The Frontier May Exist in the Interaction

If I were explaining this idea to you in person, I would probably begin with the question that has followed me through several weeks of research: where exactly is the intelligence when a human and an advanced AI arrive at something neither seemed able to reach in quite the same way alone? I have gone through scientific papers, laboratory experiments, AI research, conversations with different models and some of the arguments taking shape around human-AI collaboration, and I keep seeing versions of the same problem. A scientist brings an observation or a research problem, an AI explores possibilities at a scale the scientist cannot realistically search, the scientist sees something useful in what comes back and changes the direction, and sometimes an experiment enters the process and proves both of them wrong. The result goes back into the system, another hypothesis appears, another experiment follows, and somewhere inside that movement new knowledge begins to form.

At first, the idea I wanted to investigate was simple: the frontier may exist in the interaction. I was thinking mainly about what could happen when brilliant human insight meets machine intelligence operating at a scale far beyond normal human reach, but the more cases I examined, the harder it became to keep the idea that simple. The human was not always directing the machine throughout the process, several AI agents were sometimes interacting before a scientist saw the result, tools and laboratory instruments were sometimes making important contributions, and experiments occasionally revealed something neither the human nor the machine had anticipated. What began as a question about human intelligence meeting AI intelligence was slowly becoming a question about the larger system that forms when different kinds of intelligence, tools and reality begin acting upon one another. That is why I am deliberately saying may. I do not know yet what the strongest version of this argument should be, and I do not want to decide what I believe first and then spend the next several papers searching for evidence that makes me look right. It may turn out that interaction only reveals capabilities already present inside the model, that human judgement mainly determines where machine intelligence should look, or that some intellectual territory genuinely becomes reachable only because the interaction happens at all. Those are different claims, and this series will lose its purpose if I choose between them before the evidence gives me a reason to.

That uncertainty is also why I do not intend to treat this as one paper written once and closed. The Frontier May Exist in the Interaction will develop as a research series through scientific experiments, observed interactions, competing explanations, failures, objections and questions whose answers are allowed to change when better evidence arrives. Some conclusions may survive almost untouched, others may become weaker or stronger, and some may turn out to have been asking the wrong question altogether. I want to watch the idea while the world around it is changing because increasingly capable AI systems are no longer arriving in isolation. Scientists are placing them inside laboratories, researchers are connecting them to databases and specialist tools, multiple agents are being allowed to reason together, and physical experiments are beginning to feed information back into the systems that proposed them. The scientific evidence already gives us plenty to work with.

Google’s AI Co-Scientist offers one of the clearest early examples. Human scientists define the research problem and provide feedback while multiple AI agents generate hypotheses, challenge one another, rank ideas and refine possible explanations, and some of those ideas have then moved beyond the model and into laboratories, including proposed drug-repurposing approaches for acute myeloid leukaemia that researchers experimentally tested. Robin, another scientific AI system, makes the structure even more interesting because it can search literature, generate hypotheses, propose experimental directions and then analyse the results that human researchers obtain from those experiments. In work investigating possible treatments for dry age-related macular degeneration, the process did not end when the AI produced an idea: scientists carried out the experiments, the results returned to the system, and those results changed what the system investigated next. Look closely at what is happening there. The machine proposes something, the human decides what deserves physical investigation, reality answers, the result changes what both sides know, and another question becomes possible. Where exactly would you draw the boundary around the intelligence responsible for the eventual discovery?

That question becomes even harder when we move to other experiments. Researchers at the University of California, San Diego generated and evaluated roughly half a million DNA-sequence variants while studying the downstream core promoter region involved in initiating gene activity. The human researchers created the biological problem and generated the evidence, but machine learning was needed to identify patterns across an experimental space far too large to inspect manually, and once those patterns became visible the scientists could use them to understand the genome differently. Here the AI did not independently decide which biological mystery deserved solving and it did not create the physical evidence, yet human intelligence alone could not realistically extract everything hidden inside the evidence humans themselves had produced. Anthropic’s reported protein-binder work gives us another version of the same puzzle. A human expert first translated knowledge of protein-binder design into a scientific protocol, Claude then conducted long design campaigns with substantial autonomy, and external laboratories physically synthesized and tested the resulting proteins. The human scientist had contributed years of knowledge before Claude began operating, that expertise shaped the rules of the search, the machine explored a molecular space with a persistence and reach the scientist could not reproduce manually, and the laboratory finally decided which predictions survived physical reality. Remove one of those parts and the outcome may be completely different.

That has become one of the most important developments in my thinking since I began this research: reality itself may belong inside the interaction. I initially pictured the idea mostly as human intelligence meeting machine intelligence, but the scientific experiments keep forcing me to widen that picture because a model can reason brilliantly about a molecule and biology can reject the molecule, while a researcher can bring decades of experience to a problem and an experiment can still show that the underlying assumption was wrong. Sometimes the failure itself contains the useful information, so the human notices something unexpected in that failure, the machine explores the new direction, another experiment follows and the process continues. Reality is therefore not simply standing at the end of the chain waiting to grade the answer; it is changing what can be thought next. We are already seeing versions of this pattern in AI systems generating hypotheses that humans test, machine learning finding structure inside experiments designed by humans, AI systems designing molecules that laboratories expose to physical reality, autonomous systems beginning to operate scientific instruments directly and human researchers using whatever comes back to decide what should happen next.

