Indeed the Neurips arms race is pathetic but very funny. And I have zero sympathy for any complainers; I just laugh. The solution is trivial, and takes two steps:
1) Just stop submitting, and disengage. That is what I did. It is very easy. I used to enjoy the conference (from about 1990-2015) because I liked talking to people there, but after 2019 I vowed never again (even now it is heading to Sydney because of work I did many years ago to get it there!). I have turned down every request to be a reviewer, area chair, senior area chair, and even senior program chair (a couple of years ago) and am delighted I did. And I don't submit there any more.
2) Stop judging people in terms of whether they are named as authors on papers published there (note my careful wording). Alas, my professorial colleagues (world-wide I mean) are the worst offenders here. This too is pathetic...
[An alternative (only slightly flippant) suggestion which I have made to a Neurips board member is to decouple the papers from the conference: by all means have the conference, for people to confer, but do not tie it to any publication; separately, have a publication venue (aka a journal). Demand submitters to pay handsomely for each submission. Use the vast profits to pay for nice food etc for the conference attendees who would then attend for the sake of conversations. The conference would no doubt shrink which would be a fine thing, and it might be fun to attend again...]
Regarding your other points, well Geoff Hinton reckoned there would be no radiologists any more (I am very glad he was wrong, having interacted with them to my considerable benefit). And other technological superstars are equally terrible at making predictions (there is a large literature on this, not just an opinion, but a well documented pervasive pattern ... look at all the refs in chapter 3 of https://acola.org/sites/default/files/2026-03/saf05-Technology%20and%20Australia's%20Future%20report-full-17sept%202015.pdf ). So when I read the latest predictions, I just recall the performance of prognosticators of times past.
My free advice: Say no to Neurips. And say no to stupid chatbots, and get on with life. Easy :-)
One thing we need to address is what junior people should do. They have more pressures, but they also have more agency then us older folks. If they go away, all of academia does too!
Indeed! Thinking of them was a primary motivation for expressing my opinion so bluntly. There are many things we can do: for example, stress the value of intrinsic rather than extrinsic rewards (well documented advantages are demonstrated by Deci and Ryan); declare our (old folks) disdain of judging by appearance in a venue (sign the DORA to signal your stance); encourage them to practice articulating what they are doing and why it matters (rather than relying on a venue where it shows up as a certificate of quality); encourage them to be proud of what they do, rather than what people say of it, and have the courage and gumption to take the long view; teach the value of good scientific conversations, so that when they go to conferences, they can make the most of the experience rather than trying to impress; decouple funding approval for conference attendance from having a paper there (the normal scheme implicitly says to young folks that the reason for going is only to transmit rather than to receive); explain to them that many of the "pressures" are self-inflicted (e.g. anxious and tawdry comparisons with others); telling them that the bar is low -- they can easily do better than our generation (because they have seen our collective mistakes); and leading by example in not being a complete ranker.
It seems that writing was easier to automate than reading - or, what Tao, in a strange gastrointestinal metaphor calls (proof) digestion - I wonder why that is...
Is this true? LLMs can automate both "writing" (generation?) and "reading" (verification?) under some definition right now, but for a human to understand the output of any automated writing they obviously have to read it themselves
Feels like a cargo cult to me! Math is not about proving theorems or writing papers. We don’t stop running because cars are faster. Isn’t it all about the joy of figuring things out?
I want to work on new processes for science in the era of AI if anyone wants to join. And no I don’t think the existing ML conferences will adapt soon enough. They’ll take years of pain that I don’t want to bother trying to get involved with.
Yes. Me too! I have things I'm working on behind the scenes which are still in rough draft form. But maybe I should share those drafts here... I post a lot of my rough drafts here.
In 2014, I suggested to Philip Stark that senior people have more responsibility than junior people to change evaluation structures. He, too, said he was working on it.
Would something like Twitter but only for paper reviews work? Each researcher can curate their feed based on what they're experts in and also get feedback organically.
Bluesky community aside, AT Protocol seems like a pretty solid foundation for sharing and discussing papers. More promising than desperately trying to resurrect Twitter or Last.fm or w/e.
