The approach

We automate the cycle, not just the typing

Scientific progress is bottlenecked because the scientific method still runs on sequential human experimentation. AutoScientia closes that loop with autonomous systems that propose, run, evaluate and iterate: thousands of experiments at once instead of one at a time.

Inside a cycle See the roadmap

01 · The cycle

Four moves that used to take four weeks

A cycle starts with an objective — a metric, a constraint, an engineering goal — and a space of candidate designs. From there the loop runs on its own: experiments are proposed, executed in parallel, scored, and the survivors become the starting point for the next round.

Humans stay in the loop by setting the objective and the guardrails. The repetition — the part that never scaled — is what we automate.

01 Propose
02 Run
03 Evaluate
04 Iterate

Propose

Frontier models turn an objective into hypotheses, designs and experimental plans, with the search space declared up front.

In: objective, constraints, prior results
Out: ranked candidate experiments

Run

Candidates execute concurrently across large-scale compute, so the wall-clock cost of a thousand experiments approaches the cost of one.

In: ranked candidates
Out: raw results, logs, artefacts

Evaluate

Results are scored automatically against the objective and compared to a fixed baseline, with failures recorded rather than hidden.

In: raw results
Out: measured deltas, uncertainty

Iterate

Winners are promoted, priors updated, and a new round begins — without waiting for anyone to book a meeting.

In: measured deltas
Out: next round of candidates

People keep

Direction and judgement

Choosing which problems are worth solving.
Defining objectives, constraints and safety guardrails.
Interpreting what a result means for the world.
Deciding when a system may act outside its sandbox.

Machines take

Repetition and scale

Generating and prioritising the next experiments.
Running thousands of trials in parallel.
Scoring, comparing and logging every result.
Restarting the loop the moment a round finishes.

02 · Infrastructure

What "in parallel" actually requires

Running one experiment is easy. Running thousands of them continuously, without losing results or burning compute, is a systems problem — and it is most of the work.

Scheduling at scale

Queueing, batching and placement so thousands of concurrent runs share a finite cluster sensibly.

Checkpoint & recover

Long runs survive preemption. Failed trials are retried with the reason recorded, not silently dropped.

Evaluation harnesses

Automated scoring against held-out objectives — the part that makes an autonomous loop trustworthy.

Cost & safety gates

Budgets, resource ceilings and isolated execution, so an autonomous loop cannot spend or reach beyond its remit.

03 · Our own first customer

Machine learning research is where the loop learns fastest

We start with machine learning research and engineering because the feedback signal is immediate and honest: a run either improves a measurable objective or it does not. That makes it the right place to prove the loop — and it is also the stack we depend on ourselves.

Recipes, architectures and training pipelines searched against held-out objectives.
Systems work — kernels, precision, memory, scheduling — tuned by measurement.
Data curation and evaluation harnesses improved by the same machinery.
Every gain is spent on the next cycle, so the loop gets faster as it runs.

Talk to us about it

Candidate ranking · example

Illustrative

Ranked by expected valuerequired

Scored on a frozen baselinerequired

Re-run before promotionrequired

04 · Rigor & safeguards

Autonomous does not mean unaccountable

An automated loop is only as useful as the trust you can place in its results. These are the rules we hold ourselves to while the system runs.

Measured against fixed baselines

Every claim the loop produces is scored against a baseline that does not move during a comparison. Improvements are reported as deltas, with uncertainty where it exists.

Reproducible by default

Runs are versioned — code, data references, configuration and environment — so a result can be replayed rather than believed.

Human gates outside the sandbox

The loop is free to iterate inside isolated infrastructure. Anything that would leave it — external side effects, spend beyond budget, published claims — requires a person to approve it.

Failures stay visible

Negative results are logged with the same care as wins. A search that quietly discards its failures is not a search we would trust.

A working position, not a marketing claim. We publish this list so it can be argued with. As the systems mature, our safeguards will have to mature with them — and we would rather say that now than pretend the question is settled.

05 · Roadmap

From our own stack to the grand challenges

We are deliberately sequencing this. Each phase has to earn the next one by producing measurable results at the phase before it.

Automate machine learning research

The current focus: autonomous search across model design, training recipes, data curation and evaluation — with results measured end to end on our own infrastructure. This phase exists to make the loop demonstrably better than a manual one.

In progress

Automate machine learning engineering

Carrying the loop from research into engineering: systems, kernels, scheduling, tooling and the products built on top of them. This is where the loop starts improving the platform it runs on without a human in the middle of every change.

Grand challenges beyond machine learning

With the loop proven where outcomes are measurable, we aim it outward: medicine and health, clean energy, clean water, cybersecurity and the tools of discovery themselves. These problems share a shape — cycles of experimentation that no human team can run fast enough by hand.

Long horizon

Say it plainly

What this is not

Autonomy is a direction, not a finished state. Being clear about the gap keeps the work honest.

Not a chat interface wrapped around a model — the loop, the compute and the evaluation harness are the product.
Not a claim that discovery is solved today. We automate cycles that can be measured, and expand from there.
Not a system that grades its own homework: objectives, baselines and guardrails are set outside the loop.
Not a black box. Results come back with the runs, configurations and failures attached.

Have a loop worth automating?

Tell us about the experiment cycle that is slowing you down. We are looking for problems where progress can be measured.

Start a conversation Who we are