From machine learning to the grand challenges

Why machine learning comes first

The company began in June 2025 with a narrow target: automate machine learning research and engineering before attempting anything whose feedback takes years. That choice is the whole shape of the sequencing that follows. ML is where the loop closes fastest, where success and failure are unambiguous, and where every improvement makes the next campaign cheaper — so the capability we need to attack a grand challenge is partly built by practising on our own stack.

Everything below is about what happens after that: which domains the loop travels to next, what each of them demands that we cannot supply yet, and why the honest unit of progress is a gate rather than a date.

Automation is a rate change, not a discovery

Most claims about AI accelerating science quietly assume that a faster loop is the same as an answer. It is not. What a faster loop gives you is more attempts per unit time, evaluated consistently — which is enormously valuable when the space of experiments is large and the cost of each trial is the binding constraint.

That is precisely the situation in drug candidate screening, catalyst discovery, materials characterisation, and large parts of ML engineering. It is not the situation in problems where a single decisive experiment is the whole difficulty, or where the bottleneck is regulatory rather than experimental. Being clear about which kind of problem we are addressing is part of being useful.

The five areas, and what each one actually needs

The translation gap

Every one of these areas contains an unavoidable distance between an in-silico result and a real-world consequence. Our loop can operate on the computational side of that boundary today; it cannot cross it, and we do not intend to pretend otherwise.

The honest formulation is that automation compresses the part of the pipeline that runs on computers — hypothesis generation, candidate prioritisation, experimental design, analysis, and the choice of which physical experiment to run next. Everything after that is a real experiment in the real world, with the cost, duration and uncertainty that implies.

This is still a substantial contribution. In fields where the computational stage is the long pole, compressing it changes what research programmes are feasible. In fields where it is already short, our impact will be correspondingly smaller, and we should say so rather than implying otherwise.

The sequencing we accept being judged against

Stated plainly, without dates attached to any of it, this is the order in which the work has to happen for the eventual claims to be credible.

1. Automate our own machine learning engineering. The loop closes on targets we control, with results we can verify ourselves.
2. Make it economical. Cost per verified result has to fall as campaigns scale, or nothing downstream is affordable.
3. Make it portable. Take one complete workflow — objective, harness, budget, review gates — and run it somewhere the hardware and the data are not ours.
4. Work with one partner in one domain. Choose a problem where the computational stage is the bottleneck and the partner can check our work independently.
5. Only then broaden. Additional domains inherit a method that has already been wrong in public at least once.

The temptation is to skip to step four, because that is where the compelling story lives. Steps one to three exist to make step four mean something when it arrives.

The evidence ladder

We think about progress outward in terms of what has been demonstrated, not what has been announced. Each rung has to be reached before the next is discussed seriously.

A year into the company, we are working on the first two rungs. That is not a modest claim dressed as humility; it is simply where the evidence is.

Why we publish gates instead of dates

Dates are easy to write and impossible to defend, and in this field they are the single most common form of overclaiming. A date implies knowledge of obstacles not yet encountered and of the efficiency of solutions not yet built.

Gates are less satisfying and more useful. Each one describes something observable that would have to be true, and each one can be checked by someone outside the company. When a gate is missed, we have to say so — which is exactly the accountability a date quietly avoids.

• Gates replace milestones in our roadmap; missing a gate is reported, not quietly rescheduled.
• Any external claim is tied to the strongest evidence level it has actually reached.
• We do not describe a computational result as a medical, energy or environmental one.

What would actually convince us we are on the right track

Not a benchmark score, and not a demo. Three things: our own campaigns produce improvements that survive independent re-running; the cost per useful result falls as campaigns scale; and a partner research group, using our methods and verifying them independently, reaches a conclusion faster than they would have otherwise.

The third is the only one that cannot be manufactured internally, which is why it is the gate we are most interested in passing. If the loop is only convincing to its builders, it is not yet a scientific instrument — it is a very expensive belief.