AI and the Discoveries Hidden Between Disciplines

The recent announcement by Anthropic got me thinking more deeply about a topic I’ve been pondering for a while: the potential for large language models to connect discoveries across fields, and to find overlooked connections deep within them.

How much useful knowledge have we already accumulated without recognizing what it means when the pieces are brought together? A finding in one field may help explain an unresolved problem in another. A technique familiar to one community may offer a new way forward for a different discipline. Even within a specialty, something important may be sitting in a dataset that nobody has had a compelling reason to examine closely.

I suspect this is one of the most consequential opportunities for AI in science.

What Claude noticed

In its September 23 announcement, Anthropic described how Claude identified a previously uncharacterized biological system associated with reverse transcriptases, enzymes that copy RNA into DNA. The enzyme itself had been identified before. Claude noticed an overlooked combination of nearby DNA repeats and an accessory protein, and investigated its significance. The arrangement had features reminiscent of CRISPR.

According to Anthropic, roughly 950 agents explored the data over 21 hours. Claude examined the unusual pattern, compared it with known systems and checked the literature before reporting it for human review. Human scientists conducted the laboratory work. The system’s function remains unknown, and resemblance to CRISPR does not establish that it will become a useful gene-editing tool.

Those limits matter. This is early research. Even so, the possibility that interests me is clear: information that was already available contained a connection that had gone unrecognized.

The “Fields Medal” focus and jagged intelligence

It also raises a question about where the AI industry has been directing its scientific ambitions. Much of the most visible work has concentrated on exceptional performance in tightly defined areas: competition mathematics, formal proofs and difficult specialist problems. Google DeepMind’s Olympiad gold-medal performance and subsequent research-level mathematics work illustrate that emphasis.

I think of this as pursuing the “Fields Medal” end of intelligence: demonstrating extraordinary ability at the intellectual frontier. These achievements are substantial, and advances in mathematical reasoning can have much wider applications. But exceptional performance on a carefully specified problem tells us relatively little about how reliably a product will help a scientist or business user work through an ambiguous situation.

Competition mathematics offers a clear task and a rigorous way to judge success. Much of practical research involves deciding what the task should be, reconciling conflicting evidence, identifying missing context and recognizing that a useful idea may come from somewhere unexpected. Those capabilities are harder to capture in a headline score.

My concern is that optimizing heavily for the first kind of achievement may reinforce “jagged intelligence”: remarkable strengths alongside surprisingly basic weaknesses. Research on the jagged technological frontier has documented uneven usefulness across professional tasks. The further possibility I see is that specialist training could raise a model’s highest peaks much faster than it improves the surrounding terrain. That would make its capabilities more uneven even if none of them actually deteriorated.

For a working scientist, reliability across a sequence of activities matters enormously. A model might produce an ingenious argument yet fail to notice that the evidence supporting it comes from an incompatible experimental setting. For a business user, brilliant analysis can lose its value if the system overlooks an operating constraint. A higher ceiling on reasoning ability does not automatically resolve those gaps.

This is a concern about priorities and opportunity cost, rather than an established claim that mathematics training causes unreliability. There are already efforts aimed at broader usefulness, including Google’s AI co-scientist. Nevertheless, I suspect the pursuit of spectacular specialist results has left some of the opportunity to build dependable, broadly useful scientific and business products underexplored.

I would like to see connecting knowledge across disciplines, checking whether a connection survives scrutiny and helping people decide what to investigate next receive comparable attention in training and evaluation. That could make advanced AI useful to a much larger community. It is one reason this direction of biological research interests me.

The limits of specialization

Scientific specialization is both necessary and extraordinarily productive. Becoming expert in a field takes years of concentrated effort. Understanding which results are robust, which methods are appropriate and which apparent breakthroughs conceal familiar problems requires depth that cannot be acquired by skimming a few papers.

But there is an unavoidable trade-off. The more time someone spends developing that depth, the less time they have to explore everything else. Even an exceptionally curious scientist can follow only a fraction of the work that might eventually prove relevant to their research.

Research communities also develop their own vocabulary, assumptions and preferred techniques. Two groups may be studying related problems without describing them in terms that make the relationship obvious. Their work may appear in different journals and at different conferences. The connection can remain invisible simply because the right people never encounter each other’s results.

Human researchers have always crossed these boundaries. Some of the most interesting work happens when someone brings experience from one discipline into another. My argument is that AI could make this kind of exploration far more routine.

A capable LLM can draw on concepts from many disciplines and, with access to suitable sources and tools, investigate a possible connection in greater detail. It can move from an observation to relevant literature, from that literature to a method, and from the method to an analysis of the underlying data.

That creates the possibility of combining breadth with depth on demand. A model does not have to know in advance which specialty will become relevant. It can follow a promising lead across boundaries and bring the findings back into the original problem.

Of course, that capability is uneven. Training data is incomplete, published research has gaps, and fluent explanations can conceal misunderstanding. Access to many fields is no guarantee of mastery. The advantage I see is the ability to explore a much larger space of possible connections at a manageable cost.

Changing the economics of discovery

That cost matters enormously.

A scientist deciding whether to pursue an unusual observation has to weigh it against other demands on their time. A speculative connection might require days of reading before it becomes clear whether there is anything worth investigating. Most leads will fail. Choosing carefully is a rational response to limited resources.

AI changes that calculation. It becomes practical to investigate many more possibilities, including those that initially look too uncertain or peripheral to justify much human attention. Several lines of inquiry can proceed at once. An unsuccessful investigation can still help eliminate an explanation or reveal that a supposed discovery is already well understood.

I find this particularly interesting because we naturally focus on the dramatic moment of discovery. Much of the opportunity may lie in making the less glamorous work surrounding that moment faster and more systematic: searching, comparing, checking, reproducing an analysis and deciding whether an anomaly deserves attention.

From connections to evidence

The Anthropic result is primarily an example of investigation within biology. The broader prospect is to extend that approach across more distant disciplines. Imagine a researcher investigating an adaptation mechanism who can also draw on relevant work in control theory, materials science or ecology, without personally becoming an expert in all three.

Some proposed connections would be superficial. Others might suggest a useful experiment or a better way to formulate the problem. The value would depend on how effectively the system distinguishes between them.

This is where scientific judgment becomes even more important. If AI makes hypotheses abundant, the scarce resource shifts toward deciding which deserve testing and establishing what the evidence actually supports. A persuasive analogy can generate a research question. Demonstrating a mechanism requires further work.

The most productive relationship, I think, will combine AI’s capacity for broad exploration with the expertise needed to challenge its assumptions, design meaningful tests and interpret the results. Scientists should be able to trace a proposed connection back to its sources and see where established evidence ends and inference begins.

Beyond the laboratory

The same opportunity extends beyond science. Businesses also accumulate knowledge in separate teams, systems and professional specialties. An engineering observation, a maintenance history and a change in operating conditions may each look unremarkable in isolation. Bringing them together could suggest an explanation that no individual team had enough context to consider.

Here too, access and validation matter. An LLM cannot reliably connect evidence it cannot retrieve, and a plausible explanation still needs to be checked against how the business actually works. Making knowledge available for comparison is part of the architecture required to turn this potential into useful decisions.

What excites me about this direction is the possibility of getting much more value from what people have already discovered. We have invested enormous effort in creating specialized knowledge. AI may help us explore the relationships within that knowledge far more thoroughly than we have been able to before.

Some future breakthroughs will require entirely new observations. Others may begin when we finally recognize the significance of things we already know.