AI (as we know it) won't kill scientific research.

This post is being written on September 2nd, 2026, on my first day of classes at Harvard College. Anthropic’s Fable 5.1 just dropped yesterday, and OpenAI’s Astra may be dropping soon. Everywhere, we are awash with the unmistakable imprints of AI.

One of the first things covered in my first class at Harvard, the notorious Math 55, was the AI policy. Consulting clubs around campus are advertising their “comps” with such flagrantly AI-generated posters that even the old Ghibli-profile-picture people wouild turn up their nose. I see people walking around with open lids, Claude Code churning in their terminal. Deep in every common room poker game, late night foosball match, and Annenberg discussion over strawberry fro-yo, is this impending sense that we are in the middle of a revolution, that an undercurrent of change is passing beneath each of our noses and that the world in four years when we walk out of this university in those crimson robes will be unrecognizable.

Below, I outline my theory of change for AGI/ASI. I aim to revisit this periodically and see if any of my predictions have held up. As a general gist, I belive my views are more AI-optimistic than the average American public but more AI-pessimistic than the average San Franciscan.

Claims

Some of the points in this section are predictions, some are not; I mostly use this as a scratchpad to outline my overall thoughts. I may organize this a bit more cleanly in follow-up posts to address any updates to the predictions. A note: most of these claims relate to pure scientific research and less-so to more applied fields (e.g. AI engineering, devops, etc.) since the former is what is more relevant to me. Another note: this assumes that LLMs plateau in capability within 1-2 orders of magnitude of where they are today (also that AGI doesn’t happen / doesn’t kill us all / doesn’t radically restructure society). Anything beyond that and my confidence interval is unbounded.

1. LLMs operate on a low-dimensional (relatively) manifold.

This is a n=1 sample size of evidence, but I think that LittleLearner is a prescient example. LittleLearner is a 5B-parameter model trained only on knowledge a average fifth grader would know. The interesting finding was that while improving pre-training and post-training enhanced LittleLearner’s capabilities in fifth-grader knowledge domains, it was not able to concretely learn advanced reasoning skills outside its training data. I am perhaps very Lecun-pilled when I believe that models need to have a concrete world model before being able to generalize to OOD tasks.

Abstractly, I believe that all of current training data (~human knowledge) lies on a relatively low-dimensional manifold within a high-dimensional space. This manifold is the span of all human “thought vectors.” I believe that discoveries along this manifold will be rapidly exhausted by LLMs taking every “linear combination” of these thought vectors. As for off-manifold exploration, see below.

2. We cannot solve science operating on a low-dimensional manifold.

Every other science neolab is claiming that LLMs and adjacent foundation models will solve science. According to a significant number of them, within a time horizon of ~1 decade (or even less!), we will unveil the identity of dark matter, solve cancer, end world hunger, and cure male pattern baldness. These, and their fellow brethren, are of course immensely hard problems whose solutions may be on the same order of novelty as Einstein’s relativity or Newtonian mechanics. The solutions to many of these difficult problems will likely require approaches off the manifold (though I concede that the majority likely do lie on the manifold and will be rapidly exhausted and solved); perhaps in directions completely orthogonal to the surface. If LLMs operate simply along a span (see point 1) of all known fact/knowledge vectors, then certainly, there is a lot of science between the gaps that will inevitably be filled over (see, e.g. the AI-assisted math proofs), but they will never be able to explore that off-manifold subspace.

That space is for us to explore.

3. Academia will speed up, then slow down.

AI will not be the death of scientific research. Imagine a society of aliens where arithmetic has never been formalized (imagine they percieve the world fundamentally differently, like heptapods). The academia of this alien world largely consists of novel combinations of new numbers and novel arithmetic symbols. Then one day, a calculator is introduced into the world. Suddenly, any syntax-appropriate combination of arithmetic symbols is computable. The first aliens with access to this technology begin to churn out new papers at an astonishing pace. Soon, the technology is made public, and large organizations set up automatic calculators that automatically iterate through symbolic sequences and solve arithmetic problems. Academia is flooded, and peer review is dead. The entirety of this world’s research has suddenly been “automated.”

But there is so much more mathematics than pure arithmetic. These aliens will never discover analysis, group theory, differential geometry, perhaps even calculus with just a calculator. There needs to be an extrinsic insight to orthogonal to the calculator’s line of work.

Slowly, the journals begin to restrict papers that are just pure arithmetic “slop.” The same aliens that were publishing thousands of papers a year are now left rejected from every venue save a few select “arithmetic-only” journals. There is of course, an initial outcry from the calculator-heads. But eventually, one by one, every venue, begins to desk reject basic arithmetic papers.

This is of course, devastating for the aliens. For the first time in many years, the amount of papers published actually decreases. But there are also the Alien Einsteins and Alien Newtons, who continue to make ever-incremental steps away from the arithmetic manifold. The aliens entering academia realize that these orthogonal directions are the only way to get published, and soon, significant brainpower and mass accumulates towards pushing beyond simple arithmetic. It’s not that no one cares about arithmetic anymore; it’s now simply an unproductive and “solved” field. Eventually, by sheer pressure, like leaks in a jammed hose, entirely new fields burst forth.

Progress is incredibly slow, of course, since these orthogonal directions are far more novel than anything in the arithmetic span, but it is still finite: perhaps, each paper gets published every 2, 5 years, but it contains a field-shattering insight.

If it wasn’t obvious, this is the direction I expect scientific research to go in the age of LLMs. Currently, we are in the same situation as the “aliens when the calculator (i.e. LLMs) was first made available to the public. Eventually, journals will stop accepting “incremental” (relative to the capability of LLMs) contributions; even spotlight papers at the top conferences in the present may be of the rejected cloth in the near future. Academia will slow down, but the advent of LLMs means that suddenly, the idea space off the current manifold will be free and nearly limitless to explore. A new mode of academia may take the current one’s place, where ideas, not papers or citation count, are valued.

I think that will be a very positive outcome.

Predictions

Taking these claims (and their preceding assumptions), I wager the following about scientific research in the wake of AI:

1. We will see a rapid string of progress in the next 10 years, where the majority of unsolved problems in data-saturated fields (i.e., where there aren’t additional observations to be taken; a counterexample would be astronomy) will be identified. It is hard to put a quantitative number of this, but I believe that a qualitative outcome will suffice.

2. In the next 15 years, academia will rapidly shrink. This will be due to a) many people believing that “science is solved” and leaving the field for more concrete/corporate pursuits, and b) a selection effect due to the aforementioned off-manifold exploration requiring individuals who have developed a propensity for finding novel directions/ideas.

3. Within 20 years, the publishing pace of the average lab in most fields will become 0.25-1 papers/year, although fortunately, each of these papers will likely be significantly more substantive than the ones published nowadays.