That variation matters because I do not want to use the word interaction so broadly that it eventually means everything. There is a difference between asking a model a good question and receiving a useful answer, a scientist repeatedly steering an AI through a difficult research problem, several AI agents challenging one another, and an autonomous scientific system manipulating laboratory equipment while responding to measurements from the physical world. Whether all of those belong to one underlying idea is something I want this series to investigate rather than assume, and a conversation with Claude sharpened that problem further by introducing the distinction between elicitation and emergence. Imagine that an AI already possesses the capability to solve a problem and a human finally asks the question that brings that capability out. The interaction matters, but perhaps all that happened was elicitation: the human discovered the doorway to something that was already there. Now imagine something different. A person introduces an observation the machine would not independently have pursued, the AI connects it to another field, the human recognises something inside that connection and rejects part of the reasoning, that correction changes the next machine response, a new hypothesis appears, an experiment changes the hypothesis again, and several passes later the process has reached an intellectual location that neither participant entered the interaction already holding. Is that still elicitation, or has something emerged from the relationship itself? I do not know yet, and I want to resist choosing the second answer simply because it makes the title more interesting.

There is also something slightly amusing about how the idea itself has developed. I originally brought a relatively simple proposition into a conversation with AI: perhaps when brilliant human insight meets machine intelligence, we may encounter another kind of frontier. That interaction expanded the thought, I later took the argument to another AI system specifically to see whether it could break it, and that conversation produced questions that returned to this research. Scientific cases then forced another change by making it obvious that laboratories, tools and the physical world could not simply be treated as background. None of that proves the thesis, because a conversation producing an interesting idea is not scientific evidence for a new theory of intelligence, but it does show why I want the series itself to remain interactive. I intend to keep taking the idea outside its own comfort zone. If another model offers a stronger objection, it belongs here. If a scientific experiment looks like perfect evidence for the thesis but a closer reading shows something much simpler happened, that belongs here too. If researchers eventually demonstrate that what appears to be human-AI emergence can be explained entirely as capability already residing inside the model, we should not hide from that because it damages a good title.

The experiments are already giving us reasons to be careful because some of the same scientific benchmarks where AI performs remarkably well also reveal sharp limitations. Models that can design extraordinary candidates in one biological task can perform poorly when the problem changes, computationally convincing molecules can collapse when synthesized, and AI systems can generate plausible scientific explanations that physical experiments reject. Putting AI into science is therefore not producing a neat story where machine intelligence progressively replaces every other participant. In many of the strongest examples, its failures are part of what makes the larger system useful because those failures expose where another form of intelligence or evidence is still required. That leads to another possibility I want to keep open: perhaps the frontier is not produced because every participant becomes sufficient, but because different participants are insufficient in different ways. The human cannot search everything, the machine may not know which observation deserves attention or whether its elegant prediction will survive contact with biology, the laboratory can test reality but cannot decide which of billions of possibilities deserve testing, and an instrument can measure what happened without necessarily deciding what the result means. Connect those limitations correctly and something becomes possible that none of the components could produce efficiently by itself, although whether we should call the resulting system a new form of intelligence remains another question entirely.

There are many of those questions ahead. If two scientists receive access to the same advanced AI system and reach completely different discoveries, what produced the difference? If an AI begins generating the questions as well as the answers, where does human judgement move? If several models argue and refine an idea before any person sees it, is the important interaction still human-AI? If an autonomous agent can operate laboratory equipment, receive experimental feedback and alter its own next experiment, does the physical environment become part of the cognitive system? And when something genuinely new is discovered, how would we demonstrate that the result emerged from the interaction rather than merely being retrieved from capabilities already contained somewhere inside it? Underneath all of those questions sits another one that may eventually become even more important: are we looking at the correct unit when we measure intelligence? We usually measure the scientist, then the model, compare one model against another and compare machines against human experts, but some of the systems now appearing in science make me wonder whether another unit deserves examination as well: the capability of the whole interacting system. Not because humans and machines should be blended together whenever it is convenient, and not because collaboration automatically creates something profound, but because there may be situations where analysing the participants separately causes us to miss part of what actually produced the result.

For now, I want to leave the proposition exactly where it began: unresolved. The frontier may exist entirely inside increasingly capable machines and interaction may simply be how we learn to access it. Human judgement may continue to determine where that intelligence becomes useful. Some capabilities may emerge only after different kinds of intelligence begin acting upon one another. Reality may turn out to be an essential part of that process, or we may eventually discover that we have been placing several unrelated phenomena under one attractive idea. I am comfortable not knowing yet because what matters now is that we have a question worth following, real experiments with which to test it and increasingly capable systems creating new versions of the interaction faster than any single paper could reasonably capture. So rather than begin with an answer, this series begins with an invitation to watch the evidence with me.

The frontier may exist in the interaction. Let us find out what that actually means.

— Teff