Vision

Where Continuous Function is going, and how much of it exists today.

As AI systems take on more of the work of building AI, such as writing code and running experiments, the people responsible for that work need to be able to see what changed, understand why, question it and redirect it in time. This page describes a way of working that would make that possible, the small part of it you can try on this site today, and what is not built.

Status
A direction, not a product
Direction set
30 September 2026

The direction

Picture one workspace for each question, shared by a person and an AI system. The person would fix what must not change: a test set the AI system may not see, a limit on what may be traded away. Within those limits the AI system would explore on its own. Its proposals would arrive as drafts; only the person could approve a plan, and approval would mean that the plan may run, not that it is right. Work could continue while the person is away, inside the limits they set, and when they return the same page would show what changed, what held and which decisions need them.

Claims would be narrowed to what the evidence supports before they are shared, and would travel as packages that someone else can rerun, adapt and extend. If AI systems begin to improve their own methods, those changes would pass the same test: tried on fresh tasks, judged by a checker that the system being improved cannot alter, and adopted only where a person agrees.

The first part to build is the record of who predicted what, who corrected what, and what held up. None of this exists yet as a product. A small version of one loop runs today in the workspace: a bird classifier that fails one group, a prediction made before each run, limits that are enforced, and a record of what held up. That record holds only your own predictions, corrections and results, kept in your browser; a record shared across people and services is not built.

What can be tried today

The direction has six parts. For each, the links below open the place on this site where a first, small version already works, say where the answers there come from, and say what is not built. Everything runs in your browser, and no AI model is called.

  1. See the failure, image by image

    Look at where a model fails one example at a time, not only at its average.

    Where the answers come from: Computed in your browser from the benchmark’s numbers.

    Today: 600 of the benchmark’s validation photos drawn as anonymous dots (the photographs themselves are not included), the trade-off each option makes, and which photos each option fixed or broke, all computed in your browser.

    Not built: views of the data in three dimensions, and explanations you can direct by voice or by drawing.

  2. Ask where a statement comes from

    Ask of any sentence in a result: where does the evidence show that?

    Where the answers come from: Replies written in advance; inside a draft claim, answers computed from your own results.

    Today: select a word in a draft claim and a guide whose replies are written in advance answers from the results. The claim itself is assembled only from phrases generated from the numbers, which you choose among; your own words go in a note beside it, never into it, and are not checked. A package received from someone else shows its claim as written.

    Not built: asking an AI system a question of your own; no AI system answers on this site.

  3. Predict before the result

    Commit to a prediction before anything runs, so the result has something to be compared with.

    Where the answers come from: Your guess, set against what is then measured, beside a prediction written into the case in advance.

    Today: place your guess on the trade-off, with a reason, before anything runs; it is then compared with the measured result, beside the case’s own prediction written in advance.

    Not built: a record of forecasts over many decisions, for a person and for an AI system alike.

  4. Fix what must not change

    Set limits the work may not cross, including a test set that stays unseen until the end.

    Where the answers come from: Enforced in your browser, not merely advised: the workspace’s limit check, its compute budget and its single look at the test set; the retrieval case’s fixed settings.

    Today: keep the average within a limit, the test images unseen and the computation within a time budget. Limits are kept as a draft and take effect once you approve the plan: every candidate option is checked against your limit, overriding the check needs your acknowledgement and is recorded, refitting stops at your budget, and a separate checker opens the test file only once per workspace, after recording that look in this browser.

    Not built: limits enforced on real computation, data and agents beyond this page, and a checker on separate infrastructure whose files an agent cannot read.

  5. Read the source’s own words

    Read a paper in its own sentences rather than in a summary, and mark the word you doubt.

    Where the answers come from: From the source: the sentences are quoted from the paper and cited.

    Today: the paper’s own sentences, quoted and cited; mark the word you doubt and it travels with your work as a note to yourself: “not necessarily in our model: test it”.

    Not built: listening to a paper read aloud and marking it as you go.

  6. Approve a plan, then share work others can rerun

    A plan runs only after a person approves it, and the result is shared in a form others can rerun and extend.

    Where the answers come from: Your approval, bound to that exact version of the plan; the runs are then computed in your browser.

    Today: a plan runs only once you approve it, and the approval is bound to that exact version. Your claim, the plan’s history, every completed run, each look at the test set and every removal go into a package that anyone can rerun step by step, or adapt as a draft of their own.

    Not built: a shared record across people and services, and a common space where another research group can add a test to your evaluation.

The same idea also runs in a smaller form: train a small model on a synthetic three-feature task and look inside one checkpoint. Lessons that teach the mechanisms underneath start with attention.

Where each answer comes from

Where a page on this site answers you, as in the desk and the workspace, the reply carries one of these labels beside it:

written in advance
replies prepared before you arrived: the first exchange, the case’s predictions and its proposed drafts.
computed from your record
answers worked out from what you did here: your guesses, runs, limits and marks.
from the source
a paper’s own sentences, quoted and cited, never summarized.
model · provider · cost
shown only when a live model answers. None is connected, so no line on this site carries it.

A live model would connect where you ask: the questions you select, and a question you bring in your own words. Its proposals would arrive as drafts carrying the fourth label, and would become yours only when you approved them. The connection for it exists and is switched off. How much an AI system may do on its own is a setting on your desk; it starts at proposing, and nothing beyond proposing is built.

Films

Short films show an imagined working day of this kind. Each says on screen that it shows an imagined day, not software that exists; the newest grounds its one benchmark rerun in a real, measured result and labels everything else as illustrative. They are not published online yet; to see them, email founder@continuousfunction.ai.

What We Can Build On4:24A researcher and an AI system turn one model failure into a narrower claim that another lab builds on, with limits that come to be enforced, a test set kept hidden even from the AI system, and a measured rerun of a published result.
The North Star3:35Why people lose track of what AI systems are doing, and an imagined layer, above the models, agents, tools and cloud services, in which they could keep track of it.
The North Star, research edition4:41One machine-learning researcher’s imagined day as AI begins to help build AI: the page that shows what changed overnight, the services underneath, and a person approving each step.

How it is being built

The site itself is a small instance of this way of working: much of its code is written by AI coding assistants under our direction, and we decide what is built and published. How this site is written describes this in full.

The next step planned is a shared record of what was proposed, approved, run and shared, with limits enforced outside the browser page. It would be tried first on synthetic data and then with outside teams on their own work. It is planned, not built.

Openness is part of the direction. Today the source repository is private.

In more detail

What exists, what does not, and the status of every research question are set out in How Continuous Function works.

The open question

Everything above is one small loop: a prediction before the run, a limit that is enforced, an approval that means the plan may run, not that it is right, and a record of what held up. Whether that loop can stay this clear when the work is larger, faster and partly done by AI is not known. That is the question this project works on.

When AI begins to help build AI, can the people responsible still see each change while it is small enough to steer?