I gave talks at Caltech Friday of last week and at Sorbonne on Monday of this week, and after so many discussions with some very good scientists, the signal is unmistakable. A tipping point : many rigorous/ traditional scientists have stopped debating whether AI models (no.. I am not talking about SciML, which has had limited impact thus far outside of sweet-spot problems) are good enough for top level research. They're talking about implications in a world when most cognitive work is getting off-loaded! Even the people you wouldn't expect to say it out loud are saying it out loud, and not mincing words.
Unless one is living under a rock, or has become a static parrot*, it must be clear that the AI inflection point was a while ago. But that's not my point. As tools are getting better and better (it is not because of "more data" or "scaling". It is because of scaffolding and unhobbling), opinions and perspectives are changing faster than any time in the history of science. People used to quip that "Science progresses one retirement at a time". But this feels like a true tipping point: Opinions have changed so much even since I started this blog (and that was just November!).
Now.. Caltech is a venerable place. When you visit, you get a feeling that time stands still***, and that this university distills things. Caltech is also different in other ways: When you give a seminar at GALCIT, nearly all the faculty show up. Including the Emeritus professors. Hard-core applied mechanics people, fluid dynamicists, people who have spent decades doing things the traditional, rigorous way. But I digress..
What struck me though was how strongly almost everyone was feeling about what AI is already doing to cognitive work. Sure, there might have been some sampling bias amongst the people I met (my talk was NOT about AI, actually and my two slides closest to AI were on the limitations of SciML-type approaches), but I did meet a dozen or so distinguished professors (and only one is an AI researcher). These are people whose instincts are to be precise, and careful. And they were not debating whether the game has changed. They're debating what to do about it (yes.. I know I said it already!) and what the research enterprise will look like.
The 100-foot problem. At lunch, one of the discussion points was that "AI brings anyone within 100 feet of any problem... but not beyond it." BTW, 100 feet is still incredible, because it used to take a decade of training to get there for any problem that matters. I'd add that it also gets us within 10 foot of some, and 0 feet in a few** (that is the essence of AjoGI). But to cover that final however many feet in science, you need a good mix of certain type of skills and modern tools. Deep domain knowledge. Physical intuition. Mathematical maturity. The kind of understanding one normally builds by doing things the hard way a few times (i.e. doing the hard yards to get to within 100 ft).
Which leads to the question: How do we train students to cover the final 100 feet if we remove the friction and struggle (or friction naturally fades)? Is friction good or bad? My opinion: it's a good thing. Mostly. Always. Sometimes. Always, well... mostly, right? The struggle of deriving something by hand, of debugging code line by line, of reading a paper three times before it clicks is not wasted effort. That's how intuition is built. Take it away entirely and you get people who lack intuition or even what questions to ask, or how to spot inconsistencies. Keep it all and you could get people who are 10x less productive than they could be. The balance is everything and we don't know where it is.
What are we training them for? This is the deeper, harder question. What does the world look like in 5 years (that's the duration of a typical PhD)? What is scientifically valuable when AI can do many (though not all) cognitive tasks better than humans? I've written about this (but that was 4 full months ago!). What are the moats? Perhaps data, infrastructure and style. By style I mean the distinctive way a research group frames problems, the taste that determines which questions are worth asking, the judgment that separates interesting from important. Very shortly, we'll be customizing models (I predicted 2 years ago that this would happen in late 2025. I don't think I'll be off by a lot) --or at least instructing them in special ways-- and the value will come from what we bring to that interaction. Some aspects of what I wrote a few months ago, I now believe more strongly will happen, and sooner than I expected. I repeated this in a panel at Sorbonne as well: Thus far, we've given the highest amount of respect to academics who come up with elegant, tractable problems, develop clean solutions and insight. Nothing wrong with that. I think the value proposition will now shift to those who solve practical, long-standing problems.
The Golden age of the PI : One other topic is how much of a boost these tools are providing to PIs... at least those who do some research on their own. One professor with good taste, deep knowledge, and modern tools can now do what used to require a few students. I also met a couple of professors at a conference recently, and they say they now do more research than most (all) of their PhD students. Elsewhere, here is a professor pondering whether it is worth hiring and training a student. Perhaps that is an extreme case. The purpose of being an academic is to train students.. and well.. PhD students are more than "task completers". All of that said, it would be naive to ignore recent advances that led to the question, and the rate at which the situation is evolving.
Coding agents powered by reasoning especially have had a huge impact on my research. Over the past decade, I have averaged around 20-30 hours/week (obviously the week is much more than 40 hours) reading up on all kinds of things, doing derivations, writing (bad) code.. only to drop it because a) my real job calls, or most often b) something even more interesting came up. Among other things, I am quite shocked at how agents can take half-baked, badly written codes and track my line of thinking that gets broken frequently, and provide continuity and insight. Personally, a huge thing.
It's also, let's be honest, a little terrifying. Was it Elon Musk who said we'll have "radical abundance and rioting in the streets"? The optimistic and pessimistic versions of this future are uncomfortably close together and co-exist, and so the shakeup of the research enterprise might be both amazing and unsettling.
Here's something I didn't know: France's scientific golden age was shockingly brief. The names on the Eiffel Tower cluster overwhelmingly from about 1800 to 1830. During those three decades, Paris was the undisputed center of world science. Nothing came close.
And then, almost as quickly as it rose, by the 1850s, the center of gravity in science had shifted to Germany and Britain. By mid-century, the three great advances in theoretical physics: thermodynamics, kinetic theory of gases, and electromagnetism, were all being made by Thomson, Clausius, and Maxwell. France, which had pretty much created two of these subjects (Carnot in thermodynamics, Ampère and Coulomb in electromagnetism), contributed essentially nothing to the breakthroughs that followed (Ben-David & Herivel).
The French system was brilliantly designed for the era in which it was created (and Ecole Polytechnique had a big part to play). France had the mathematical/theoretical talent and a head start. What France didn't have (relative to Germany & England, that is) was the flexibility to pursue new kinds of questions in new kinds of ways.
I see some parallels now.. The question is whether we will build structures flexible enough to evolve with the sudden burst of incredibly powerful AI? The French case also speaks to the friction question. The mechanistic tradition, with its insistence on understanding why (molecular models, force, deeper explanations) was the tradition that turned out to be aligned with where physics was actually going.
From my reading of some literature (I did chat with a Professor of the history of science at Sorbonne as well) England and Germany overtook France because they combined theory with practice, cognitive with empirical, and apparently institutional structures that rewarded both. The lesson for today: if AI collapses the cognitive work of science, but we lose the empirical grounding, the physical intuition, the willingness to ask why and not just what, most bottlenecks won't be going away anytime soon.
It felt fitting, because I think we're at the start of another Golden era, potentially just as consequential for science, and we can't get carried away by the early gains in purely cognitive acceleration. We need to
1. Add `friction' as necessary
2. Build Agentic AI systems around instruments, simulators, verification mechanisms, and of course scientists. This will give the empirical grounding and make non-computable leaps accessible. A conceptualization of such a system, can be found in this article.
3. Build the right institutions to manage this new era of science.
I spoke about some of these things in my Keynote at Sorbonne, at the kickoff of the Paris AI for Science Cluster (SCAI)
* Not that there isn't a ceiling to these models. But if people don't agree that these tools are powerful now (and hide behind weird / outdated ideas) "static parrot" sounds appropriate.
** Right now, in some math, formalization and coding problems mainly
*** Speaking of time standing still, in one of my discussions, a Caltech colleague spoke about a problem he was working on.. and I remembered a slide from my PhD defense which I was able to pull up (wow!), and that led to an animated discussion