Yes, each of these, but individual refusal remains difficult in corporations as well as universities, esp when careers still depend safety and publication, respectively. What collective actions could change the incentives without asking junior employees and researchers to sacrifice themselves first
While true, this ignores the consequences of doing something else. The individual consequences of choosing not to publish (or submit) can include perishing, which eliminates the author's power to continue publishing with the affiliation of the institution that executed the perishing---or maybe any future institutional affiliation.
The institutional power structures (deans, hiring committees, P&T committees) that promote/require publishing will likely advocate for scaling the productivity systems (e.g., AI reviewers) over reflection upon why we publish followed by policy changes.
Wisdom does not need as many words as knowledge, but only understanding can create wisdom, and it is, unfortunately, a derivative of learning knowledge. But learning knowledge does not necessarily lead to wisdom.
I suspect those who do not play the AI arms race might find they have an insurmountable advantage in a future iteration where those who purport to be credentialed gatekeepers in their hierarchies are but naked kings wholly reliant on the mediocre recursive knowledge amalgam spat out of AI.
The agency you're referring to is lacking in those structures because for a while now we've been replicating hierarchies where those who are promoted do not embody ethical virtues. Talk is cheap etc. I dare submit that ethically minded intelligent people progress more arduously in unethical environs and are more likely to exit fraudulent games, thinning the quality of those players who remain.
Personally I didn't pursue a PhD over a decade ago because it initially seemed pointless, and later, when I finally found a worthwhile question to consider doing a PhD on, the idea of passing between three and five years of my life in university was a turn off. I'm not suggesting I would be a who knows what, but do consider the analytical inputs about what's wrong with the gamified model of university this essay is critical about is biased by those who are within the game and doesn't ask for the opinions of those who took a look and decided to walk away.
Great piece, Ben. So many opportunities to escape the vortex to the bottom and work in orthogonal directions to escape the positive feedback cycle of absurdity. Step one is to name it and remind us that new ways are possible.
Hallelujah!
Indeed the Neurips arms race is pathetic but very funny. And I have zero sympathy for any complainers; I just laugh. The solution is trivial, and takes two steps:
1) Just stop submitting, and disengage. That is what I did. It is very easy. I used to enjoy the conference (from about 1990-2015) because I liked talking to people there, but after 2019 I vowed never again (even now it is heading to Sydney because of work I did many years ago to get it there!). I have turned down every request to be a reviewer, area chair, senior area chair, and even senior program chair (a couple of years ago) and am delighted I did. And I don't submit there any more.
2) Stop judging people in terms of whether they are named as authors on papers published there (note my careful wording). Alas, my professorial colleagues (world-wide I mean) are the worst offenders here. This too is pathetic...
[An alternative (only slightly flippant) suggestion which I have made to a Neurips board member is to decouple the papers from the conference: by all means have the conference, for people to confer, but do not tie it to any publication; separately, have a publication venue (aka a journal). Demand submitters to pay handsomely for each submission. Use the vast profits to pay for nice food etc for the conference attendees who would then attend for the sake of conversations. The conference would no doubt shrink which would be a fine thing, and it might be fun to attend again...]
Regarding your other points, well Geoff Hinton reckoned there would be no radiologists any more (I am very glad he was wrong, having interacted with them to my considerable benefit). And other technological superstars are equally terrible at making predictions (there is a large literature on this, not just an opinion, but a well documented pervasive pattern ... look at all the refs in chapter 3 of https://acola.org/sites/default/files/2026-03/saf05-Technology%20and%20Australia's%20Future%20report-full-17sept%202015.pdf ). So when I read the latest predictions, I just recall the performance of prognosticators of times past.
My free advice: Say no to Neurips. And say no to stupid chatbots, and get on with life. Easy :-)
One thing we need to address is what junior people should do. They have more pressures, but they also have more agency then us older folks. If they go away, all of academia does too!
