Research roadmap · Learning properties
Three levels of generalisation
“Generalising” is not a yes/no box. It is a climb in three stages. The first is initiated, with a verified partial result; the second is observed, though it cannot be credited to the system alone, since a simple similarity method reaches it too; the third — grasping an abstract rule never seen before — remains open. Here is the exact map, and why determinism makes it possible to draw.
Two axes, not a race
Other projects simulate the fruit fly with remarkable biological fidelity — complete connectome, realistic body, natural behaviours. On that axis of fidelity they are ahead, and I am not trying to catch up.
My axis is a different one: verifiable learning. A faithful replica of a connectome is a snapshot: its connections are those of a given moment, and on its own it does not learn. My prototype learns, and that learning can be checked: on a circuit modelled from part of the fruit fly connectome, it learns a task, then a second one, and partly retains the first, above chance; on the protocols tested, its recorded decisions replay identically, bit for bit, in Python as in Rust, and resuming after an interruption gives exactly the same state. What I have not yet shown is that an experienced consequence is enough to change a later choice. The question is therefore not whether I have a better fly, but how far this learning goes and where it stops today.
The three-level scale
map drawn from the testbed’s tests
Level 1 — local / associative generalisation
This is learning from direct experience: associating an object with what followed it (“this object was favourable to me”) and using that association afterwards. On my testbed, the prototype learns one association, then a second one, and partly retains the first: its performance on the first task stays above chance (about 0.62 on average over a study series of 12 simulated individuals, against 0.50 for chance; one of the twelve stays below chance), while dropping clearly from the level reached right after the first learning phase. On the protocols tested, its recorded decisions replay identically, bit for bit, with two implementations (Python and Rust).
This is learning, not a reflex wired in advance. Initiated, within a bounded perimeter: the learning protocol is supervised, the first task is only partly retained, and it remains to be shown that an experienced consequence is enough to change a later choice.
Level 2 — generalisation by similarity
One step higher: recognising or transferring to new objects when they share a sensory core with objects already seen. The prototype does this — but a simple similarity baseline (KNN) does it too, which marks the boundary: it is useful, yet it amounts to recognising a resemblance rather than grasping a rule.
Observed and bounded — useful, but not “intelligent” in the strong sense, and not attributable to the system alone.
Level 3 — strong compositional generalisation
The summit: “I understand that A+B implies X, even though I have never seen this combination.” This is where an abstract rule genuinely comes into play, and where the most recent tests are unambiguous.
The pre-registered tests run so far show no learned compositional generalisation that beats a baseline. The cause has been identified internally; it is not detailed here, because it concerns the design of the prototype and the next steps of the work.
Open — still some way off, and the obstacle is identified.
A diagnostic instrument
A well-established negative result is valuable: it avoids searching in the wrong place. A deterministic system makes it possible to see it clearly, to replay it and to check it, test after test.
It is the kind of limit a deterministic system makes clearly visible, where an opaque probabilistic system leaves room to hope indefinitely that more training will eventually work.
What comes next
The plan of work that follows from this diagnosis is established and tracked internally. It is not detailed here: a published roadmap would tell a competitor where to look, and above all what to avoid. What will be published are the results, once reviewed, negative ones included.
The day the prototype reaches level 3 — a learned compositional generalisation (not a rule programmed in advance), on an integer substrate, beating a baseline — this page will move from “objective” to “reached”. That result would be submitted for peer review. Until then, this page describes the objective as it stands, and what has been acquired as it has been measured.