Twitxy Saved

Mathys

Verified account This account is verified. Learn more

@GrapotteM

Founded and building Nextflow Frontier at Seqera
Barcelona nextflow.com/frontier Joined February 2019
368 Following 53 Followers 100 Posts

Latest posts 5

Mathys Verified account @GrapotteM
Definitely an interesting xAI method to try
Daniel Murfet Verified account @danielmurfet Aug 5, 2025
Neural networks are grown, not programmed. What does that growth process look like? Like this!

This is a small language model (3M) across training, visualised with a new interpretability technique: susceptibilities. We call this handsome critter the rainbow serpent.
View quoted post on X
Mathys Verified account @GrapotteM
For people coming from pure dl, I will try to vulgarize the bio part, a 馃У
Arc Institute Verified account @arcinstitute Jun 26, 2025
Register today for the Virtual Cell Challenge and use AI to solve one of biology鈥檚 most complex problems.

Announced in @CellCellPress, the competition is hosted by Arc Institute and sponsored by @nvidia, @10xGenomics, and @UltimaGenomics.
View quoted post on X
Mathys Verified account @GrapotteM
I am on board with this, but still remain optimistic regarding the impact of what has been done in this paper.

Specifically excited about what comes out of figure S6.B i.e. "if model is good at next token prediction in arbitrary sequence, then mistakes are impactful variants"
color me skeptical. Actually, given the incredibly broad and bold claims in this paper, color me *extremely* skeptical
View quoted post on X
Mathys Verified account @GrapotteM
This is such a lifesaver!
Is there a way to access the ressourceLimit values in the nextflow process ? (for instance, to detect if we should run our python code with GPU or CPU enabled?)
1/ The @nextflowio community has long sought an effective solution to address scenarios where resource requests exceed the capacity of the infrastructure. Some of you may be familiar with the check_max() Groovy function developed by the @nf_core community to tackle this challenge
View quoted post on X
Mathys Verified account @GrapotteM
This is why we are building nf-core/deepmodeloptim this way and why nextflow is a fantastic piece of software for this.
Jack Morris Verified account @jxmnop Nov 4, 2024
i decided you don't need lots of GPUs to do good research when i read "Are Emergent Abilities of Large Language Models a Mirage?" (won award at NeurIPS 2023)

if you buy that loss scales smoothly and predictably from small to large scale (both in model parameters and sequence length) then you should be able to validate and test most hypotheses on small models with small sequences

both my most recent successful research results (language model inversion work and contextual embeddings) came from scaling down and running hundreds of tiny experiments with a tight, short feedback loop

if you're thoughtful, clever, and persistent, you can get good research results on a single GPU
View quoted post on X