HYPERPRIOR

Working papers

HyperPrior Research
London

Five working papers are in preparation, one from each of the group’s programmes. The programmes cut across the seven research areas on the home page. Each paper will carry its date, its version and how far we trust its main claim.

  1. No. 001 Probabilistic inference

    The prior is the product

    P-01 Prior Engineering

    Pretraining data is a prior, whether or not anyone meant it to be. We argue for designing that prior deliberately, and for measuring what a model actually absorbs from it.

    Version
    Date
    Confidence
    Revisit

    In preparation

  2. No. 005 Reinforcement learning

    Forgetting on purpose

    P-02 Filtration

    A model that learns from a stream has to keep deciding how much of the past still applies. We make the case for models that forget on purpose, and for measuring when and why they let go.

    Version
    Date
    Confidence
    Revisit

    In preparation

  3. No. 002 Environments and graders

    Scoring the future

    P-03 Out-of-Sample

    The only benchmark that cannot leak is one that has not resolved yet. We set out a protocol for scoring forecasts committed before their outcomes exist, including when the forecaster can move the outcome.

    Version
    Date
    Confidence
    Revisit

    In preparation

  4. No. 004 World models

    Where beliefs live

    P-04 Belief Geometry

    A model that predicts well has to track the hidden state of whatever produced its data. We look at reading those beliefs off its activations, and at what their geometry says about how the model represents that process.

    Version
    Date
    Confidence
    Revisit

    In preparation

  5. No. 003 Complex adaptive systems

    Other people’s world models

    P-05 Shared Worlds

    In a market, most of the environment is other agents. We ask what a world model has to learn there, and how to plan with a model that goes stale once the other agents adapt to it.

    Version
    Date
    Confidence
    Revisit

    In preparation

Close

HyperPrior Research Ltd, registered in England and Wales, No. 14046272. Registered office 60 London Wall, London EC2M 5TQ.

Notice

HyperPrior Research Ltd is a private research and development company. The firm undertakes applied research, modelling, simulation, and systems development on a selective basis. This website identifies the company and describes its work in general terms. It is not a complete statement of the firm’s interests, capabilities, methods, or current work, and it is deliberately partial. The interactive elements on this website are illustrative only and describe no real results or clients. Nothing on this website constitutes investment, legal, regulatory, tax, technical, or other professional advice, nor does anything on this website constitute an offer, solicitation, recommendation, inducement, arrangement, or invitation in any jurisdiction. No person should rely on the contents of this website in connection with any transaction, investment activity, research engagement, commercial arrangement, or other decision.

Privacy

HyperPrior Research Ltd is the controller for personal data processed through this website and related correspondence. To reach the controller about privacy, use the communicator on the home page and choose Other, or write to the registered office. This website does not provide accounts, comments, public user submissions, or payment facilities. Its communicator sends and stores nothing itself. It composes a message in your own email application, and nothing leaves your device until you send it from there. The website operates without cookies, advertising trackers, behavioural profiling, or analytics technologies. One small note is kept in your browser’s session storage for the length of your visit, listing the calibrator statements you have already seen, and it never leaves your device. The calibrator’s graded slab is drawn on your device, saved only where you choose to save it, and never uploaded. Posting it is done by you, from your own account, on a site this website does not control. A link you share carries the number that chose your five statements and nothing else.

Hosting, security, network, and email providers may nevertheless process ordinary technical information required to deliver and protect the website and associated systems, including IP address, browser information, request metadata, timestamps, and similar operational records. If you contact the firm by email, the firm may process your email address, name, organisation, message content, metadata, attachments, and any information you choose to provide. Personal data is processed for the purposes of receiving, reviewing, recording, and responding to correspondence; assessing possible research or commercial discussions; maintaining appropriate records; protecting the security and integrity of systems; and complying with legal, regulatory, accounting, or administrative obligations. The lawful bases relied upon may include legitimate interests, compliance with legal obligations, and steps taken at your request before entering into any arrangement.

Personal data may be processed by infrastructure, hosting, email, storage, professional service, and administrative providers acting for or assisting the firm, and may be disclosed where required by law, regulation, court order, public authority, or competent regulator. Correspondence and related records are retained only for as long as reasonably necessary for the relevant purpose, unless a longer period is required or justified for legal, regulatory, security, accounting, archival, or dispute-management reasons. The website does not make solely automated decisions about visitors with legal or similarly significant effects. You may have rights under applicable data protection law to request access to, rectification of, erasure of, restriction of, objection to processing of, or portability of your personal data. You may also complain to the Information Commissioner’s Office, the UK supervisory authority (ico.org.uk).

Colophon

Set in Austin, drawn by Paul Barnes and published by Commercial Type, and Xanh Mono, drawn by Lâm Bảo and Duy Đào of Yellow Type. The few glyphs they lack come, as subsets, from Noto Serif, Zen Old Mincho, Noto Sans JP and STIX.

Static files, served as they are. The liquid chrome and the pool run in WebGL2, with the simulation in a worker, and each stops when it is off screen or the tab is hidden. Firmware 0e3cb7e. Page weight 1.3 MB, from the card to the footer.

Three devices, HPR-C1, HPR-C2 and HPR-C3. The calibrator holds two banks, 400 general statements and 200 on machine learning, five to a round, each one checked. A screen reader gets the receipt in full, jokes included.

No cookies, no trackers, no requests to other servers, no sound. Made in Shoreditch, London.