[kylebenzle]

This is such an amazing insight and hits the nail right on the head (for why the above project and 99% of "AI" fears are nonsense).

Reading any scientific paper usually takes me about 1 day, if I actually want to understand it. I've been in my field a decade but still, to read one paper usually means reading AT LEAST one other paper along the way, but I don't know which of the 100s of citations I will need until I understand what I don't understand, AI can't do that for me.

AI is like the crypto hype but for the HN crowd, except with basically no real world use cases.

[robbomacrae]

Do you think it is completely out of reach for the AI to follow those rabbit holes automatically and tie in the useful information? Could it not also be personalized to the users knowledge of the subject?

I'm actively working on the first problem. The second is in my todo list.

[godelski]

Currently? Yes. This is a challenging problem for someone with decades of experience. I'm not sure you can train an LLM to appropriately do this because I can't even begin to describe how one would generate an adequate cost function. I don't think even RLHF can resolve that aspect because the truth of the matter is that I don't know what's important in that rabbit hole until I spend time working on the problem, replicating, or have sufficient experience. All too common a single line can make or break an algorithm and that line is 3 papers back. All too common there's nuances that radically change results that aren't even in the papers themselves.

I hope you succeed, but personally I don't know how this could be solved. The problem is that I don't actually need better summarization, its that I need more nuance and technical aspects. The problem exists because we're writing to larger audiences as competition increases and the quality of reviewing decreases (we even have a shortage which only exacerbates this problem). I'm not sure AI solves existential problems that are built around reward hacking, in fact everything I've seen suggests they explicitly do the opposite. I mean we literally train them to do that...

[godelski]

As another ML researcher I'll second this. When I review papers for conferences it takes me hours to review a work that's in my niche. It's because you can't skim a paper unless it is exceptionally good (lol) or has major flaws. A single sentence often holds the magic keys to making an algorithm work, and I don't think many would realize this unless they try to reproduce works from the paper alone (not using lucidrains or the official implementation). Even lucidrains makes some of these mistakes. And yes, even in my own niche, I go back and reread papers that are key references to make sure I didn't forget key details and understand the exact limitations the authors are addressing. The main thing is that the closer it is to my exact same niche the fewer reference papers I have to read (because I know which ones matter) and the faster I can read those papers. It's ensuring I am not forgetting key nuances of specific datasets or specific metrics that are used, because not keeping these in mind will trick me into wrong conclusions. This is what's required if I want to give a quality review. It's what's required if I see myself as on the same team as authors (team science) and helping them make the best work they can.

But I'll admit that there's a lot of pressure for me to stop doing this. A big part being that it's very clear my reviewers are not prescribing to this tactic. Rather I think many reviewers are not concerned with the rigor of their reviews. That they do not see themselves on the same team but rather antagonistic (team conference/team journal) and that their job is to filter. But I think an issue is that in ML you get an advantage if you are reject heavy and lazy in reviewing. Not only do you save on the time it takes to review but since it is a zero sum game you ever so slightly increase the odds of your own work being accepted. Honestly I do not feel the process is very scientific. Even the new CVPR LLM rules are a joke. More signaling than solutions. I just wonder if people care about the science anymore.

[hackideiomat]

Have you ever used a summary service? Have you ever build or even done something with ChatGPT or competitors?

There's tons of real world uses and you're falling behind if you think there aint.