End-to-end is not the end goal.

There is a powerful idea shaping the development of AI for science: give an AI system a scientific problem and let it carry the work all the way through.

Generate the hypothesis. Search the literature. Design the experiment. Analyse the results. Write the paper. Move from question to answer with as little human intervention as possible.

This is an extraordinary technical ambition. It will produce important discoveries. Task-shaped AI systems that can solve defined problems, search enormous possibility spaces and automate difficult scientific workflows are going to change research.

But Lou does not think end-to-end automation should always be the implicit end goal.

Science is not only a process for reaching answers. It is also a process through which researchers learn what the better questions are. The route matters. The false starts matter. The strange paper found while looking for something else matters. The observation that does not fit the original hypothesis matters. The conversation with someone working on an adjacent problem matters.

If we optimise only for reaching a predetermined destination, we may become extremely good at getting scientists to the wrong place efficiently.

The helicopter and the hike

The mathematician Terence Tao has a useful metaphor for this.

Imagine that you want to reach a waterfall in the mountains. Traditionally, you walk there. The hike is slower and more demanding than simply arriving. But along the way you see wildlife, notice another place you might want to visit, learn the shape of the landscape and perhaps meet other hikers. Some of those encounters are accidental. They are not part of the stated goal of reaching the waterfall. They may nevertheless be the most valuable part of the journey.

AI can offer a helicopter. It can drop you directly beside the waterfall. You reach the destination quickly and get the photograph you came for.

But, as Tao argues, the destination was only part of the value.

This is not an argument for refusing the helicopter. Sometimes you need to get somewhere quickly. Sometimes the route is already well understood and walking it again adds little. Sometimes automating the journey frees people to work on more important problems.

The mistake is assuming that arrival is the only thing that counts.

Research goals are often written as if they are more precise than they really are. Test this hypothesis. Estimate this effect. Identify this mechanism. Solve this problem. Produce this output. A human researcher pursuing one of those goals usually encounters nearby questions at the same time. They develop intuition. They notice anomalies. They learn which assumptions are carrying too much weight. They find that the original question was badly framed or that a supposedly minor result is more interesting than the planned one.

An automated system given the destination has no necessary reason to value any of that. It can capture the flag without mapping the surrounding territory.

Task-shaped AI will transform science

It is important not to turn this into a weak argument against automation.

Task-shaped AI is already useful across science. It can search chemical spaces that no human team could inspect manually. It can help prove mathematical results, generate code, classify images, extract information from millions of papers and identify patterns in datasets too large for ordinary human attention. As these systems improve, they will take on longer and more complex workflows.

That is good.

There are also parts of research that should become substantially more automated. Repetitive data processing, routine checks, formatting, transcription, conversion between systems, standardised searches and other forms of avoidable labour do not acquire scientific value merely because a researcher currently has to do them by hand.

Tao himself is not arguing for untouched, pre-AI mathematics. He uses AI for literature search, code generation, calculations and proving lemmas that would otherwise consume time. His point is more careful: the best level of automation is not automatically 100 per cent. Some friction is artificial and can be removed. Some friction is the intellectual work through which understanding develops.

The challenge is knowing the difference.

Task-shaped AI is built around a bounded objective. That makes it powerful. The task can be specified, attempted, checked and improved. In areas where outputs can be verified reliably, highly automated systems may make discoveries that would otherwise have remained out of reach.

But a research career, a laboratory or even a single substantial project is not one bounded task. It is a shifting collection of ideas, evidence, collaborators, uncertainties, deadlines and decisions. Questions split into other questions. Findings change the interpretation of earlier work. A grant becomes a study; a study produces a methodological problem; the methodological problem becomes a paper; a reviewer comment exposes an assumption that changes the next grant.

That is a different shape of work.

Portfolio-shaped AI supports the journey

Portfolio-shaped AI is designed around that wider shape.

Rather than taking one task from beginning to end and then disappearing, it stays with the researcher across tasks. It understands the body of work they are carrying: the papers, proposals, analyses, notes, collaborators, unfinished questions and decisions that connect one project to another.

