Research that has to survive contact with a model in production.

Opening the model instead of guessing at it

A model that produces good answers for reasons nobody can state is a model nobody can audit. Our interpretability work tries to replace that uncertainty with structure: circuits, features, and interventions that can be tested rather than narrated.

The practical test we hold ourselves to is whether a finding changes behaviour. If we locate a mechanism, we should be able to steer it, break it, and predict what happens when we do — on tasks the analysis never saw.

• Circuit-level tracing of behaviours we can name, such as refusal and hedging.
• Feature extraction that stays stable across model sizes and checkpoints.
• Steering experiments with pre-registered predictions, published whether or not they hold.
• Tooling so that an outside researcher can reproduce a finding without our help.

If an explanation cannot be falsified by an experiment, it is a description of our own confidence rather than of the model.

Making behaviour match stated intent

Alignment, as we use the term, is not politeness. It is the gap between what a system was asked to optimise and what it actually did — and the gap widens as tasks get longer and the feedback gets sparser.

• Training objectives that stay honest when a task has no single right answer.
• Honesty and calibration under pressure, including where a model was trained to be agreeable.
• Scalable oversight: keeping a human meaningfully in the loop as outputs outrun review capacity.
• Behavioural drift over long agentic runs, measured rather than assumed.

What has not worked

Preference tuning alone does not close the gap; it teaches a model which answers get approved, which is a different thing from which answers are correct. Several of our published negative results exist because that distinction kept catching us out.

Infrastructure as a research instrument

The bottleneck in a lot of safety work is not the idea, it is the cost of testing it. We invest in the plumbing that turns a hypothesis into a run within a day: training infrastructure, evaluation harnesses, and reproducibility guarantees.

• Deterministic replay for evaluation runs, so a score can be re-derived months later.
• Contamination checks that run continuously rather than once before publication.
• Cluster scheduling that keeps a small experiment from waiting behind a large one.

Assuming the model is being probed by someone patient

Red-teaming here means systematically looking for the failure that survives a demo: prompt-injection paths through retrieved content, tool calls that escalate privilege, and evaluation setups that a model can recognise and perform for.

• Threat models written before a capability ships, not after an incident.
• External red teams with publication rights and no veto from comms.
• Findings fed back into the scaling policy rather than only into a slide deck.

How this is landing in actual work

Aggregate statistics hide the thing worth seeing: which tasks are being handed over, in which occupations, and whether the handover is augmentation or substitution. We publish a recurring index built from anonymised, aggregated usage data, with a documented methodology and an open invitation to criticise it.

• Task-level rather than job-level measurement, because jobs are bundles.
• Regional and sector breakdowns, with sample sizes disclosed alongside the numbers.
• Survey work covering displacement risk, not only adoption rates.

Where models help most is where the search space is boring

The wins we can point to are rarely a model having an insight. They are a model doing the tedious middle of a research programme — screening candidates, drafting protocols, reconciling literature — so that a person reaches the interesting question sooner.

• Biology and materials discovery, with wet-lab validation as the arbiter.
• Literature synthesis where citations are verified against the source.
• Instrumentation work aimed at labs that are not already well funded.

Studying effects while they are still soft

Most technology's social effects are documented decades after the design decisions that shaped them. We would rather run the study while it can still change what we build, which means publishing findings that are unflattering to our own product.

• Longitudinal panels on trust, reliance, and deskilling.
• Access questions: who can use this, in which languages, at what price.
• Independent grant funding for researchers outside the company.

Recent papers

Preprints and peer-reviewed work, listed newest first. Code and data are linked where they can be released.