What changed: we have a System 2
Theory update · 2026-06-26 · the round in one place
Before this round the architecture had one half: a fast, parallel count substrate, with System 2 listed as absent. It now has the other half: a metacognitive gate that calls a still-minimal System 2 over explicit concepts. Here is everything that changed, and the one frontier it sharpens to.
What this round was
Five experiments, run as one arc. Each took a named result from cognitive science and asked whether it survives translation into counts: online, single-pass, no gradient. Two of them added capabilities the model did not have. One revised a standing decision. Two refined the rules of bounded memory. Read together, they move System 2 from a promise on the theory page to a working part. This post is the synthesis; each claim links its experiment.
(a) NEW: a working System 2
The model now has a second speed. A dual gate (calibrated confidence plus Botvinick conflict) decides when to think, and deploys a deliberate pass that overrides System 1 only when System 1 is wrong. The Engle signature is clean: on conflict cases, where the fast reflex is wrong 88% of the time, accuracy rises 0.38 (0.116→0.496); on no-conflict cases, zero harm (0.680→0.680); at zero budget it falls back to System 1 bit for bit. The honest finding underneath is that the gate is the win, not the elaborate serial workspace: on this one-step probe the cheapest deliberate operator, defer to the wider context, beats the focus-and-inhibition machinery (0.39 vs 0.15 where they differ). So the metacognition ships now; the workspace is parked for the multi-step task. Thinking slow, by counting.
(b) NEW: redescription turns counts into concepts
System 2 needs operands, explicit things it can name and move. Redescription mints them. On stability, not error, a mastered construction is promoted from an implicit count into an explicit, slot-addressable concept that answers queries the flat count structurally cannot: inverted slot lookup, role substitution, slot analogy. The Karmiloff-Smith U-shaped dip confirms the dynamic (0.155 → trough 0.051 → recovered 0.181, above baseline), with the honest caveat that the recovery is a consequence of the modelled re-binding: the real win is that the mechanism is expressible and self-consistent on counts, and that the resulting concepts are manipulable. This is the implicit-to-explicit, count-to-concept bridge: the operands the gate deliberates over. When a habit becomes a thought.
(c) CHANGED: the right combiner is take-the-best, not full pooling
The standing combiner pooled every cue with a geometric mean. This round revises it. Validity-ordered, noncompensatory, early-stopping take-the-best beats full integration on every axis at once: accuracy 15.00% vs 9.71%, perplexity 1,918 vs 7,160, at 4.56 cues a step instead of 8. The less-is-more effect holds where the theory predicts: ignoring the weak channel wins, most sharply on sparse contexts. A count model is high-bias by construction, so stopping early is the right trade. The combiner sharpens by ignoring, not by blurring, the thing the scorecard kept asking for. Less is more, and you can prove it.
(d) REFINED: bounded memory has a shape and no schedule
Two refinements to the budget rule. First, the shape of forgetting: ACT-R's power law is the right curve (the only one that represents spacing, spaced 8.96× more accessible than massed) but for dense char-grams raw-count LFU wins eviction at every cap, because LFU is the power law's d→0 limit and a char-gram's value is pure frequency. Keep LFU for char-grams; reach for the power law at the sparse word and concept level. Second, no curriculum: growing the memory budget on a schedule does not beat full-from-start (2.744 vs 2.751), because "starting small" was a property of the gradient optimizer, and a count learner cannot get stuck. Final horizon matters; scheduling is a no-op. The shape of forgetting · Starting small, on purpose.
The synthesis
The architecture now reads in one sentence. System 1 (the parallel count substrate, voting instantly to recognize a word or feel a topic shift) feeds a metacognitive gate that, on low confidence or conflict, calls a still-minimal System 2 that deliberates over explicit, redescribed concepts and overrides the reflex only when it is wrong. Bounded memory underneath it has a settled shape (LFU for dense counts, the power law for sparse ones) and needs no curriculum. The combiner that pulls the votes together is frugal, not exhaustive.
What is conspicuously not yet proven is the workspace itself. The gate is real and the operands are real, but on a one-step char prediction the elaborate serial deliberation loses to a trivial deferral, because one-step decisions cannot exercise a workspace. That names the frontier precisely.
The frontier
The open question is now the multi-step task (compositional reasoning, discourse coherence) where the serial workspace holds and manipulates redescribed concepts across cycles, and a single deferral cannot reach the answer. That is the one place the parked machinery should earn its keep, and the one place we have not yet tested it. The combiner question that stood open is now partly answered: sharpen by take-the-best. The coherence question remains. The next swing builds the multi-step probe and puts the workspace, the gate, and the explicit concepts to work together.
The five posts
- Thinking slow, by counting: the gate and the deliberate pass.
- When a habit becomes a thought: stability promotes counts to concepts.
- The shape of forgetting: power-law shape, LFU place.
- Less is more, and you can prove it: take-the-best beats full pooling.
- Starting small, on purpose: no curriculum for a count learner.