Its value is not that it removes the researcher from the process. Its value is that it reduces the cost of remaining deeply engaged in it.

A portfolio-shaped AI can help a researcher follow several lines of inquiry without losing the thread. It can search an unfamiliar adjacent literature and bring back the papers that challenge the current framing. It can compare a new idea with earlier grants and abandoned drafts. It can trace why a methodological decision was made and ask whether it still holds. It can surface a contradiction between what the team planned, what the code now does and what the manuscript claims. It can preserve a promising tangent that would otherwise be lost when the next deadline arrives.

This is not end-to-end automation. It is sustained intellectual collaboration.

This is the kind of relationship Colleague is designed to support: not a system that takes the whole research process away, but one that stays with the work across its changing forms. The individual tasks still matter. The difference is that they are held inside a continuing understanding of the research portfolio rather than treated as disconnected requests.

The researcher still chooses what is interesting. They still decide whether an anomaly is noise or a clue, whether a detour is worth following and whether an argument is scientifically honest. They remain responsible for the evidence and the claims.

But they can explore more.

The usual fear is that retaining a human in the loop makes the process slower. That assumes the only alternatives are a person walking one path slowly or an AI flying directly to the endpoint.

Portfolio-shaped AI offers another possibility: helping the researcher meander faster.

The journey remains a journey, but the cost of exploring each branch falls. A researcher can inspect several neighbouring literatures rather than one. They can test multiple framings before committing to a manuscript structure. They can return to an old idea with the relevant context already assembled. They can pursue a speculative connection without spending the first day reconstructing what they knew six months ago.

They do not merely reach one destination faster. They cover more of the landscape.

Discovery needs continuity

Serendipity is sometimes described as pure accident. In practice, useful scientific serendipity usually depends on preparation and continuity.

The unexpected paper matters because the researcher recognises its connection to a question they have been carrying. The odd result matters because they remember an earlier assumption that it unsettles. The conversation with another researcher matters because someone can connect it to the larger programme of work.

That means supporting exploration requires more than producing more options. An AI system that generates a hundred disconnected ideas may increase noise rather than discovery. The useful system has to understand which ideas connect to the work already underway, which assumptions they challenge and which earlier threads they reopen.

This is where the portfolio matters.

A portfolio-shaped AI does not treat each prompt as a fresh beginning. It can hold the evolving map: where the researcher has been, which paths were rejected, which remain open and how today’s question relates to everything else. That continuity makes iteration cumulative rather than repetitive.

It also makes disagreement more useful. A system that knows the wider work can do more than comply with the immediate request. It can notice that the new framing conflicts with an earlier decision. It can ask whether a convenient conclusion is actually supported. It can point out that the quickest route to finishing the paper may bypass the most interesting unresolved question.

The aim is not frictionless science. It is removing the friction that prevents researchers from engaging with the friction that matters.

We should build both

Science will benefit enormously from AI systems that automate bounded tasks and complete increasingly ambitious workflows. We should build them, test them and use them where their outputs can be responsibly verified.

But that should not become the only vision of progress.

A system is not necessarily more valuable because the human has been removed from a larger proportion of the workflow. Full autonomy is one design objective, not the natural endpoint of every scientific tool. In work where judgement, reframing, interpretation and connection are central, deep collaboration may be more valuable than maximum automation.

Task-shaped AI can take us rapidly to a destination. Portfolio-shaped AI can help us explore the landscape.

Science needs both.

The future should include AI that solves a well-defined problem while researchers sleep. It should also include AI that stays beside a researcher for years: remembering the abandoned idea that has suddenly become relevant, connecting a reviewer comment to the next study, opening several possible routes and helping the researcher decide which one is worth walking.

The goal is not to preserve inefficiency for its own sake. It is to preserve the parts of the journey where understanding, judgement and discovery are formed.

End-to-end is not the end goal.

The better goal is to help scientists travel further, notice more and return with a richer map.