Paul Tune

Machine learning engineer @canva, occasional photographer, one-way time traveller.

[Paul Tune @ptuls]

@torchcompiled Coding agents are starting to get good around large codebases. I’m inclined to think it’s the processes and engineering culture around the codebase that have to change, and verification of code is now the bottleneck

[Paul Tune @ptuls]

I kinda miss the rawness of base models. Harder to use, sure, but more surprises. We’ve tamed all the tails to mediocrity

[Paul Tune @ptuls]

@torchcompiled which more than likely also inherit a whole bunch of practices outside the codebase itself: organisational inertia etc.

[Paul Tune @ptuls]

@torchcompiled probably way easier to speed up on a codebase around a model than to tack it on to work in exisiting codebases

[Paul Tune @ptuls]

finally, a model with taste

[Josef Chen @josefchen]

Launching our new paper on arXiv: we trained the largest multilingual food model ever built.

4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.

All of human cooking compressed into 2 megabytes.

[Paul Tune @ptuls]

The coding agents are more high agency than some people I know. They just never give up

[Paul Tune @ptuls]

@cargoshortdad64 school needs to reorient around teaching students to ask questions effectively. it’s a (meta)skill that’s missing from current education

[Paul Tune @ptuls]

@cargoshortdad64 my conversations have so much redundancy that shannon would be rolling in his grave

[Paul Tune @ptuls]

@cargoshortdad64 I approached it differently by using gumbel noise, which you can formulate as the continuous analogue to the absorbing state discrete diffusion process. An argmax operation converts it back to discrete tokens from logit space, and used the gumbel max trick to train it

[Paul Tune @ptuls]

@cargoshortdad64 yes, on that note, I agree. thankfully it bodes well for the rest of us gpu poors that constraints do unlock creativity

[Paul Tune @ptuls]

The Jensen Huang leather jacket collection

[Paul Tune @ptuls]

Trumpian nightmare

[Paul Tune @ptuls]

definitely better adherence to the ending

[Paul Tune @ptuls]

Not even a year later, and OpenAI Image 2 gave us a step jump

https://t.co/8LaEHLhQfp

[Paul Tune @ptuls]

We're getting there, we should be able to go from short story to comic

[Paul Tune @ptuls]

they shipped so fast that competitors get advanced features today

[Chaofan Shou @Fried_rice]

Claude code source code has been leaked via a map file in their npm registry!

Code: https://t.co/jBiMoOzt8G

[Paul Tune @ptuls]

@cargoshortdad64 mate, aussie taxes might be high, but it's not labyrinthine like american taxes

[Paul Tune @ptuls]

https://t.co/lTK1WFT9kC

[Paul Tune @ptuls]

This, and Maverick