AutoScientia builds autonomous AI that runs the whole experimental cycle: propose, run, evaluate, iterate. Work that has always taken weeks of human attention becomes thousands of runs executing side by side.
Automating machine learning research and engineering first — then the grand challenges.
Illustration of a cycle in progress. The figures shown are examples, not measurements.
Automating discovery is not one invention. It takes a proposer that can write the next experiment, infrastructure that can run a thousand of them side by side, and a way of scoring the result that resists wishful thinking. All three arrived at roughly the same time, and a loop needs all three — remove any one of them and there is nothing to build.
A frontier model can read a result, form a view about why it failed, and write the next experiment in a form a scheduler can execute. That step used to be the bottleneck of the whole method; it is now the cheapest part of the loop.
Parallel infrastructure turns a hundred questions into a single wait. The economics of an experiment change completely when the thousandth trial costs little more than the first, and when failure is cheap enough to be worth trying.
A faster loop is worthless if it measures the wrong thing. Frozen baselines, objectives the proposer never sees, repeated runs and a taxonomy of failures are what let volume mean something instead of merely increasing it.
Put the three together and the change is not a better tool for scientists — it is a different rate of experimentation, and the rate has been the constraint all along. The argument is in Why science is still hand-cranked.
The scientific method still runs on repetitive, sequential human experimentation. One hypothesis, one experiment, one wait: every result gated behind the last, every insight costing weeks of attention.
That structure does not scale, and the problems worth solving — medicine, energy, water, security — will not wait for it. Automating the loop is the only way we know to change the rate at which answers arrive.
Sequential by design
One experiment at a time. Every answer waits for a person to design the next question.
Parallel by construction
Thousands of experiments in flight, scored automatically, feeding the next round.
AutoScientia runs the whole cycle without a human hand-off in the middle — hypotheses in, evaluated results out, then straight back around.
Frontier models draft hypotheses, candidate designs and experimental plans against a measurable objective.
Frontier modelsExperiments execute in parallel across large-scale compute, so a thousand questions cost one wait.
Parallel computeEvery run is scored automatically against the objective, with results compared to the existing baseline.
Measurable outcomesThe strongest candidates are promoted, priors updated, and the loop starts again — by itself.
No hand-offFour stages describe the cycle, though a single pass involves rather more than four steps. The plainest items on the list below are the ones that make a result worth believing — each of them is here because a loop without it produces confident numbers nobody can defend.
The proposer returns candidates ranked by expected value against cost. Anything without a measurable prediction attached is rejected before it consumes a single GPU-hour.
Ranked by valueCompute, wall-clock and spend limits are attached to every trial before it runs, and the scheduler refuses work that would exceed them. Hitting a ceiling is a normal outcome, not an incident.
EnforcedEach trial runs in a fresh container with no shared writable state, narrow credentials and outbound network denied by default. Two experiments cannot contaminate each other's results.
SandboxedResults are scored automatically against a baseline pinned for the whole campaign, and each run records the data version it actually read rather than a floating reference to “current”.
Held outA candidate is re-run before it is promoted. A single favourable draw is not evidence, and reporting the mean without the spread would hide exactly the thing worth knowing.
Repeated seedsCode version, configuration, cost, failure classification and score are written once, append-only. If a number cannot be rebuilt from its record, it does not get reported — internally or otherwise.
Append-onlyBefore chasing someone else's problem, AutoScientia uses its automated ML research and engineering to improve its own stack — the compute it trains on, the pipelines it feeds, the models it runs, and the products around them. Every cycle makes the next cycle faster.
We start where the loop can be measured end to end — machine learning research and engineering. The same machinery is aimed, next, at problems where progress is measured in lives, gigawatts and clean litres.
Faster search across hypotheses, assays and designs — compressing the distance between an idea and a result that matters.
Long horizonMaterials, chemistries and control problems explored in parallel instead of one lab bench at a time.
Materials & gridsSeparation, filtration and treatment candidates evaluated automatically against real constraints.
Applied scienceAdversarial search at machine speed — finding and closing weaknesses faster than they can be repeated.
Autonomous searchInstruments, simulators and evaluation harnesses that make the next experiment better than the last.
CompoundingEvery cycle is scored against an objective, so progress is a number you can inspect rather than a story you have to trust.
How we measureEvery claim this company makes names the level it reached. Most things stay at the first two for their whole life, and that is the correct outcome: a pipeline in which everything reaches the top would be unselective rather than rigorous.
A hypothesis, a design or a configuration has been generated. Describing it is not evidence of anything.
IdeaA run completed and produced artefacts. This proves the machinery worked, not that the idea was good.
ArtefactThe run was scored against the objective, on a frozen baseline, attributed to the data version it actually read.
ScoredThe result survived re-running under different seeds and agreed with an objective the proposer could not see.
Reproducible“A faster loop that measures the wrong thing does not produce discovery faster. It produces confident nonsense faster — and it does so with an audit trail that looks reassuring.”
That sentence is the reason this website spends more words on measurement, budgets and failure classification than on models. The proposer is the part everyone asks about; the harness is the part that decides whether anything the proposer finds is true.
It is also why the company is deliberately early. We are building the instrument before using it on a problem where being wrong is expensive — machine learning engineering first, and the grand challenges after the method has been shown to work somewhere we control.
A date is the easiest claim to make in this field and the least likely to survive contact with reality. We publish conditions instead — observable things that would have to be true, each checkable by someone who does not work here.
Automating discovery takes more than a model in a chat box: it takes the silicon the experiments run on, the infrastructure that schedules them, the models that propose them, and the products that make the results usable.
We work across all four deliberately: an advantage in one layer is spent to improve the next, and the loop is what spends it.
Search that reaches down to kernels, precision and scheduling instead of stopping at hyperparameters.
Thousands of concurrent experiments, with queueing, checkpointing and evaluation that hold up under load.
Frontier models generate hypotheses and experimental plans that stay worth running.
End-to-end products so the output of the loop is something a team can actually act on.
Each of these deserves a straight answer rather than a slogan, including the one about results.
No. A sweep runs a list somebody wrote and learns nothing mid-flight. The distinction that matters is iterate: results have to change what gets proposed next. If deleting the evaluation step would not change the system's behaviour, it is a search, not a loop — and we test that by ablation rather than by assumption.
It does not have to. People choose which problems deserve a campaign, write the objective, set the budget and decide when a system may act outside its sandbox. The loop owns repetition and scale; judgement and authority stay with people. That split is the design, not a stage we are passing through.
The risk is real and it is why the outer boundary is hard-coded: sandboxed execution, enforced budget ceilings, deny-by-default egress, append-only records, and a human approval for anything with an effect beyond the environment. The safeguards are described in public — safety, security — and they are open to challenge.
There are none to publish yet, and we would rather say that plainly than dress up a demonstration. What exists is a method, a harness, an orchestration layer and a list of ten public gates with their real statuses. When a result reaches verification it will be published with its methods and its failures — and until then, our own stack is the customer.
Four doors, in the order most people find useful.
Silicon, orchestration, proposers and evaluation — the four layers, and what each one has to get right.
Explore the stack →Ten public gates, their real statuses, and the evidence level behind each one. Updated monthly.
See where we are →Seven essays written between June 2025 and September 2026 — the thesis, the loop, the measurement and the horizon.
Read the essays →Research collaborations and the five criteria that decide whether a problem is worth automating.
Work with us →Research collaboration, engineering partnerships, or a role on the team — start with a conversation.