J'ai eu Guerres oubliées à Noël, et j'ai adoré. Très agréable à lire, encore plus pour moi care je peux lire avec vos voix. Les chapitres sont très accessibles et c'est une bonne introduction aux livres d'histoires militaires ! Merci à vous deux : )
Est ce que l'ia utilisé par les militiaires sont uniquement des llms ? J'ai l'impression que d'autre architecture serait bcp plus efficace dans certains cas
And if it is done by a LLM/agent, so much fine understanding is lost. Which is useful to produce good results.
On the other side, if a experience researcher can produce results in less time why not use it.
I did no such things you are either misunderstanding on purpose or not interest in understanding . Anyway that is the extend I can suffer this
Then AIs are such a detriment to learning that would advice early career researchers to not used or to a limit. There is so much that increase understanding and is helpful later in code detail, in how to write or how to analyse a SDS PAGE for example.
I will answer here to most points : )
To me process and the results are the same thing. If one want great and consistant results he should obsess over the process. Otherwise results can come but it is less likely and less consistent.
how is that close to the original point that ai solved protein folding ?
I am quite astonish how you can assert truth about field that from your own accord you don't know very much.
Did human scientist place one by one the coordinates of the protein for CASP ? They had tools, just like AF is a tool.
I said before that it is not AF that produce the science it just produce data.
And even if AI could do the job of the scientist, as a society is that something that we should desire ? Just like art, food, books, I find science beautiful because human made it. I would not care at all about science produced by AIs.
What we want is sometimes the extremum of the distribution and sometimes completely outside of the distribution so LLMs cannot generate that?
That original point is untrue. Even for alphafold, it is not the ai which is doing discovery. The model on its own it worthless. It is the researcher who give meaning to the results. And LLM are particularly ill-suited to do discovery are they give you the mean of the distribution.
that the protein explore. All of that without evolutionary information encoded in an MSA or in the weights of a LLM. AI models are still far from being able to do that.
While alphafold was a big advancement in structure prediction it did not solve a the folding problem. AF is capable of one folded structure if provided with the evolutionary informations. Solving the folding problem would mean going from the sequence to the structure and the different conformations
You can have the some dataframe vizualization with data wrangler (vscode plugin) in jupyter. But it does not seems nearly as nice as the marimo's : )
I should test, it might fit nicely with my use of jupyter. The improved reproducibility seems like a big win
is the change from jupyter to marimo worth it ?
My reason for a very limited use of LLMs/AI is that I learn very little while using them. They make me more productive yes but that productivity come at the cost of understanding and getting better at the task. And coding is a great examples of that .
I followed with interest your tribulations with AI agents. While it increased your productivity, does it increased your understanding of the biological question ?
With AI we could produce a lot of new models and new papers, but how much would it increase our understanding of biology ?
I cannot grasp how one can trust openai with important data, or any llm for that case. How many story of delete code base or database do we need ?
on préfère macron en empereur ou Le pen en président ?
isn't that libre office ?
how when the UK depends on the us for they nuclear missiles ?
what about defense ? The main issue with Trump is that the US are the only security garantee for EU contries.
le cout économique n'est pas suffisant pour empêcher une guerre, cf l'invasion de l'Ukraine. Sachant la position de la chine par rapport à Taiwan le coté économique est vraiment secondaire.
How much do you trust this kinda of software that will literally (from what I understand) read your entire computer ?
that being said, I would still use bindcraft for binder design
I might sound picky, but to me 24h of compute of BindCraft and 24h of compute of RFdiffusion are not equal. With same GPU memory I can run parallel RFdiffusion runs and sample much more which might help. So I don't know how much the comparaison holds.
I fully agree with that :)
What part of Bindcraft outperform Rfdiffusion-PMPNN ?
The 4-stage sampling protocol or the filtering process ?
To me it is quite hard, with the current data, to know what part of Bindcraft make for the success and then conclude more to hallucination vs diffusion.
