HYPERPRIOR

Positions

HyperPrior Research
Shoreditch, London

Three roles are open. If one is close to what you do, write before you are sure you fit.

HyperPrior is small by design. People usually join by project at first, and a project that goes well tends to become a longer arrangement. The work is environments and graders, the seven research areas on the home page, and the engagements that come from them.

We care more about what you have built and what you have been wrong about than about where you have worked. The bar is high, and it is the same bar for each role.

Open now

  1. 01

    Member of Technical Staff

    Builds the instruments

    The work

    You build the instruments. Environments that run at the scale a training run needs, graders with their reasoning written down, the infrastructure under both, and the tooling that lets a researcher ask a question in the morning and have an answer by evening. Some weeks the work is a distributed training run, some weeks a simulator that has to be fast and honest at the same time, and some weeks an engagement with a client.

    What we look for

    Someone who has shipped systems that ran unattended for a long time and knows why they did. Fluent in Python and in at least one fast compiled language, at home with GPUs and the stack around them, and able to read a paper before lunch and have its method running after. Reinforcement learning, simulation or large-scale data pipelines in your past, and a habit of measuring before believing.

    Go to the communicator

  2. 02

    Research Fellow

    Brings a question

    The work

    You bring a question of your own that sits in one of the seven areas, and you pursue it with the group’s compute, its network and its data, with a paper as the usual output. Fellows also check the graders. A grader is a scientific claim about what good behaviour is, and we want people who can state the claim, test it and say how sure they are.

    What we look for

    A piece of work we can read, published or not, that shows you can take a question from vague to answered. Mathematical fluency, in probability above all, and the engineering to run your own experiments. Calibration, which here means you have been wrong in print or in code and said so. A background in machine learning, statistics, physics, economics or the quantitative end of another field.

    Go to the communicator

  3. 03

    Wizard

    Does the other thing

    The work

    The role we send when the problem has no name yet. A wizard works across domains, picks up a field in weeks, and does the possible and, now and then, what the client was told could not be done. Some of the work is research, some is engineering, and some is being the person in the room who knows what the system is doing. On an engagement you are there until it is finished.

    What we look for

    There is no checklist. We look for a record of hard things done in more than one field, the taste to know which problem matters, the speed to be useful in the first week, and the honesty to say when a thing cannot be done. Quantitative fields help, because many of the environments we build come from them.

    Go to the communicator

Terms

Location
Shoreditch, London, or remote
Form
Full time, or by project at first
Start
After a conversation with the person you would work with
Confidentiality
NDA by default

What we have been wrong about

  1. We thought the hard part of reinforcement learning was the agent. Usually it is the reward.
  2. We expected a model that was well calibrated to stay that way. Calibration drifts, and we now check it as often as accuracy.
  3. We assumed more data would settle most arguments. It settled some, and made the others more expensive to have.

How to apply

Use the communicator on the home page and press Joining. Name the role, say what you have built and what you have read lately, and tell us one thing you were sure of that turned out to be wrong. A short note beats a long one, and a link to work beats both.

Go to the communicator

Before you write, try the calibrator. Mention your grade if you like.

Try the calibrator

These roles are open now. When one is filled, this page will say so.