Product releases, research papers, and policy updates in one feed. When a claim in an earlier post turns out to be wrong, we edit the original and say so rather than quietly republishing.
Two models built for work that runs across days rather than turns. The accompanying system card reports a 9-point drop on one refusal benchmark after we rebuilt a contaminated evaluation split — the earlier figure was wrong and we withdrew it.
New interpretability work localising the mechanism that suppresses an answer under uncertainty. We include the three steering experiments that did not replicate, because the failures are the useful part.
Capability tiers now carry named safeguards and a defined sign-off path, including what happens when a threshold is crossed partway through a training run.
We published the methodology, the sampling frame, and the known weaknesses of our task-level exposure measure, along with an open call for researchers to contest it.
Credits, evaluation tooling, and engineering time for academic groups working on safety-adjacent problems, with no requirement that results be favourable to us.
Instruction adherence degrades measurably past a horizon our tooling had not previously instrumented. Six mitigations tested, two of which held up.
Rather than report problems only when asked, we maintain a running public log of safety-relevant incidents and what changed as a result.
Repeated context now costs a fraction of a fresh call, and batch jobs run at reduced rates with a 24-hour completion window.
Nothing matches that filter yet.
For interviews, comment on published research, or brand assets, contact the communications team. We answer on the record, and we will tell you when we cannot.
A short monthly digest of what we published, what we got wrong, and what we are working on next. No drip campaigns.
This is a static demo — the form does not send anything.