Kosmos: An AI Scientist for Autonomous Discovery | Hacker News Reader

Kosmos: An AI Scientist for Autonomous Discovery

arxiv.org

60 pointsbelter20 comments

It seems like these "discoveries" are mostly "We provided a dataset and Kosmos found the same conclusion as the scientist." This is an advancement but those datasets are not random in any sense. They were created to support a specific hypothesis which lead to the shared conclusion of the scientists and Kosmos. Discovery 7 seems to be of a different flavor in that a novel conclusion was arrived at from existing data.

I really think this is not "Autonomous Discovery". There is so much thought and science behind deriving the hypothesis and determining what experiments to do that is not captured in what Kosmos demonstrated here. It is exciting to see the reasoning capabilities and look forward to next steps but at this point a bit oversold in my opinion.

The challenge with comparing AI to humans is that the bar keeps shifting up.

It’s pretty impressive that Kosmos can reproduce the conclusions that human scientists came to de novo. Especially when it does so much faster than a human.

If the goal is to accelerate scientific discovery, this is what success looks like.

The GP says this is helpful but not autonomous discovery, you then reply we're holding AI to increasing expectation (both highly debatable and the fault of AI hypers) and say this is success. They are not mutually exclusive, and actually converge on what many have promoted with little reception: this is a useful tool but no silver bullet.

How did I say that? This is really cool and advancing but calling it "Autonomous Discovery" is very bold claim and needs strong evidence. I don't see it. They could have claimed it differently but they chose those words and that is what I am commenting on.

You are changing what they claim from "Autonomous Discovery" to "Accelerating Scientific Discovery". I agree with the latter.

To me it seems like an exercise in semantics. As per Carl Sagan, you could argue that fully Autonomous Discovery requires that it first "create the universe". I read "Autonomous" there as "working independently without being closely directed", and from my reading, it had passed that bar.

> The challenge with comparing AI to humans is that the bar keeps shifting up.

Exactly. There is no standard, humans will adapt and find how to use AI as a tool, and the bar will never and should never be fixed.

The beauty of Turing's Test (which he strangely seemed to misunderstand) is that it is almost impossible to pass.

The thing you are gatekeeping seems to have been cracked some months earlier by a different group:

https://sakana.ai/ai-scientist-first-publication/

Okay well I am not commenting on what that group did so I don't see how I am gatekeeping anything.

I am looking at this paper and saying I don't see the Autonomous Discovery claim but I do see novel AI contributions to science.

Includes 4 novel discoveries:

https://lifearchitect.ai/asi/

this site both looks legit and gives off "ufo believer" vibes

It is unfortunate that accepting the existence of alien life, which it is hard to argue is not a statistical fact, gives off those vibes.

How exactly extraterretial life is a statistical fact?

How is it not? Other than extremely anthropocentrically?

There are at least 200 billion trillion stars in the universe that we are aware of. That is a number beyond our comprehension. Stars generate elements. Elements form molecules. Life is built on some of these molecules.

The statistical argument is basically:

Multiplying a number beyond our comprehension by an unknown probability >= 2

Right?

I would say the statistical argument is the null argument. To invalidate it you should instead need to come up with a reason why in a billion trillion structurally relevant constructions we must be the only one where life emerges.

I too always claim my position is the null argument.

Curious to see where these go. The world model, and other enhancements seem helpful but effectively become hard coded rules. We know that human specified rules generally underperform learned relationships (at least in previous ML work) so I wonder if we’ll get to a regime like I think we were in older AI booms where we bump up against the limitations of rules again.

this is a joke... to even call this a scientist is an insult.

Why? I don't recall hearing of any airline pilot who felt insulted by an Autopilot, or a cleaner insulted by a Roomba. People who see parts of their jobs replaced often have a range of feelings, but I don't see why insult would be one of them.

How much slop will it discover?

Reply on news.ycombinator.com