Sébastien Boisvert

ARC Prize 2024 Puzzle 0ca9ddb6 youtu.be/9d_mKu8mQqs?si… via @YouTube

Rich Sutton’s new path for AI | Approximately Correct Podcast youtu.be/NvfK1TkXmOQ?si… via @YouTube

The Bellman equation is all you need ! - en.wikipedia.org/wiki/Q-learning - en.wikipedia.org/wiki/Bellman_e… - nature.com/articles/natur… - storage.googleapis.com/deepmind-media… #artificialintelligence

Big Shoe Companies LIED TO US!🤯

Training with large minibatches is bad for your health. More importantly, it's bad for your test error. Friends dont let friends use minibatches larger than 32. arxiv.org/abs/1804.07612

Canada should revoke access to our free healthcare to non citizens If citizens can't get family doctors, why are we giving access to so called "asylum" seekers?

I moved the Dropout unit at the correct place.

All Time Warmest Barefoot Winter Boots – Zero Drop, Snow, & Waterproof anyasreviews.com/warmest-barefo… #barefoot

ARC prize progress on puzzle 3aa6fb7a arcprize.org/play?task=3aa6… github.com/sebhtml/Arc-Pr… youtube.com/watch?v=7kMemb… @arcprize #AI

@nath_simard I see. Thanks.

@sebhtml Not really, it's built in with CUDA, but it requires an external dependency if we want to use the C++ API when compiling (github.com/ROCm/rocWMMA).

Working on CubeCL is so satisfying! For a long time we only had WebGPU as our in-house backend for Burn, but in the last few months we added support for SPIR-V, @nvidia's CUDA and @AMD's Rocm/HIP.

@nath_simard Is enabling WMMA (Warp Matrix Multiply Accumulate) actually the hard part ?

The ROCm/HIP backend is extremely new (merged yesterday) and we're still working on enabling WMMA. We haven't faced any huge driver/compiler issues to date. It's better than my expectations, so that's great!

Learning to solve the example 1 of puzzle 3aa6fb7a in the ARC prize @arcprize dskernel.blogspot.com/2024/10/learni…

For the ARC price arcprize.org, today I have increased the number of layers from 4 to 8. To do this on the NVIDIA A40 GPU on runpod.io, I changed my batch_size from 1024 to 512 and changed num_layers from 4 to 8. #AI @arcprize

For the ARC price arcprize.org, today I fixed a bug that caused my neural network to not see all the input text of a sample because the context was 140 characters but the samples were 211 characters

You can tell the RL is done properly when the models cease to speak English in their chain of thought

I made a meme image. #AI #LLMs #annoyedbird

Vive la République, vive la France !

I am using @runpod to train my AI for the @arcprize I think that I should have eventually something working for puzzle 3aa6fb7a using pytorch and xformers and transformers. hehe #AI #puzzles

Rust is such a beautiful language for working with genomic data. Clean syntax, functional programming, and extremely fast.