Source
Freeborn (2026), LIT-131 (arXiv 2604.04171v1), read in full in NOTE-123. The companion notebooks were not run.
What was actually shown
The criterion (§4.2, p. 15). A system S systematically understands a property p of a target T insofar as (1) S contains a subsystem M that is an adequate model of some aspect of T; (2) M tracks p without memorizing it; (3) bridge principles connect M’s terms to T’s; and (4) S can use M to derive properties of p. Bridge principles count as part of S only if fixed at deployment, executed automatically, part of closed-loop competence and functionally integrated (p. 19). Non-memorization needs stability of the prediction under a family Π of structure-preserving perturbations (Eq. 7).
The demonstrations (§6). A 4,481-parameter ReLU network trained on local values of a torus function yields a closed surface with Euler characteristic 0, hence genus 1. A network with Kepler’s hypothesis space built in fits Mars’s eccentricity at about 0.096 (modern about 0.093). A small transformer groks addition mod 113, and the Fourier spectrum of its logits concentrates on few frequencies. Othello-GPT is reported from Li et al. 2023.
The hypothesis (§§7.4–7.5, pp. 45–46). Understanding in deep networks is symbolically misaligned (loss-driven coordinates), non-reductive (patchwork is a stable resting point), locally fractured (piecewise structure), proxy-dependent (any predictive dependence is rewarded) and representationally fragmented (no pressure to factorize). Patchwork solutions are common because they exist, local gradient descent finds them, and they are stable once the loss is low. The paper offers this as one cause of the “jagged frontier”.
What could have come out otherwise. The demonstrations could have failed the criterion. They mostly did not, but they bear on whether networks can understand, not on whether their understanding is fractured. Nothing in the paper tests the hypothesis.
What this does not say
- That networks often understand. “Often” in the abstract is not measured. Every demonstration is a toy, a single run, with no seeds or ablations.
- That Othello-GPT understands the board by the paper’s own test. The bridge principles cited (activation to decoded board) are post hoc probes, which fn. 50 and p. 19 exclude from the agent. By its own criteria the case shows only that board information is present and used.
- That the Kepler result is about deep learning. The conic form, heliocentrism and the area law were built in, and the author concedes it is “debatable” whether this is deep learning at all (p. 34).
- That weight decay is needed for grokking. Footnote 47 says that without it the model “remains in the memorization regime indefinitely”. The anthology’s reading of the grokking literature (ANTH-LIT-538) records weight decay as the strongest intervention, changing how much data is needed, not whether grokking happens.
- That understanding verdicts are observer-free. The perturbation family Π is “necessarily interest-relative” (fn. 19).
- Anything about large language models. They are excluded by design.
Connections
- The reading names the bridge-principle conditions (p. 19) and the stability test (Eq. 7) as the paper’s most usable export, more than this hypothesis (NOTE-123).
- Angius et al. (LIT-153) pair naturally with it on explaining network competence, as does López-Rubio (LIT-148) on computational functionalism.