The connection was lost. Trying to connect again.

Post by Thomas Kahle

Thomas Kahle @tomkalei@mathe.social

This is very interesting to read:

openathena.ai/blog/there-is-no

Progress in AI is made in post-training these days. Programmers and mathematicians have worked largely in the open (and were also cheap to hire) so that their behaviors could be used to train the machine.

The next years of this industry will probably consist of getting biologists, lawyers, etc. to also go through this. Create many scenarios and record how human experts act in that scenario.

Open Athena There is no posttraining data commons The LLM revolution ran on a pretraining commons that humanity built for other reasons. Posttraining consumes tasks, not documents, and no such commons exists. If the open models community wants...

· Public

Replies

Christian Meesters @rupdecat@fediscience.org

@tomkalei
A Nobel prize has already been awarded for the prediction of protein structures. At no time it was perfect nor did we learn anything from it regarding the nature of the problem.

And people were cheering nevertheless.

Where in the past careers were built upon crystallization projects or similar endeavours in NMR or EM, now people turn to AI. Yes, this has been tremendous progress. And we made progress since the prize. And yet, striving for verification became harder.

Which means the risk of faulty conclusions is ever increasing.

In other words: biologists are already there... and nobody seems to be bothered.

· Unlisted

Thomas Kahle @tomkalei@mathe.social

@rupdecat

Hmm. Maybe I’m just very ignorant but I still think of alphaFold as very different from LLMs. In terms of maths, more like an optimization problem rather than the general imitation of all human behavior.

But the research-sociological consequences you describe do sound very similar.

· Unlisted

Christian Meesters @rupdecat@fediscience.org

@tomkalei ah, yes.

I was trying to make a point about AI usage in general. And the impact on a culture of insufficient scepticism. (Or rather on a publish-or-perish culture where a prezi presentation has more impact that paper with equations: It is a long-established fact, that in bio* papers with equations are cited less often than other papers. And I witnessed a prezi presentation which gave a mathematically illiterate person a professorship.)

Speaking about LLMs I could rant about vibe coding in bioinformatics. Brings us plenty of slop applications. Yet, just lately I attended a symposium in medicine. There just is some overlap of what I do.

Anyway, it was AI here and AI there. Not only are people doing diagnostic predictions with their models (where back in my day, work was slower and with plenty of pitfalls, we were doing things analytically - which means, in the end we could interpret why which weights are associated with which term if we were lucky. Not so any more.) Also, LLMs are heavy in use in that sector: particularly with the hope of easing work in the clinical sector.

What really bothers me: it seems as from conceptualization to paper LLMs are all over the place. Already >> 80 % of all biomedical papers are written with the help of LLMs. This already is debatable.

Also, yes, people ignore "the maths" and resort to an LLM instead.

· Unlisted

Thomas Kahle @tomkalei@mathe.social

@rupdecat

Insufficient scepticism … yes. 😞

You say it _was_ AI here and there. Then what happened? How has the community adjusted?

Math is going through fundamental changes and people seem way too depressed and not sceptic enough.

And those who are sceptic, are too sceptic, some even fundamentalist and don’t want to see technology or even reality.

· Unlisted