Possessing intelligence and using it are different things
Take a simple question: what is a white car? A small model that only knows cars answers it as well as a model that also knows physics, law and literature. Yet the larger model does vastly more computation to say the same thing.
A human brain knows aviation, mathematics and a childhood home, and recognises a white car almost instantly, at almost no extra energy cost. We want to know whether an artificial model could have that property, or whether the extra cost is intrinsic.
We are not building a smaller model, compressing one until it forgets, or caching answers. The aim is the full capability, kept available, with only a small part exercised at a time.
What P(f) means
Poised
A working research concept: the model's learned capability would remain ready between thoughts. Whether a small evolving state can support that availability remains an open question.
Flow
A working research concept for how a thought might move through what the model knows. Whether staying in an area or moving between areas can be cheap remains to be tested.
Purpose
To find out whether a model's size and the cost of one thought must rise together, or whether that link is a habit of how today's systems run.
The target, in words: keep the model's full capability, and measure the cost of one thought honestly, counting compute, data touched, data moved, time and energy. No system we have built meets this yet. "f" is our working reading of the name: flow.
How we work
Every idea gets the cheapest experiment that could prove it wrong, with success defined before we look. Positive and negative results are recorded transparently, with their appropriate evidence grades. Negative results help identify which specific routes did not work. Each finding carries one of three evidence grades.
| Confirmation | Pre-registered, independently reproducible test with a frozen pass rule. |
|---|---|
| Development | Real measurement used to understand a mechanism. Informs the next question, not a claim. |
| Scout | Quick, exploratory look that earns a closer one, or a decision to stop. |
Findings repository
Short, plain-language summaries. Each states what was shown and what was not. Designs, parameters and code are deliberately omitted. Additional technical information may be considered for qualified collaborators under appropriate confidentiality arrangements.
Scientific summaries are provisional pending founder/CTO review. PDF summaries are pending publication.
Where the evidence stands
Sensitivity is not necessity
Changing a part of a model changes its output, yet removing part of it was often harmless. Which parts move the answer is a different question from which parts are needed.
Predicting is not avoiding
Knowing in advance what a model will compute does not remove that computation.
Cheap and capable do not come free together
Cheap replacement parts ran fine and lost the capability. Preserving what the model can do is the hard part.
State is not the missing ingredient
Modern models already carry persistent state. The open question is whether useful state can evolve cheaply.
Shown
- Within a restricted behavioural test, source-established information and an evolving compact state had causal effects while the large model was absent.
- At one model layer, reduced expert sets were locally adequate in a restricted, hindsight-based single-step test.
- Specialist execution was independently demonstrated; selection and chaining of separate learned capabilities were demonstrated in bounded laboratory composition.
Not shown
- That a large model's full capability can be kept available at a fraction of the computation.
- That the small state carries useful information beyond what it was trained to carry.
- A working system, a speed-up, or any saving that survives measuring the whole cost.
- Anything at frontier scale. The programme is open and could still fail.
Current focus: why does capability need expensive computation to become usable?
A model may hold a capability, yet reaching it takes a long chain of computation that builds the right internal conditions first. We are working out which parts of that chain matter, and whether a much cheaper process can produce the same conditions while the capability still works.
Our latest experiment tested whether a real model can keep computing usefully with no new input. The state genuinely changed, but we measured no benefit at the scale we tried. We count that as a result.
Work with us
We are looking for people who find this question as interesting as we do, and who are comfortable with results that say no.
Researchers
Working on conditional computation, sparsity, recurrence, interpretability or memory-efficient inference? We would value critique of our methods and are open to joint work.
Startups and engineers
Building towards large models on modest hardware? We can share what we have ruled out so you do not repeat it, and compare notes on what is still open.
Contact details pending founder confirmation.