Knowledge Was Never the Bottleneck

For a long time, a serious scientist could be a generalist.

Newton could move between mathematics, mechanics, optics and theology. Faraday crossed chemistry and electricity. Maxwell connected electricity, magnetism and light. The early modern scientist did not know everything, but the frontier was still small enough that one unusual mind could hold several maps at once.

Then knowledge became too heavy.

Modern science did not kill the polymath because people became less curious. It killed the polymath because specialization was the rational response to depth. To reach the frontier, you had to spend years entering one corridor. Then you had to work with people in neighboring corridors: the instrument specialist, the statistician, the domain expert, the code person, the publication insider.

That was not a failure of science. It was progress.

But it had a cost: fewer bridges.

AI Makes It Cheaper to Cross Disciplines

The important thing AI changes is not that it makes people faster at writing.

It lowers the cost of entering an unfamiliar field.

A good model can map a literature, explain vocabulary, compare frameworks, write exploratory code, translate one domain into another, generate first-pass simulations, find adjacent papers, and prepare better questions for a human expert.

That is not expertise.

It is scaffolding.

But scaffolding changes what kind of person can do useful work. Before AI, broad curiosity was expensive. Every detour into a new field cost weeks or months before you could even ask a non-embarrassing question. Now the first map is cheap. The dangerous part is that cheap maps are often wrong. The useful part is that a disciplined researcher can now hold many provisional maps at once.

The new independent researcher is not a return to the gentleman scientist. The gentleman scientist worked alone in a small corpus. The AI-augmented researcher works inside a vast corpus through a collaborator that can keep more of it in reach.

The role is closer to a director: one motivated human, AI as a transdisciplinary collaborator, and a small number of experts, tests, and formal tools acting as gates of quality.

This is the first thesis.

AI makes polymathy operational again.

The second thesis is more uncomfortable: once agents can produce the work, the real bottleneck moves to trust.

Software is already showing the pattern. Agents write code, agents review code, agents run tests, agents patch regressions. Humans increasingly specify the target, inspect the evidence, decide what ships, and take responsibility when the system is wrong. In that world, the hard question is no longer "who typed the code?" It is "why should we trust this change?"

Science will face the same question. If AI agents can generate hypotheses, run simulations, search proofs, optimize experiments, and draft papers, then discovery becomes easier to produce and harder to evaluate. Which result is interesting? Which proof is correct? Which benchmark was gamed? Which experiment reproduced? Which apparent breakthrough is just a rediscovery, a local optimization, or a beautiful hallucination?

So the new research problem is not only how to make agents discover things.

It is how to build the trust infrastructure around machine-generated discovery.

Recent AI Research Results Show the Shift

The clearest early signal is mathematics, because the outputs can be checked.

In 2023, DeepMind's

paired an LLM with an automated evaluator and found new solutions for the cap set problem and bin-packing heuristics. The point was not that the model "understood" the whole problem like a mathematician. The point was the loop: generate, score, keep what survives.

In 2024, DeepMind's

solved four out of six International Mathematical Olympiad problems, reaching silver-medal level. In 2025, an advanced Gemini Deep Think model under official IMO grading, solving five out of six problems.

In 2025, DeepMind's

moved from contest-style reasoning toward algorithm discovery: matrix multiplication improvements, code optimization, and progress on open problems such as the kissing number problem.

Then in May 2026, OpenAI announced something sharper. An internal model

around Erdős's planar unit distance problem by finding an infinite family of point sets with polynomially more unit-distance pairs than the long-standing conjectural limit allowed. A group of external mathematicians published a explaining and digesting the result.

That last step matters.

Solving olympiad problems is impressive. Improving search heuristics is useful. But a model producing a new route through a real open problem, one that external mathematicians can verify, is a different kind of signal. It suggests that AI is beginning to participate in the creative part of research: not only summarizing what is known, but proposing bridges humans had not prioritized.

The same pattern is now appearing in AI research itself. In June 2026,

published early results from an automated AI research system on NanoChat Autoresearch, NanoGPT Speedrun, and GPU-kernel optimization tasks. Shortly after, announced results from its autonomous research system across NanoChat, ProteinGym, and NanoGPT Speedrun. The details differ, but the shape is converging: define a measurable task, launch many hypothesis branches, evaluate candidates, audit for reward hacking, and keep the improvements that survive.

This is still early. Benchmarks can be overfit. Evaluators can be gamed. Some gains are rediscoveries, implementation details, or local optimizations rather than deep scientific insight. But that is precisely why the loop matters more than the single result. The interesting object is not just a model that proposes ideas; it is a research system that can generate, test, falsify, archive, and harden them.

What Has to Be Built Next

This does not mean anyone can now do science.

AI does not give experimental access. It does not make a bad question good. It does not remove the need for expertise. It hallucinates, overstates, imitates consensus, and sometimes produces beautiful nonsense.

Knowledge and computation are becoming easier to delegate. Taste is not. Persistence is not. The decision to stay with a problem after ten false starts is not. The ability to know when a result is real, when it is merely plausible, and when to ask a senior expert is not.

AI gives you more surface area.

It does not give you a center.

So the practical agenda is not "ask the model for discoveries." It is building systems where guesses are cheap and verification is strict:

• ingest sources, but keep them traceable
• let the model propose connections, but log contradictions
• write code, but rerun and reproduce it
• use formal systems when claims are mathematical
• audit attribution and novelty before calling anything new
• ask experts when the local verifier is not enough
• publish only what survives contact with reality

That is the shape we are exploring at Yuki Capital: AI for breadth, literature, computation and candidate generation; humans and verification systems for taste, judgment and proof.

Some of that work is already visible on our

. The public projects are only the surface: behind them are experiments in AI-assisted literature review, data pipelines, reproducible analysis, expert review, and research workflows where models propose directions but the claims still have to survive evidence.

That internal lab engine now has a public name:

. It is our workbench for the loop described here: broad candidate generation, candidate archives, benchmark adapters, replayable runs, evidence gates, negative-evidence logs, promotion decisions, hardening reports, and artifacts another researcher can inspect. The point is not to make the machine declare itself right. The point is to make it easier to produce many plausible routes while making it harder for weak claims to escape. Some implementation details stay private for now, but the operating principle is simple: randomness in discovery, determinism in verification.

Scientific Gatekeeping Has Not Adapted Yet

There is one more bottleneck AI does not remove: legitimacy.

The production side is changing fast. A small team can now read more, code more, test more, and draft more than would have been plausible a few years ago. But the distribution side is still built for the old world.

arXiv is open in principle, but posting in many categories still requires endorsement from someone already inside the system. Journals still use affiliation, citation networks and editorial judgment as soft filters before a paper is even sent to referees. Funding often requires an institution. Conferences and seminars still run through networks of trust.

Those gates exist for a reason. If AI makes it cheap to produce plausible-looking papers, removing all filters would flood the corpus with noise.

But affiliation is now a worse proxy than it used to be. The ability to produce serious work is starting to move outside the institutions faster than the validation infrastructure is moving with it.

That is the uncomfortable middle: independent AI-augmented research is becoming technically viable before it is socially legible.

The answer is not "no gates." It is better gates: reproducible artifacts, public code, data releases, formal verification when possible, expert review, and a track record that can be inspected directly instead of inferred from institutional address.

The next polymath will not be someone who knows everything.

It will be someone who can cross fields without getting lost.

Less romantic than the Renaissance.

Much more useful.