The train is sliding through the dark of the night, and as the internet connection has been down since before dinner, I finally have time to myself. Trains are now seen as slow travel, though they were seen as killing the journey when they were introduced. An industrial revolution that changed our world and created a bubble that ruined families when it burst. And yet we adopted trains and moved on to something even worse with planes. Is travelling useful? Most of our discoveries could be made at home, and there is a sort of consensus that what matters is not the destination but the journey. The true journey is the return.
As the train gently moves, the comparison with LLMs is too easy: another bubble that steals our time and cuts the journey short, as many see it. I have to say that despite all their flaws, I am fond of LLMs. I went into machine learning when I could still feel my professors’ disdain for choosing this path. There were no lectures on the topic, so I moved to another university where, by contrast, the hype was well established and obvious to everyone. Very few would have predicted the rapid growth of LLMs, though. NLP was still a very unsatisfactory topic. On the informal evaluation sheet, someone had written:
the NLP lecture is to lectures what English cooking is to gastronomy.
(Yes, we are French.)
Machine learning was already magical, though, and I convinced myself that with good supervision or feedback loops, enough data and compute, anything was possible. There is a lot of joy in seeing how far we, as human beings, have come in constructing LLMs. I hope other researchers also have time to feel this joy.
However, I hear much more despair, fear and dismay than curiosity. Negative sentiment tends to dominate on social networks as they are currently structured, but in this case, I believe it also reflects researchers’ feelings, at least in Europe (where we are exposed to Silicon Valley’s existential crisis). LLMs may just be the perfect current scapegoat, but I do feel an unusual sense of doom as well. I have observed that all researchers starting a permanent research position question themselves harshly, so I would have gone through this regardless of LLMs, but I still link it to LLMs to some extent. I recently endorsed one of the many texts warning about the dangers of AI for its creators, mathematicians and ML researchers, and calling for a slowdown. However, while slowing down is well defined when driving a car, and usually an easy way to increase safety, it is not an actionable agenda for a research programme. So here is my own contribution.
I like that machine learning research sits at the intersection of maths and computer science, because to me they represent two extremes in the construction of knowledge.
Computer science is the realm of bootstrapping. From hardware to assembly to low-level code to code to libraries to pipelines. If you are not comfortable with black boxes, you are not a good computer scientist. You have to accept trial and error and work with abstract entities you will never fully open.
Mathematics is pure. Only what you have built alone in your head matters. You slowly construct this private garden, and memorising results, accepting them without proof or leaving parts in shadow comes back to haunt you.
Both worlds exist, and both are interesting. Maths tends to be a good proxy for training and measuring the ability to reason correctly about complex problems, and is thus often used to select or train people. Mathematicians usually complain about this, emphasising the beauty and interest of the results themselves. I always find this complaint valid but overstated: most mathematicians are mathematicians because they took enough pride in being the best at maths to keep going. Competition in maths is fierce, and opposing elitism sounds like a pose. The fact that almost no one understands what is going on has also allowed false theories to survive for quite a while. LLMs might give mathematicians a useful check. I do not believe that some subfields are necessarily broken, but I think some work may tackle problems solved long ago in other domains, and the difficulty of learning unfamiliar notation is enough to prevent researchers from other fields from checking it.
I also feel that the fear of lagging behind LLMs comes from the habit of being at the top. I have always assumed that I was close to being too stupid to stay in the field, and would thus be surrounded by cleverer people. This brings shame, but also constant opportunities to learn, which for me outweigh the downside. This position is, I am afraid, quite gendered. It is very close to Simone de Beauvoir explaining that when she met Sartre, she dreamed of a man who would dominate her intellectually. This made sense to her, as her feminine education had obviously slowed her learning compared with men’s, so it was natural to follow in order to learn. Because I expect a constant fight just to be clever enough to stay in research, I do not feel a drastic change with LLMs. On the contrary, being able, at any time of the day without availability issues, to close a chat after asking an LLM to explain a mathematical proof I do not understand as if I were five is an upgrade.
Of course, this is not the only consequence. What researchers complain about is the speed. Everyone working more, submitting more papers, quality lost to quantity. Machine learning conferences are growing exponentially, and this predates LLMs. Tenure is getting more competitive, grants scarcer; even securing PhD funding is competitive. To some extent, this pressure has made selection more rigorous. (Yes, older professors, sometimes you should remember how unlikely you would be to secure tenure today, rather than believing you are the most competent and should keep making decisions for the whole field.) But today, the process is not only more selective, but also noisier, messier and counterproductive. We are in Fantasia:
Stehe! stehe!
Denn wir haben
Deiner Gaben
Vollgemessen! –
Ach, ich merk’ es! Wehe! wehe!
Hab’ ich doch das Wort vergessen!
