Hacker News Reader

Comment by augment_me | Hacker News Reader

I am currently generating about 10 papers per month for marginal increments/architectural additions that are autoreviewed in a loop by Claude + Codex until those tools don't spot any major issues and then sending them in to various conferences. Since more than half of the reviewers are just using these tools to review, it's really easy to get a good score. This year I got into ICLR, ICML, NeurIps, SIGGRAPH and ACL with 2-authors papers (me and my supervisor).

I feel completely disillusioned with academia, this is apparently the "reward" that the funder wants. In the end I have time to do whatever I want because this business it running itself.

View on HN

Yes, ICLR submissions have gone 11k to 20k to 60k in 2025, 26, 27.

Most reviewers hate reviewing, don't get rewarded for it, the papers they reward are usually bad (as acceptance rate is around 20-25%), so they are forced to do it via reciprocal reviewing rules where submitter must also review, but they review with AI, the area chairs write their meta review with AI, they (area chairs) can arbitrarily reject a paper even with "accept" reviews, delaying people's graduation by half a year or more. Also, meanwhile much of the research in eg AI ML application areas where you tweak the architecture and beat a benchmark (vast majority of the tens of thousands of submissions) can now be done better with AI.

Yes there are elite academics who do deep foundational thoughtful stuff but the system is totally subverted because most participants don't give a shit and don't feel any long term allegiance to the "research community". They take what they can, and are out very soon. Of course this means the signal and prestige of papers on your CV will tank. And having too many may have the opposite effect l already today.

> Also, meanwhile much of the research in eg AI ML application areas where you tweak the architecture and beat a benchmark (vast majority of the tens of thousands of submissions) can now be done better with AI.

This is an underappreciated point. I was telling someone in the lab the other day after we saw our ICLR submission ID... that it more so revealed how much existing AI/ML research was very nearly slop.

I think when push comes to shove, academics will have to admit that this volume of papers even pre-AI just wasn't all about novel groundbreaking work. Now, mathematicians are starting to admit that oh the theorems and proofs don't actually matter, they are just steppings stones and learning opportunities for PhD students to hone their skills, and to form a community around etc.

Same way, most published papers are much more about a junior researcher gaining experience, learning how to speak clearly in a talk, how to summarize work in poster format, how to write, how to make figures, etc. etc. even if the final result is meh, they kind of got initiated into all this song and dance and that's what the academics really want to do, to pass on their culture basically. When this experience actually gets "truly" deployed is a good question though, because once you reach that level, you basically become a manager and all your later papers will be again written by the juniors honing their skills that you hired, because you no longer have time to do hands-on research once you were actually starting to become good at it, because now you just have meetings, write grant applications, sit on committees, teach courses, go to faculty meetings, help co-organize conferences etc. etc.

I'm convinced that engineering centers where after a point, you head a center for something something raise some money for it and it is much better than typical late-career academia work. One of my advisors does this and they enjoy it a lot. It's like a academia-startup hybrid. One other is a HoD and while they like it, it's more like what you describe.

Anyways the point is that, students working in that long term goal oriented project oriented setting is in many ways nicee than the typical "paper mill" approach. They still do get all the skills you mentioned since publishing something by yourself is mandatory still, but they spend a lot of time on practical problems.

I'm not sure about your goals, and whether your comment is ragebait, but typically this won't help you with academic careers or industrial positions where people know academia eg because they themselves have done PhDs. If you are the sole author of dozens of papers a year, it's a huge red flag for most good labs. It's just not plausible. And if they suspect you did it with AI, which they increasingly will, why would they need to hire you instead of just paying for AI subscriptions? Why the need for the middleman?

I think AI can be very effective in research but spending more effort on making one or two really great papers per year gets you better career results than 10 mediocre ones per month.

LOL I love this and also hate it. You are the best and also the worst. I am actually laughing :D Who cares about deepening our understanding? Curiosity? Hoping to understand some deeper connections/meaning? Let the AI write 10 papers to top conferences per year, and be done with it. If this is what people think science is, that's what we should do. Incentives -> outcomes, in the purest form. Loving it. And hating it!

Reply on news.ycombinator.com