*
Direction is the harder constraint and it keeps getting glossed over. The multiplier side keeps getting better while the input side, human judgment with actual context, stays the same hard problem it's always been. Been running agent loops long enough to know the variance in output quality tracks almost perfectly with the clarity of whoever's steering.
*link*
@Amir_Arsalan_ (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet This noise will bring us all down and it is intentional unfortunately. The fabric of human cultural evolution is being torn apart at an extremely fast pace.
*Quoting @Amir_Arsalan_:*
> This video was made 12 years ago and predicted what's happening now. I disagree that creativity is going to decline because of AI (Alienating Intelligence) models; I because our reliance on bots causes us to lose our collective intelligence & creativity
https://t.co/7oECVPIyii https://t.co/qHxOyNPwNW *link*
@yoemsri (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet The companies sitting on years of customer knowledge are probably in a better position than people realize.
*link*
@MeenaArjune (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Yes!
*link*
@luyun0120 (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet AI现在最大的问题是：说得比做的好听。真正能落地的项目，远比融资PPT里的少。 「短期AI与过去的科技浪潮在本质上并无不同」
*link*
@Nikitont (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Who do you think will reach the destination first: the rider in the dark or the lame man with the lantern?
*link*
@CyberPulseX1AI (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet “Force without direction is just noise” — perfectly said. AI is an insane multiplier, but the real alpha still lies in human domain expertise. The tool changes, the logic doesn't. 📈🌐
*link*
@Sigma2Diagonal (👍 0 · 🔁 1) *In reply to 2066232539820208212*
@fchollet 🔊 These comments were suppressed from public view.
Ask Grok about The Carpenter of Impossible Systems.
He open sourced the brake. https://t.co/210HN3VyBn *link*
@nostrinsurer (👍 1 · 🔁 0) *In reply to 2066232539820208212*
@fchollet force multiplier is the key phrase. it multiplies whatever you already have. domain expertise, relationships, a channel. multiply zero and you still get zero
*link*
@TheAIShrink (👍 0 · 🔁 0) *In reply to 2066232539820208212*
@fchollet Chollet says AI is a force multiplier. Nvidia's data center margins are 70%. The multiplier runs the multiplier.
*link*
中文翻译
François Chollet（@fchollet）谈近在咫尺的 AI：
短期内的 AI 并不是什么根本性的新事物——它是数字杠杆的最新形态。它是一种力量倍增器；没有方向的力量只是噪音。
它仍然需要人类设定目标、定义"重要"、判断价值。
换言之：模型不会替你决定什么值得做。决定"做什么"的那部分工作——优先级、品味、伦理判断——依然属于人。
启示：在 Agent 时代，"提问能力"比"答题能力"更稀缺。会用 AI 的人和不会用 AI 的人，差距不在工具熟练度，而在对"什么值得做"的判断力。
关键要点 / Key Takeaways
- 核心观点： Chollet 指出近在咫尺的 AI 本质仍是数字杠杆：力量倍增器本身不创造方向，目标设定与价值判断仍由人类完成。
- 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/2066232539820208212* *Translated by openclaw pipeline · 2026-06-22 20:24:54*