How do we stop the agents before we drown with them? In Fantasia, the apprentice is saved by the sorcerer, who stops the disaster and puts him back in his place. This is not an isolated example but a recurring theme in children’s stories, especially those for girls. I remember reading Les Malheurs de Sophie: the story tells girls that submitting and obeying adults brings peace and happiness. This obviously helps them believe in patriarchy later. Current proposals to pause AI often rest on the hope that an adult will come and enforce the pause: the idea that we will restrain AI until we are mature enough for it. This won’t work. One cannot be ready for LLMs without using them, just as one cannot be ready for freedom without experiencing it. We are supposed to be the adults in the room. The fact that most researchers ignored problematic applications when they were not directly affected does not help, of course.
A third aspect that annoys me is that researchers in my field seem to believe their situation has no precedent. This is obviously wrong. From the applications of CRISPR to the fear of destroying the planet with nuclear weapons, the crisis is hardly new. Even the old printing revolution changed the speed and structure of knowledge and was seen as out of control. More recently, nuclear physics has had a rich history of researchers engaging with the social implications of their research and going through severe existential crises as well. I think we can learn from these precedents. And since I cannot yet think of anything better than what already exists, I would like to propose a few ideas others have had before me:
- Stop imposing a race for quantity. Biology experienced citation inflation before us and pushed for the DORA recommendations. For years, CNRS has asked researchers to limit their reported contributions to three papers. We can easily stop reporting citation counts and abandon metrics that encourage quantity over quality. Some fields are organised much more around large projects than paper publications.
- Develop new ways to share knowledge. As the number of researchers grew, letters became too unwieldy for sharing information, so journals were created to organise the field. Now we need better ways to store our results. This could mean building verified libraries in Lean to cover entire research fields, or creating more interactive content combining proofs, videos and animated figures, now that web development is cheaper. LLMs are also, to some extent, a new way to store data.
- Develop active learning for researchers. With more resources online and new challenges in sustaining attention, pedagogy researchers recommend making lectures more interactive and having participants develop answers themselves. With LLMs, the cost of building new proofs or quick proofs of concept is low, so we can adopt this approach too. We should hold mathathons instead of small workshops: agree in advance on the open problems we want to solve (the part that takes time), solve them during the meeting, then spend a long time consolidating the results.
Of course, all this reshapes authorship and how researchers become famous. Here, once again, I am privileged. First, I never really hoped to discover anything important enough for the question to arise. Even if I had done so by chance, as a woman, I know very well that the Matilda effect would have protected me from fame. So a lack of individual recognition as research becomes more collective is unlikely to harm my desire to do research. On the contrary, it could even improve the community. I would also like to warn against those who say they will make workshops invitation-only. Good riddance if they want to spend time with friends, but beyond reproducing existing biases and removing the ladder after climbing it, I don’t see how this can be a serious proposal. I have to confess that I am on the train to a workshop that, to my dismay, adopted this approach: papers were accepted without review, in a quick decision by the three programme committee members. The keynote speakers were chosen by the main organiser. The result is that I already know all the authors, and the three keynote speakers come from the US (for a European workshop!) and are co-authors. Need I say they are all men as well? I really believe this is a dead end, and that double-blind review and randomness, which open fields to new people, are crucial to research quality. I hope I can raise awareness of this.
There is still a problem: how do we find interesting problems if solving them is cheap? This is hard because, until now, solving problems took enough time for new ideas to emerge naturally along the way. Many researchers also never find new problems, but simply solve those others have written down. If I had the solution, I would be writing wonderful papers instead of this. But let me try:
- I will write proofs. We need to understand the solutions to previous steps. To keep building the realm of maths in my mind, I will try to teach and present my results as much as possible, in diverse ways. When teaching, I will demonstrate the pain of working through a proof on the blackboard, with all its mistakes and uncertainty.
- I will maintain trust with my co-authors and colleagues. If LLMs make everyone more productive, collaboration becomes even more demanding, so we need to be involved in fewer projects at a time. I will not increase the number of projects I work on simultaneously.
- I will keep my papers under 40 pages. As generating material is cheap, a paper’s quality depends on what we leave out. As a reviewer, I will start suggesting that less interesting parts be removed. What we do not include in the paper is part of the contributions. We should aim for impact rather than completeness. Before sharing any writing, I will take the time to reread it in full.
- I will listen to other fields. Mathematicians now look quite ridiculous panicking when they did nothing while artists went through a similar phase. As we need more problems, we need more diversity, and other fields can provide it. Most fields are also much less affected, as many aspects are harder to automate (field observations, physical constraints on experiments, more diverse output formats). Most humans are still convinced that LLMs are useless.
- I will keep putting the younger generation first. Universities are places for both research and young minds to blossom, inspired by rare research successes. Educating the next generation is most researchers’ most tangible contribution, and it should remain so. In particular, I will continue to hire PhD students even if they are less "productive" than an LLM on a well-defined project. I will spend more time chatting with them than with an LLM, and I will try to help them use LLMs efficiently and appropriately as well. This means I aim to transform LLMs into convivial tools. Currently, this is not the case: typical use involves individual prompting and sometimes bombarding someone with a far too long chat. But this can change. Maybe I can even help build such tools.