Questions, answered plainly
The questions we actually get asked — including the awkward ones about autonomy, failure and what we are not claiming. If yours is missing, write to us.
About AutoScientia
What is AutoScientia?
An autonomous AI company. We build systems that run the entire experimental cycle — propose, run, evaluate, iterate — so that scientific and engineering progress stops being limited by how many experiments a person can shepherd by hand.
When was AutoScientia founded and where is it based?
AutoScientia was founded in 2025 and is based in Germany. Our registered website is autoscientia.com.
What does “autonomous experimentation” actually mean?
It means the system owns the whole loop: forming hypotheses, designing experiments, executing them in parallel, scoring results against an objective fixed in advance, and letting those results decide what to run next — without a human in the middle of every turn. Our glossary defines the terms we use.
Are you a research lab or a company?
Both, deliberately. We do research on how automated experimentation should be built and evaluated, and we do the engineering to make it run reliably. The loop is the product; the papers come second.
Who leads AutoScientia?
Our founder and the story behind the company are on the About page, which is the right place for a question about people rather than technology.
How large is the team?
Small on purpose. We stay deliberately compact so that decisions about the loop are made by the people who build it — and we hire when a role removes a real bottleneck rather than to fill an org chart.
What does the name mean?
Scientia is Latin for knowledge, and the prefix needs no translation. The name states the ambition: knowledge that advances itself rather than waiting on human throughput.
The technology
How does the loop work, step by step?
Four moves: propose candidate experiments against an objective, run them in parallel across large-scale compute, evaluate each result automatically against a frozen baseline, then iterate from what survived. The approach page walks through it; technology lists the eight operational steps behind a single cycle.
Which models do you use?
We build on frontier models as proposers rather than training a foundation model from scratch — that is not where the bottleneck is. On top of them sit our own fine-tunes and, more importantly, the surrounding system: the search logic, the scheduler and the evaluation harness do most of the work.
How many experiments can it run?
The design target is thousands of trials in parallel per cycle. The real number depends on the objective, the size of each trial and the compute available — we would rather quote a measured throughput than a marketing figure.
How do you stop the loop from fooling itself?
Objectives and failure conditions are recorded before a cycle runs. Baselines are frozen. Promoted candidates are re-run across seeds. Proposals are scored against objectives the proposer never sees, and every discarded trial is logged with the reason it lost. Nothing gets reported that cannot be replayed from the run record.
What about hallucinated ideas?
We expect them, and we plan for them. The proposer only suggests; measurement decides. A hallucinated experiment that cannot be built, or that fails, costs compute and a line in the log — not correctness. The harness is what protects the result, not the model's confidence.
Do humans stay in the loop?
Always, at the points that matter: objectives, budgets, stopping conditions, and any effect outside the sandbox. The loop takes the repetition and the scale, not the judgement.
Does the system improve itself?
It is aimed at our own stack first — models, pipelines, infrastructure, tooling — which is the fastest way to make the loop better at everything else. Improvements are reviewed and shipped like any other change; there is no unattended self-modification of the system that approves its own work.
Safety and responsibility
What stops the loop from doing something harmful?
Structural controls rather than good intentions: isolated execution with no ambient credentials, restricted network egress, enforced compute and spend ceilings, an append-only audit trail, and a human gate on anything that has an effect outside the sandbox. The safety page lists all six.
How do you handle dual use?
By screening objectives before compute is committed and declining work that fails the screen. Security-related experiments only ever run against systems we control. Where a method would be more useful to a misuser than to a defender, we publish the defence instead of the method.
What happens when something goes wrong?
It becomes an incident: contained, documented, and fixed at the cause rather than the symptom. Budget overruns, unexpected data access, drift in the loop and results that fail replication all follow the same path.
Are you worried about AI risk?
We take the concrete risks seriously: objective misspecification, reward hacking, goal drift over long campaigns, and the misuse of general experimentation capability. We treat them as engineering constraints and test for them. We do not claim to have solved alignment, and we would be suspicious of anyone who does.
How do I report a vulnerability?
Write to security@autoscientia.com. Good-faith reports are welcome, acknowledged, and never met with threats — that is a commitment, not a courtesy.
Working with us
How do we start a collaboration?
Email research@autoscientia.com or use the contact form with three things: the objective you care about, what a single iteration costs you today, and why the cycle cannot simply be parallelised by hand. Those three answers tell us whether automation would help at all.
Can you run experiments on our data?
Often, yes — with classification on entry, an egress policy per dataset, and a written agreement covering handling and deletion. Trials only ever see the data their objective requires. The privacy policy covers personal data; project data is governed by the contract.
Do you have a product or API I can use?
Not publicly. We are in the first phase: proving the loop on our own stack. When something is genuinely useful to outsiders, it will be described here before it is sold.
Do you take investment or acquire companies?
Nothing on this website is an offer, a solicitation or investment advice, and we do not run fundraising conversations through a contact form. If you have something concrete and relevant, hello@autoscientia.com reaches a person.
Who handles press enquiries?
press@autoscientia.com. We are happy to explain the work or correct how it has been described.
Where are you willing to work?
We are based in Germany and collaborate across Europe first, with partners elsewhere when the measurement access is worth the distance. Physical proximity matters most where hardware and lab work are involved.
Careers
Are you hiring?
Yes — current openings are listed on the careers page, with the hiring process written out in full.
Can I apply if nothing matches my profile?
Yes. We would rather meet the right person than fill a slot: email careers@autoscientia.com and show us something you built. Links beat adjectives.
Do you take interns or students?
We take on a small number of research collaborations with students whose interests match the loop itself. Write to careers@ with what you would want to work on, not a general request.
Is remote work possible?
Roles are based in Germany. We are a small team working on coupled systems, and the honest answer is that we prefer real proximity for the core team, with flexibility where the work allows it. We say so up front rather than after three interview rounds.
Data and this website
Does this website use cookies or analytics?
No. No cookies, no analytics, no advertising pixels, no third-party trackers. Fonts are loaded from Google Fonts, which necessarily sees the request your browser makes to fetch them.
What happens when I use the contact form?
The form composes an email in your own mail client. Nothing is transmitted to a server of ours, and no data is stored by this website. If you send the email, we use your details only to answer you.
How do I exercise my data protection rights?
Write to privacy@autoscientia.com. Access, correction, deletion, restriction, portability and objection are all described in the privacy policy, along with how to contact a supervisory authority.
Where can I read the rules of using this site?
The Terms of Use cover permitted use, intellectual property, disclaimers and liability. Provider information is in the Impressum.
Still unanswered? Ask directly
A specific question gets a specific answer. Vague ones get a friendly reply and a link back to this page.