Nathan

Training foundation models

[Nathan @naatyu0]

@ZipMind_AI @liquidai @MistralAI My approach is to send the screen to the decison model to know where he is but he seems to have no spatial awareness

[Nathan @naatyu0]

Wanted to try decisions models so I made D1 @liquidai destroy the minesweeper ! All built with @MistralAI large 4 (really strong model btw, but current inference speed is low 😢)
I tried to make D1 play pokemon but even with a vlm supervisor for planning it was unsuccessful.

[Nathan @naatyu0]

@ZipMind_AI @liquidai @MistralAI Plain decision can't even pass the starting screen. I added a vlm for short horizon planning like, enter the house, take the stairs but even with that the decision model couldn't properly move in the game.

[Nathan @naatyu0]

@resiliem Yes of course, the simplified visuals are awesome ! I'm a chrome tracing user, but looking to switch, perfetto seems cool.

[Nathan @naatyu0]

@Aurelien_L_ Really cool, I need to try it on pokemon

[Nathan @naatyu0]

@antoine_chaffin Really cool ! Will you publish maybe a blog post on how you trained it ?

[Nathan @naatyu0]

@helloiamleonie @adithya_s_k @huggingface Awesome work! Have you considered data-mixing strategies for harnesses? Weighting rollouts from different harnesses could maybe lead to some interesting results

[Nathan @naatyu0]

@jbhuang0604 Amazing video as always !

[Nathan @naatyu0]

@maximelabonne Wow, d1 is based on lfm 2.5 models ?

[Nathan @naatyu0]

@Aurelien_L_ Amazing ! Blog post incoming ?

[Nathan @naatyu0]

@rasbt I also saw that they did red teaming on their rl envs. This could also have helped producing better env and hence less reward hacking possible.

[Nathan @naatyu0]

@tugot17 What is this for ?

[Nathan @naatyu0]

@DlRMCfUXI8K0k62 @reach_vb Heavy in inference ? Due to the 2 loops it would make sense. I never tested it

[Nathan @naatyu0]

@vincentzed_cuda @stanfordnlp @tatsu_hashimoto Nice ! Lectures will be uploaded in september ?

[Nathan @naatyu0]

@AnthropicAI This seems highly related with the current trend of Astra bding looped model and people saying it will abstract the reasoning. With this you could possibly try to interpret what happen between layers and also between loops.

[Nathan @naatyu0]

@hkproj I Love the video, 2h30 through it. I saw that in arithmetic intensity you used sparse peak flops instead of dense and wondered if it was intentional?