If ARC-AGI required visual perception, you'd see VLMs outperform strict LLMs — by a lot. Everyone tried VLMs during the 2024 competition — no one got better results. Every single top entry used a strict LLM. As we've said many times: ARC-AGI is a 2D symbolic reasoning
@fchollet
-
Comparing compute efficiency with o3 low-compute version
By
–
It's in the same ballpark as the low-compute version of the o3 entry — so way less than the high-compute version
-
Jeremy Berman open-sources ARC-AGI solution with LLM genetic programming
By
–
Jeremy Berman, former #1 on the ARC-AGI-Pub leaderboard, just open-sourced his solution as a template on Params (link in next tweet) It uses LLM-driven genetic programming, which has turned out to be way more powerful than anybody expected. You can book a consulting call with
-
Scalability Requires Simplicity, Efficiency, and Strong Generalization
By
–
So you should certainly work on scalable AI ideas (which generally implies *simple* ideas). But you should also pay close attention to efficiency & generalization characteristics of your ideas. Data & compute on their own aren't going to save your bad ideas from eventual failure.
-
Twitter’s Open-Source Algorithm Still Uses Keras Without Updates
By
–
When Twitter open-sourced "the algorithm" on GitHub, it used Keras. That was only one year ago and we've had no update since. https://
github.com/twitter/the-al
gorithm
… -
Chollet launches consulting profile on Params platform
By
–
I started a profile on Params, where you can book a consulting call with me to chat about recommender system plans: https://
params.com/@fchollet Params is a platform for open-source devs to publish project templates — project users can then book calls with template authors. -
KerasHub Generative Models Now JAX-Based with TPU Support
By
–
All of the generative models on KerasHub run on JAX and are TPU compatible, with model parallelism built-in, so check it out
-
TPU v5 Colab: JAX Keras ModelParallel for Full Core Utilization
By
–
TPU v5 in Colab! Best way to make full use of all the cores: JAX + Keras + keras.distribution.ModelParallel API
-
Humanity’s Computing Power Consumption Will Grow Every Year
By
–
The most reliably predictable trend of the next 100 years: every year, humanity will use significantly more computing power than the previous year. Someone should start an ETF based on that thesis (it's not just $NVDA and $AMD, it's cloud services, the data center industry,
-
Waiting for Public Model Release to Obtain Important Data Point
By
–
This would be a very valuable data point, but we didn't have access to the base model — I suppose we'll have the answer when the model goes public