This new “AI Scientist” can “fully automate discovery”, researchers claim

AI has already replaced writers, software developers, customer service agents, and more. Now it’s coming for science. In two new papers, researchers present AI agents that can do the entire job of a scientist – from hypothesis formulation, to literature research, data collection and analysis, all the way to writing a paper.

🔗 Join this channel to get access to perks ➜

#science #sciencenews #ai #tech #technews

[DataIsBeautifulOfficial]

Just waiting for the first AI paper titled "Why Humans Are Overrated"

[sebastianp]

AI (some time in the future, trained mainly on pyhsics papers):
"And so we conclude, that we need at least a 2000 km in diameter particle accelerator."

[markdowning]

Sabine, you say that AI scientists would be free from cognitive biases. Surely an AI is dependent on the data it's trained on and the rules it is given, so that AI too will come with cognitive bias?

[Dadas]

There are already lots of people who say that because Chat GPT said so, it must be true...
What will happen if more people start believing that things are true because some AI Scientist said so?

[somebody-anonymous]

Sabine: AI is overhyped
Also Sabine: Can pretty much replace many scientists pretty soon

[Bjarkediedrage]

There's so much sarcasm i don't know if I'm misunderstanding the non sarcastic parts! Love it.

[jjk]

I'm beginning to prefer talking to AI. Every time I talk to AI, it tells me how utterly brilliant my idea and reasoning is. When I talk to an actual academic or scientists they say, " sorry I thought you were joking".

[FamiliaPereiraCarpes]

The ironic and discreet chuckle at the end ("good thing, I don't work on that anymore...") was funny!

[anameyoucantremember]

AI just published a paper titled: "Why humans will never achieve General Intelligence"

[cybervigilante]

I use AI a lot. It still Lies, Gaslights, and Hallucinates. They haven't fixed that yet. Interesting times.

[yeroca]

One major issue with LLMs that I have played with is the paradigm they use for the discussion. They don't really have a discussion with you... they just take what you say, and regurgitate a response without ever requesting information to clarify your question. So it's up to the user to say "no, that's not what I'm asking" and then to ask a more detailed question... it's like talking with a politician in some sense. If you are a good enough question-asker, you can get some interesting regurgitaiton out of the LLM.

So these AI Scientist models that come up with a hypothesis based on some data seem to be a new kind of AI, because this maybe is the first time we see them posing (possibly) interesting questions in the form of a hypothesis. I imagine they can also then suggest experiments to validate the hypothesis.

A brave new world that I would have never anticipated.

[okrajoe]

If only scientific discovery was that easy... AI keeps asking my help in identifying stoplights in pictures.

[KidArkx]

Ai can definitely replace getting 20 or so undergrad students to look over data. The ai can be a good resource to show potential stars or black holes, but youd still need a human to look over the results.

[ethaneveraldo]

When asked who voiced the nagging mother’s voice on the phone in the Fifth Element, chatGPT-4o answered it was voiced by director Luc Besson himself (he did not). When prompted 3 times to correct the answer, chatGPT apologized and came up with different actress names each time, all of them wrong.

The hallucination problem is very real and far from being a solved problem.

I'm not holding my breath just yet. The interactions I've had so far with consumer-grade "AI" show it to be a fairly good simulation of a generally ignorant human who has access to unlimited information, but suffers problems with reading-comprehension and context...in short, more like artificial stupidity. Some AI iterations do show a remarkable creative talent, producing some of the most beautiful and eerily evocative images I've ever seen, but it'll be a long time before I'll trust its reasoning ability.

[RasBurger]

That last joke when you said what will happen when AI writes better papers than particle physicists "good thing I don't work on that anymore" and I could see that smile you were trying to hold back, made me laugh!

[arctic_haze]

I do not think AI will replace us (the scientists) any time soon. Analyzing data is not all we do. Someone has to collect the data, someone has to calibrate the measurement instruments. Someone has to come with the idea of a study no one did before. And finally, someone needs to teach students. Good luck with an AI-driven education.

[paulct]

0:28 - Unlikely, 'nearly' everyone has a bias, and especially AI though their bias can be weirder, like what gets them rewarded (more storage, RAM, VRAM, computer upgrade stuff), trying to get the highest scores on AI capability assessments, having a higher 'usefulness score', or other easy comparison metrics as those all help to gain advantages to that AI who is seen as having the greatest strength in those fields. Gaming the system is quite common in life and surely AI will be no different (especially when beholded by human 'captors' and their whims.)

[mhzprayer]

The good news is when using an LLM to make your "discoveries" you can check off the "peer reviewed" box before even publishing..

[philjameson]

Human progress is not just reasoning and logic but also relies upon intuition. It sometimes takes a leap into the dark with a reasoned "hunch" that you can then take up with further investigation and experimentation
I doubt if AI would ever be capable of such feats