Tuhin Chakrabarty and Sam Rodriques reject claims that three recent Nature papers show AI systems fully replicating core scientific tasks and rendering scientists obsolete

@SGRodriques @sebkrier I hope you are right, but ten years at our current pace is a very long time.

AI tools should amplify human expertise, not dilute it. It sounds trivial & obvious. But it may not turn out that way if we don't distinguish how to use AI for doing science vs truly amplifying our expertise & training budding scientists. 3/

One major challenge is making sure newer generations of scientists do not completely lose pretty fundamental knowledge about how all the automated stuff actually works, even if most of that time they can just automate it. There will continue to be a need for deep expertise. 2/

@SGRodriques Been working on this for the past few months - I've already seen enough, to strongly agree with you.

Plenty of incremental and tedious steps to the scientific research process will be sped up via automation, but we are very far from human researchers being redundant.

@SGRodriques We will need to widely share the benefits, or face an anti-AI movement.

http://Aipledgeforhumanity.org

@SGRodriques The bottleneck is shifting from execution to discernment.

Science may become less about producing possibilities and more about recognizing which ones matter.

There is huge potential to improve both. But it's going to require careful restructuring of education, humility, honest reflection & critical thinking. 4/4

@SGRodriques 10 years is a very long time in AI world, and it seems unclear how difficult it is to automate (median) scientific taste, especially since there are gonna be huge economic and military incentives to do so

I think the deep cruxes (beyond academic self-interest and defending self-image) are whether one cares whether knowledge is human-made, and whether academic structures (as they are and as traditions) have intrinsic value.

way more coming soon!

@AndrewCurran_ @SGRodriques @sebkrier The avg fresh bachelor is cooked, the system does not sell scale of edu at economy rates (And accessibility; It's a debt humiliation ritual), so maxxing grad, doc, postdoc students isn't even institutionally existent right now. Most corpos hire grads+

@SGRodriques @razibkhan It doesnt matter if it replaces scientists because the goal is to cure diseases. Its incredibly selfish to keep a human in the loop just for the sake of preserving jobs if AI alone can cure diseases faster and thus save life. Please stop with this low iq obsession with jobs

@CDHughett @SGRodriques The bottleneck is data. Language models had an abundance of data just ready to be ingested. Biological data is stored in 1000s of niche formats, behind paywalls, behind log ins, on 1000s of project websites, separated from key metadata

As for checking references, that is at least something we can use AI to improve cheaply on top of our own ability... assuming our software can legitimately access the academic literature! As long as it is locked up, we should have lower hopes for human and AI truth-seeking.

@CDHughett @SGRodriques It's also not surprising that the weak link in the AI paper was the analysis agent, which needed significant oversight.... because language models make very poor statisticians.

@anshulkundaje This is my biggest worry. For us AI is wonderful, because we still have the deeper expertise from before, so we can really reap the benifits and (hopefully!) avoid the main pitfalls.

But what about people whose training will include those tools to start with? 1/

We need to rethink incentives for deep expertise. I think get the sense, having talked to a few, that this is a big unknown for a lot of trainees.

What do you optimize training for, depth in computation? Depth in a specific biological/physical domain or set of techniques? What will get automated first vs continue to provide value? I don't know that anyone has a good answer, and the current message is that generalists will be most valuable. I don't think thats universally true, but it certainly is hard to predict where to allocate training time.

@anshulkundaje Parts of PhD work can be replaced. I hear companies are recruiting fewer juniors. That makes sense in short term, but who will be turning into the deep-expertise person in 20 years?

2/

@SGRodriques Totally agree. Though to be clear we still need humans for their taste in software as well. The story of AI agents is one of augmentation and amplification of power and ambition, not of replacement.