Thomas Guiziou

So dvpt originally made for the private side will have direct implication for science as well. Google will not invest in transformers without return. But then scientist can do a lot with those. The underlying question is more, is that science and other good application of ai worth the "bad" usage?

But now it is not really about how precise the term LLM is, it is an anti-AI take. Which might have some merits but a different debate.
While there is a lot of grifters and overhype in the AI techbros. It is also important to recognize that AI is very pervasive 1/2

I mean you call how you want chat GPT and ESM but both have a transformer and are trained by masking. Wether the tokens are words or amino acids do not really change the underlying architecture.
The idea, that we could debate, is that there a "language" to "life" thus LLM for biology.

LLM are used in biology have more or less the same structure as chat GPT, and are trained on genomic sequences or protein sequences. Which are letters, so the use LLM is not out of place.
Then there is many different architecture for different applications, from diffusion model to Graph model.

On a side note what is a "modeled hypothesis" ?
It is in line with the overuse of "model" for a wide range of meanings (not specific to this paper). It would bring some clarity if the use-cases of "model" were restrained.
2/2

Statistical significance does not always equate biological significance, very clear paper arguing against the use of p-values and null hypothesis test. Instead agues that confidence interval should be used. 1/2
rupress.org/jcb/article/...

You are welcome ! I really enjoy your post on the sequencing space !!

Personally I think a better comparaison is the Boson of Higgs experimental detection which was an expected validation of the theory as the human genome project. 2/2

I really disagree with this comparaison. The Humane Genome Project is a great technological feat but not a new theorical paradigm as quantum theory was. The humane genome project did not change our understanding of biology. It was a confirmation of what we previously thought biology was. 1/2

Congratulations !!

I agree that we won't be able to fully trust those paper. But if the review is only done by a LLM, I don't mind too much doing some type of prompt injection.
The issue is more that the prompt injection benefits the authors

My main interrogation from this is: should we let AI peer review articles ?

I agree with most of the point here.
AF2 one of the best example of good prediction but no understanding. Most people think structure prediction is solved and don't see the point of structural biology anymore.

I don't understand why they are not kicking him out. Did the royal society explained why not ?

AlphaFold has said nothing about how protein fold and has not made progress on the folding problem. We can say that and still say that AF is super useful and a worth a nobel prize.

In 1890, the X and Y chromosomes were discovered. It was found that the men who were tested had 46 chromosomes, including an X and a Y, while women who were tested also had 46 chromosomes, including 2 X chromosomes.
So obviously the conclusion was that the Y chromosome defined masculinity.

Super interesting, thanks for the summary !

The "reproducibility crisis" in science constantly makes headlines. Repro efforts are often limited. What if you could assess reproducibility of an entire field?

That's what @brunolemaitre.bsky.social et al. have done. Fly immunity is highly replicable & offers lessons for #metascience

A 🧵 1/n

That is a shame, I learned a lot watching last year's workshop.

Thanks a lot for the tutorial, it is very well written and informative

Are you doing a flexible fusion ? If they don't really interact, you might not need to redesign the interface in the first place.

ProteinMPNN should work quite well if you have a structural model of the fusion.

(As a gen z) I think it is the case and it is the same with alcool. But i also think we should decorrelate the alcool/drugs consumption to the social isolation debate.

C'est déjà arrivé chez nous, un maire a eu sa maison brulé parce qu'un centre d'acceuil pour demandeur d'asile aller ouvrir dans sa ville.

The coming of age of de novo protein design - Nature Nature - The coming of age of de novo protein design

www.nature.com/articles/nat...
Otherwise the review/perspective is a good starting point for design : )

For the design part, the rosetta notebooks are a great starting point. rosettacommons.github.io/PyRosetta.no...
They can be run in colab. Overall they are a great starting to do protein design.

The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity Recent generations of frontier language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes…

I am not particularly surprised for those results. it again highlight the need for better benchmarks for deep learning models.
The timing of the paper is also quite interesting, after the sequence on the absence of the LLM Siri at the WWDC.
machinelearning.apple.com/research/ill...

A general substitution matrix for structural phylogenetics. Abstract. Sequence-based maximum likelihood (ML) phylogenetics is a widely used method for inferring evolutionary relationships, which has illuminated the

New paper from the lab from Sriram Garg in my group. We introduce a general substitution matrix for structural phylogenetics. I think this is a big deal, so read on below if you think deep history is important. academic.oup.com/mbe/advance-...

I got fooled by AI-for-science hype—here's what it taught me I used AI in my plasma physics research and it didn’t go the way I expected.

Going to the hospital because I broke my wrist smashing the endorse button:
www.understandingai.org/p/i-got-fool...