François Chollet
@fchollet
Followers
568K
Following
10K
Media
1K
Statuses
24K
Co-founder @ndea. Co-founder @arcprize. Creator of Keras and ARC-AGI. Author of 'Deep Learning with Python'.
United States
Joined August 2009
Big news: we just released Keras 3.0!. ▶ Run Keras on top of JAX, TensorFlow, and PyTorch.▶ Train faster with XLA compilation.▶ Unlock training runs with any number of devices & hosts via the new Keras distribution API. It's live on PyPI now! 🚀
169
730
3K
AI (not LLMs, but search based AI) will be superhuman at the first type relatively quickly.
3
2
69
In general there are two different kinds of methodology to produce progress in any science or engineering field. both are important and can lead to transformative progress. There's the "Edison way" where you brute-force a large predefined design space and you keep what works,.
This is an unwise statement that can only make people confused about what LLMs can or cannot do. Let me tell you something: Math is NOT about solving this kind of ad hoc optimization problems. Yeah, by scraping available data and then clustering it, LLMs can sometimes solve some.
23
44
376
Back in 2017 my book had an entire chapter on generative AI, including language modeling and image generation. I wrote that content in 2016 and early 2017. This was some of the earliest textbook content that covered generative AI. All the way back in 2014 I was convinced that AI
9
1
166
I have been consistently bullish on deep learning since 2013, back when deep learning was maybe a couple thousands of people. I have also been consistently bullish on scaling DL -- not as a way to achieve AGI, but as a way to create more useful models.
4
5
147
People also ask, "didn't you say in 2023 that LLMs could not reason?". I have also never said this. I am on the record across many channels (Twitter, podcasts. ) saying that "can LLMs reason?" was not a relevant question, just semantics, and that the more interesting question.
6
5
149
People ask me, "didn't you say before ChatGPT that deep learning had hit a wall and there would be no more progress?". I have never said this. I was saying the opposite (that scaling DL would deliver). You might be thinking of Gary Marcus. My pre-ChatGPT position (below) was.
Two perfectly compatible messages I've been repeating for years:. 1. Scaling up deep learning will keep paying off. 2. Scaling up deep learning won't lead to AGI, because deep learning on its own is missing key properties required for general intelligence.
30
67
991
By the way, I don't know if people realize this, but the 2020 work-from-home switch coincided with a major productivity boom, and the late 2021 and 2022 back-to-office reversal coincided with a noticeable productivity drop. It's right there in the statistics. Narrative.
51
147
2K
RT @PardonMyTake: Tuesday night max woke Big Cat up with a flashlight at 2am because he thought we were going to get sued. @forthepeople ht….
0
10
0
LLM adoption among US workers is closing in on 50%. Meanwhile labor productivity growth is lower than in 2020. Many counter-arguments can be made here, e.g. "they don't know yet how to be productive with it, they've only been using for 1-2 years", "50% is still too low to see.
LLM adoption rose to 45.9% among US workers as of June/July 2025, according to a Stanford/World Bank survey. Inference demand will continue to surge, not just from more users and more usage per user, but as newer, more advanced GenAI models require far more inference compute.
340
567
5K
JAX on GPU is basically as good, actually.
Jax on TPU will solve all your problems.
9
30
403
Being pro-technology doesn't mean being blind to the negative effects of new technology. It's an exercise in pragmatic optimism -- maximizing the upside while managing the downside.
55
135
1K
RT @arcprize: Analyzing the Hierarchical Reasoning Model by @makingAGI. We verified scores on hidden tasks, ran ablations, and found that p….
0
196
0
Importantly: there was in fact no data leakage at work in the codebase, unless what a first read of the code suggested. The model is trained on the *demonstration pairs* of the evaluation tasks, but never sees the *test pairs* of those tasks, which is what it gets tested on.
1
6
275
We were able to reproduce the strong findings of the HRM paper on ARC-AGI-1. Further, we ran a series of ablation experiments to get to the bottom of what's behind it. Key findings:. 1. The HRM model architecture itself (the centerpiece of the paper) is not an important factor.
45
290
3K
Interesting findings from this post:. 1. It should be obvious to anyone who has interacted with LLMs before that the writing style of the tweet is a conspicuous caricature of AI slop (e.g. em dashes, the "it's not. it's. " construction, rambling, florid prose, etc.). Yet, many.
45
22
513
Crypto is the most profitable asset class for traders. But it's maturing fast and the edge won't last forever. Here’s how you can build, test, and deploy systematic portfolios that survive every regime:
145
202
2K