Indeed! Thinking of them was a primary motivation for expressing my opinion so bluntly. There are many things we can do: for example, stress the value of intrinsic rather than extrinsic rewards (well documented advantages are demonstrated by Deci and Ryan); declare our (old folks) disdain of judging by appearance in a venue (sign the DORA to signal your stance); encourage them to practice articulating what they are doing and why it matters (rather than relying on a venue where it shows up as a certificate of quality); encourage them to be proud of what they do, rather than what people say of it, and have the courage and gumption to take the long view; teach the value of good scientific conversations, so that when they go to conferences, they can make the most of the experience rather than trying to impress; decouple funding approval for conference attendance from having a paper there (the normal scheme implicitly says to young folks that the reason for going is only to transmit rather than to receive); explain to them that many of the "pressures" are self-inflicted (e.g. anxious and tawdry comparisons with others); telling them that the bar is low -- they can easily do better than our generation (because they have seen our collective mistakes); and leading by example in not being a complete ranker.
It seems that writing was easier to automate than reading - or, what Tao, in a strange gastrointestinal metaphor calls (proof) digestion - I wonder why that is...
Is this true? LLMs can automate both "writing" (generation?) and "reading" (verification?) under some definition right now, but for a human to understand the output of any automated writing they obviously have to read it themselves
Feels like a cargo cult to me! Math is not about proving theorems or writing papers. We don’t stop running because cars are faster. Isn’t it all about the joy of figuring things out?
I want to work on new processes for science in the era of AI if anyone wants to join. And no I don’t think the existing ML conferences will adapt soon enough. They’ll take years of pain that I don’t want to bother trying to get involved with.
Yes. Me too! I have things I'm working on behind the scenes which are still in rough draft form. But maybe I should share those drafts here... I post a lot of my rough drafts here.
In 2014, I suggested to Philip Stark that senior people have more responsibility than junior people to change evaluation structures. He, too, said he was working on it.
Would something like Twitter but only for paper reviews work? Each researcher can curate their feed based on what they're experts in and also get feedback organically.
Sometimes I feel that's what Twitter and Bluesky are trying to be.
Though the community can be annoying, the curation tools on Bluesky for what you describe are pretty great.
Bluesky community aside, AT Protocol seems like a pretty solid foundation for sharing and discussing papers. More promising than desperately trying to resurrect Twitter or Last.fm or w/e.
Yes, each of these, but individual refusal remains difficult in corporations as well as universities, esp when careers still depend safety and publication, respectively. What collective actions could change the incentives without asking junior employees and researchers to sacrifice themselves first
1) The senior people have agency, too, even though they act like they don't.
2) if the junior people go away, the whole thing collapses.
I know systems are maddening to navigate, but I still believe collective action is possible.
"We all have the power to do something else."
While true, this ignores the consequences of doing something else. The individual consequences of choosing not to publish (or submit) can include perishing, which eliminates the author's power to continue publishing with the affiliation of the institution that executed the perishing---or maybe any future institutional affiliation.
The institutional power structures (deans, hiring committees, P&T committees) that promote/require publishing will likely advocate for scaling the productivity systems (e.g., AI reviewers) over reflection upon why we publish followed by policy changes.
I hear you, but those power structures are made of people too! They can also adapt to new means of evaluation and promotion.
Wisdom does not need as many words as knowledge, but only understanding can create wisdom, and it is, unfortunately, a derivative of learning knowledge. But learning knowledge does not necessarily lead to wisdom.
I suspect those who do not play the AI arms race might find they have an insurmountable advantage in a future iteration where those who purport to be credentialed gatekeepers in their hierarchies are but naked kings wholly reliant on the mediocre recursive knowledge amalgam spat out of AI.
The agency you're referring to is lacking in those structures because for a while now we've been replicating hierarchies where those who are promoted do not embody ethical virtues. Talk is cheap etc. I dare submit that ethically minded intelligent people progress more arduously in unethical environs and are more likely to exit fraudulent games, thinning the quality of those players who remain.
Personally I didn't pursue a PhD over a decade ago because it initially seemed pointless, and later, when I finally found a worthwhile question to consider doing a PhD on, the idea of passing between three and five years of my life in university was a turn off. I'm not suggesting I would be a who knows what, but do consider the analytical inputs about what's wrong with the gamified model of university this essay is critical about is biased by those who are within the game and doesn't ask for the opinions of those who took a look and decided to walk away.
Great piece, Ben. So many opportunities to escape the vortex to the bottom and work in orthogonal directions to escape the positive feedback cycle of absurdity. Step one is to name it and remind us that new ways are possible.