Not Slow, But Doesn’t Fit in a Data Center

On watching an ethical debate morph into an economic one, and the remaining roles of the human mind in guiding progress.

Two Aperiodicals in 3 days? It is really aperiodic!

I write this — I have a vantage point on this and some time to set aside — because my own current science does fit in a data center, cosmological simulations belonged on the supercomputers since the beginning. AxioNyx to this day has FORTRAN components. At the same time, however, I have also had training in sciences that do not.

What do I mean by “fits in a data center”?

Let me lay out my observations and attempted characterization of some recent events. I hope this helps answer that question.

This month, arXiv, the largest preprint service for quantitative sciences, announced a two-submissions-per-month maximum for all registered users, in an effort to combat the piling on of AI-assisted and AI-generated works. I am not sure the rule is already in effect, but the pressure behind it is real: the human quality assurance volunteers who uphold the standards on ethics and technical consistency are already far and hopelessly outpaced by the production of papers. Yet more voices argue that arXiv’s decision only underscores that the entire journal system belonged in the previous generation, and papers weigh less than a GitHub readme in some cases. I am at least aware of the debate. Other than 2703 I run no risk of hitting the submission ceiling doing conservative, as-though-Claude-Fable-did-not-exist, work.

2026 will be for AI in science what 2025 was for AI in software engineering.

Kevin Weil, Open AI. 26 Jan 2026

The same sensation witnessing AI-led changes in AI-adjacent fields in 2025, in aspects like scalable computing, web safety and operating system kernel debugging, is seeping into noticeably more broad aspects of society in 2026. An earlier naive hope that LLM and LLM-powered agents will converge to and settle in a niche of helping with code iteration, email sorting, and job application filing (or, reviewing) is overthrown. Discontent being merely assistants — debatably a chatbot has no free will but money flux drives things similarly to how glucose flows drive the basal ganglia¹— are heading to the front line of research, blowing a raspberry between context window compressions. Somewhat fittingly as the involved might say, or giggle at, Altman’s meme “collapsed in a chair watching the nuclear explosion”. Decades of human effort are being overtaken.

Such change has most recently taken the form of a mining operation on unsolved mathematical problems: speedrunning Hilbert’s and Erdős’ lists, and headed for the Clay millennium ones.

Why math?

Because it fits in a data center. I think I am ready to come to explaining my thesis.

Broad strokes, mathematics is a highly formalized system of knowledge². Proving and proving false are immediate checks. More so thanks to decades worth of human altruism making a culture of such problem-answer pairs and checks (often nontrivial) available as part of the field’s culture and openly accessible.

This is really helpful for reinforcement learning, which this generation of AI depends on. How accurately, immediately, and cheaply a reward signal can be generated is key to rapid evolution of the models, and mathematics happened to tick all the boxes.

The state-of-the-art chatbots took less than a year to rise from Silver Medal level at the International Mathematical Olympiad to a promising contestant to a full mark. Yes, tens of millions of dollars worth of compute spent on this — becoming carbon dioxide from generators and hot vapour from the cooling system — but also undeniable progress, fitting of the public attention and speculation.

Before AI can drive beam lines and launch and orient new space telescopes, formally, at least, physics still is a human-safe field. Bots can write papers too, but the controls for rewards and factuality checks are hidden away on top of a high shelf. With that I can still sleep sound.

But that’s not what I wanted to discuss. Not comfort for self nor the world.

Again, this really marks that the relationship of flesh-and-bone humans and AI has long evolved past a feasibility debate based on philosophy and values. Sufficient minds swung and money flowed with them, enabling physical change. This is economical. The next field they uproot will simply be defined by the ease of access of a reward function. Again, how accurately, immediately, and cheaply can the models iterate.

There are sciences that are not happening slowly and yet do not fit this model. AI capitals currently have no interest or ability to develop them. Long-term medical studies, macroeconomics (not limited to AI speculation and aftermath themselves), climate change, history, all look very “off” when handed to AI. A feedback may or may not come in decades, and when it does it isn’t a Boolean true or false answer.

Immediately checkable problems and important problems overlap, but they are not the same set.

And hence I voice my concern.

1. The neural structures most correlated with human autonomy ↩︎
2. I am too lay and lunch-deprived to describe Lean in detail, but that concept pretty much captures all discussed here. ↩︎