AI researcher

Writing the training code, the evaluation harness and the infrastructure that lets an idea be tested at all.

Diagnosing why a run died, why the cluster is idle, why the numbers from two machines disagree — the plumbing between an idea and a result.

Watching a long agent task, noticing it has gone the wrong way, and stepping in — repeatedly, before the result is worth anything.

Deciding whether the number moved for the reason you think, whether it will hold at a larger scale, and whether it is worth anything outside the benchmark.

Picking which direction is worth a quarter of a team's compute, and which promising thing to stop.

Deciding that something must be paused, restricted or not shipped — and being the person who says so while the work is going well.

One laboratory's disclosure about its own research organisation, with measurements running to mid-August 2026, and the laboratory sells the agents it is measuring — the numbers and the framing both point the same way, which is the direction that suits the seller. Read three limits with it. The 3.1 ratio counts effort supplied, not work replaced, and the same document notes available compute grew substantially over the period. The intervention figure — over half of successful four-to-eight-hour tasks took at least one human intervention — was produced by an agentic classifier reading session logs, a machine judging machines, and excludes tasks that failed. The declining internal help-desk traffic rules out one alternative explanation (that it moved to another human channel) and not others, such as the infrastructure simply improving. Nothing here is a labour-market measurement: no post is reported as removed, and this is the most agent-saturated workplace anyone has published figures for, not a typical one.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

A signed essay, not a measurement, and that distinction is the entire reason this record exists in a stage that can change nothing. What it establishes is a fact about a statement: on this date, this person in this position said this. Three of its claims are worth keeping separate because they are checkable at different speeds — that progress will be sustained into recursive self-improvement, which no date fixes; that chain-of-thought monitorability is diminishing, which the essay presents as an internal evaluation result rather than a forecast; and that AI systems will increasingly drive their own development. Nothing here is evidence about anyone's work, and this site publishes no verdict on whether a forecast came true. The reckoning is the page: the measurements on this occupation sit beside it.

A named person with standing publicly predicted something, on a date, in an attributable statement. It is recorded so that who said what, and when, stays checkable — and it never moves a task's assessment, because a prediction is not an observation. Its value arrives later: the record sits on the same page as the evidence about that occupation, so anyone reading the forecast reads the record of what happened next beside it. That is the reckoning; this site publishes no verdict on whether a forecast came true.

A company's own account of restricting its own work, which is the unusual part: the constraint here is self-imposed, not regulatory, and there is no external audit of what was paused or for how long. Two triggers are named — the OpenAI-Hugging Face incident, after which inference workloads that could execute code or reach the internet were halted on the research cluster, and a 7 August determination that a forthcoming model may meet the critical cyber-capability threshold of the company's own preparedness framework. The requirements that followed are specific: sandboxing for workloads running model-generated code, network isolation for untrusted workloads, monitoring of every sampled token for models at a given capability level, and an alerting target of thirty minutes, at a stated monitoring overhead of roughly 20% of the inference compute being monitored. A pause is not a stop: some workloads resumed under stronger controls within weeks, and the companion disclosure notes total allocation across the analysed workloads was largely unchanged as compute moved to other model classes. It says nothing about any laboratory other than this one.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

The part of the job you are being trained for — writing the training loop, building the harness, getting the cluster to behave — is the part with a published number on how far it has already been handed to agents. That is not a reason to skip learning it, because you cannot steer a four-hour agent task through work you have never done yourself. It is a reason to be honest about what it buys you: it is the entry ticket, not the position. The position is being the person whose read on whether a result is real gets believed, and the fastest way toward it is to go and be wrong about a result in public, early, where someone senior will tell you why.

Your leverage moved from how much you can build to how many concurrent lines of work you can hold in your head well enough to notice one going wrong. That is a real skill and it is not the one you were hired for. Two things are worth watching in yourself. The first is whether you still read the logs of runs that succeeded — the intervention data says the failures are visible and the quiet wrong answers are not. The second is what happens to your judgement when you stop writing the code: the people who can tell a real result from an artefact are, so far, the people who built enough of the measuring apparatus to know how it lies.

Stay and strengthen

Own the evaluation, not the model

When writing the experiment is delegated and reading it is not, the scarce position is the one that decides what counts as a result. Evaluation design is also the part the field's own disclosures name as lagging: capabilities that are easy to measure improve faster than the ones that are hard, which is a statement about the measuring instruments being the bottleneck.

Reshape the role

Become the one who runs many lines at once

The measured shift is toward researchers holding several concurrent agent sessions, and the reported failure mode is not the agent being incapable but the person not noticing where it went wrong. That is a supervision skill with no established training route, which is exactly why holding it early is worth something.

Adjacent move

Move to where the constraint is the judgement

Fields where an experiment is expensive, slow or irreversible — biology, materials, medicine, hardware — cannot delegate the run to an agent, so the whole value sits in choosing which experiment to do. The method skills transfer; the cheapness that made this occupation's build phase delegable does not exist there.

No year, and in this occupation a year would be especially dishonest, because the people best placed to know disagree with each other in public. What you can measure yourself is more useful. Over your last ten pieces of work, count how many ended in code you wrote and how many ended in a judgement about work something else produced. Then count, among the agent tasks that succeeded, how many you had to interrupt. The first ratio tells you which half of this job you currently hold; the second tells you whether the half you hold is still load-bearing. Both numbers exist in your own week and neither requires anyone's forecast.

Treat the statement and the measurement as two different things, because they are published in different documents and only one of them is checkable. The measurements — agent-workdays per human workday, intervention rates, the share of output that is high-level planning — are specific, dated, and could be contradicted by a later disclosure. The expectations about what comes next are expectations: they come from people with unusual visibility and an unusual stake, both of which cut in the same direction. This site records the first kind and not the second, and that is a deliberate limit rather than a view about who is right.

The question is usually asked as if the PhD were training in building things, and it is mostly not. What a PhD is actually four years of practice at is the task on this page marked as staying with people: deciding what is worth investigating, and defending a claim about a result in front of people trying to break it. That is the scarce half. The honest caution is different from the usual one: the cheap half of the job is where a student now proves themselves to a hiring committee, and if that is all a programme trains you for, the training and the market have drifted apart. Ask the group you are joining how often a student's result gets taken apart in front of them, and how recently a direction there was abandoned.

No, and the site keeps them apart on purpose because their exposure differs. A machine learning engineer delivers a system that runs on real people, and the thing that displaced much of that work was a purchasable substitute — a model somebody else trained. A researcher delivers a conclusion, and what is being delegated there is the labour of producing evidence for it, not the conclusion. One consequence matters for anyone choosing: the engineer's displacement is reversible, because a vendor's price or licence can move; the researcher's is not that shape at all. If your work ends when something ships, read the machine learning engineer page instead.