Definition
hy·per·pri·or
/ˈhaɪ.pəˌpraɪ.ə/ n. 超事前分布
- 1. stats. A prior distribution on the parameters of another prior.
- 2. A small private research group in London. It builds environments and graders for training and scoring agents, and it studies how learning machines and complex systems come to hold beliefs and act on them.
Origin From the Greek hyper, over, and the Latin prior, earlier. A prior on a prior, and so a way of writing down how sure one is about how sure one is.
See also prior, posterior, calibration, hyperparameter
prior stats. A probability distribution over an unknown quantity, written down before the data are seen.
posterior stats. The distribution over an unknown quantity after the data are seen, which is the prior updated by the evidence.
calibration stats. The agreement between how confident forecasters say they are and how often their forecasts prove right.
hyperparameter stats. A parameter of a prior distribution, which a hyperprior in turn treats as uncertain.
The question
How sure are you about how sure you are?
Asked of a model, it is where its beliefs came from and how it would notice the world leaving them behind. Asked of an agent, it is what its weights encode and how far that holds once it acts. Asked of a human, it is the set of implicit or explicit beliefs that underpin our actions and feelings.
Research agenda
Our focus is mostly on reinforcement learning and world models. We study what emerges when many learners interact and adapt to one another in synthetic and complex adaptive environments.
Research interests subject to change without notice. Beliefs likewise.
Engagements
When time allows we can selectively take on a few outside engagements to look at deeply technical items.
For each one we put together a small, hand-selected team from our network: senior people who still write the code (or know how to extract it efficiently from agents), and engineers from trading firms and AI labs. We like to operate at the tip of the spear, and are usually the first people in.
There is no account team, so you talk to the people doing the work. We get results, and quickly. Sometimes our methods are deemed unusual, but they are highly effective.
-
A written opinion on a team’s systems, protocols, data and claims, from people who have built the same kinds of thing. It says what is real and what is fragile, in time for the decision, and it covers AI teams, quantitative ones and the teams bringing digitally native assets into the rest of the financial system.
-
A plan with the hard parts named and costed and the first of them started, a direction specific enough for your engineers to argue with, and a product imagined whole and fitted to the people and the compute you have. It is for technical firms deciding what to build, what to buy and what to leave alone, and the same people take the call when parts of it go wrong.
-
An environment that runs at the scale a training run needs, a grader with its reasoning written down, and a record of how agents tried to game it and what we did about that. We build in domains we have worked in for years, quantitative ones among them, and the environment is yours to keep.
-
A test suite that is yours to run again, a report on where the agent breaks and why, and a measure of how far its confidence can be trusted. Miscalibration, brittle failure modes and reward hacking are the usual findings.
-
Models built for data that changes under them, the evidence that they hold on data they have never seen, and monitoring that says when they have gone stale. We have spent many years on series like these.
-
The pipeline from data to a trained model to a served one, built so your own people can run it after we leave, and profiled until we can show you where the time goes. We run it on our own compute or on yours.
Terms
- Start
- A conversation with the person who would do the work
- Form
- Fixed scope or retainer
- Confidentiality
- NDA by default
- Conflicts
- Checked before we say yes
- Compute
- Our own, on premises in London, with rented scale when a question earns it
- Location
- London, remote or in person worldwide
- Capacity
- Limited
Who we are
We learned our trade in global markets, sharpening our tools against adversarial uncertainty and designing models for action and understanding.
HyperPrior Research is a private research group in London, founded in 2022 and led by Patrick Steele. Before founding it he spent more than a decade in systematic and quantitative trading, most recently co-building a systematic desk at DRW, and before that trading APAC Delta One at UBS in Hong Kong. He read computer science at the University of Auckland and holds an MBA from Cambridge.
We are small by design and privately funded, with no outside money.
Quantitative finance is where we come from, and one of several domains we know well enough to build environments and graders for. We also know the plumbing of the financial system from the inside. We have worked on joining replicated state machines, and the digitally native assets they carry, to the market infrastructure that was already there, with some of the largest asset managers, exchanges and post-trade providers. Some of our analysis of that plumbing has been published by a G7 treasury.
We work from Shoreditch, an area where industrial and thinking processes have thrived and iterated for centuries. Thomas Bayes worked nearby and now rests in Bunhill Fields, just a short walk from here.
Positions lists three open roles, if you are interested in collaborating.
Contact
Write with a problem or a question. If you want to work here, Positions says how.
