Careers
Help build the other half of AI.
Models can reason. Keeping them reliable over weeks of work, as the facts change, is the other half, and most of it is unsolved.
Open problemsKeeping decisions current when evidence changes
Which evidence should govern the next decision, and what must change when that evidence changes?
Make agents revise every decision a changed fact affects, and nothing else. Measure how reliably they repair the affected work and keep what remains true.
Keeping what a model learns on the job
How can a model keep what it learns on the job through interruptions, restarts and new models?
Find out why a capable model loses command of a task partway through, and restore it without damaging anything else the model can do. Then map where the restored command holds and where it breaks.
Replacing a model without restarting the project
What does a replacement model need from its predecessor to take over unfinished work?
Help a stronger successor take over unfinished work without losing what its predecessor knew. Judge the handover by the decisions the successor makes, even when it comes from a model family that reasons differently.
The Valley: abilities that outlive their agents
Can a population develop abilities that survive its members and exceed what any member can do?
Study how groups of agents in a persistent world acquire conventions, divide labour and build shared methods. Find out whether a capability survives the agents who developed it, and whether the group finishes work no member could.
Making experience improve the next AI
Can a system use the methods it has learned to improve the AI that finds its next methods?
Turn experience into lasting gains: better tools, policies and specialist models. Show which of them hold up on unseen work, keep the system’s earlier abilities and repay what they cost to find.
The cost of finishing the work
Which system finishes useful work for the least total effort, including the work people must repeat?
Measure what finished work costs across a whole project: productive compute, retries, recovery, model changes and the human effort of rebuilding decisions. Compare systems on verified useful work per unit of total cost.
Who we want to hear from
We want to hear from researchers and engineers working on long-horizon agents, continual learning, multi-agent systems, evaluation, or how models keep what they learn on the job. You may have measured where an agent degraded over weeks of work, built recovery for a system that failed mid-task, or trained models to learn from one another.
Tell us where it broke, the experiment you ran and what happened next. Work that changed your mind is useful here. Everyone works in person: the lab team in London, the growth team in San Francisco.
How to write to us
Pick a problem on this page and write to contact@qwntl.com. Tell us what you’ve built or investigated, the part you’d own and the first question you’d test.
Include a paper, repository or short account of your work if it helps explain the idea. We review every message.