h t t p s : / / p a t h w a y . c o m / r e s e a r c h / t h e - e q u a t i o n s - o f - r e a s o n i n g

Ingest: The Equations of Reasoning (Pathway)

Type: types/ingest-report.md

Classification

A programmatic research-blog essay. It argues for a research direction and summarises formal results proved elsewhere, in the BDH paper (arXiv 2509.26507). It reports no new experiment of its own; the one empirical item is a sparsity figure reproduced from that paper. The capture lost the equation symbols, so the essay's formal content is readable only as prose. Author: the Pathway Team (dated 2026-08-06), the company that develops the BDH ("Dragon Hatchling") architecture. The essay promotes the company's own model, so the author has a commercial interest in the framing.

Summary

The essay argues that AI development is "phenomenological": benchmarks, scaling laws, and interpretability findings record regularities in tested regimes but cannot predict behaviour outside them. It calls for an "axiomatics of AI" modelled on statistical mechanics and Hilbert's sixth problem: local microscopic rules from which macroscopic reasoning behaviour (memory, abstraction, adaptation, stability) can be derived. It takes the brain as the inspiration for those rules: sparse, local, Hebbian, online adaptation on a large graph. It then presents BDH's "Equations of Reasoning" as that axiomatics. Each token activates neurons; activity reads fast synaptic state as a weighted modus ponens and writes it by a Hebbian outer product; a trained graph routes activity. The cited formal result is a simulation chain: attention can be simulated by BDH-GPU, and BDH-GPU by local neuron-synapse dynamics, each with controlled computational overhead. That is an expressiveness and equivalence result. From it the essay claims that model growth, long inference, continual adaptation, and composition become open to mathematical analysis, and that safety can become a design property: restrict deployment to regimes where memory dominance, error amplification, and sparsity collapse stay bounded. The essay derives no such macroscopic law or bound. Those claims describe what the programme hopes to deliver.

Quotes

No source quotes have been retained yet.

Connections Found

The essay's main role for this KB is a contemporary test case for what a derivation from formal rules licenses beyond tested cases. It lands on the reach-assessment chain. Causal and proof obligations are two formal routes to assessing explanatory-reach is the closest comparison. BDH is a proof-route programme for ML architectures, but its proved content is simulation equivalence. The extrapolation and safety claims would additionally need a proved correspondence between the axiomatized dynamics and deployed behaviour, and the essay does not address that boundary. Reach-assessment already states the essay's central complaint in the KB's terms: observed predictive fit reaches only across the tested shift class.

Derivation and inheritance give starting warrant; discriminating evidence or proof earns scope separates what the essay has from what it claims. Its axioms are chosen by inheritance from neuroscience, with the brain as "existence proof". That gives a conditional starting warrant. Proof then earns scope only within the formal domain of the simulation result. The bitter lesson selects against unearned reach, not against structure makes the essay a live case. It rejects brute-force scaling and argues for derived structure, but its cross-scale reach is currently asserted and earned only for simulation equivalence. Warranted autonomy is bounded by oracle domain compares with the safety proposal, which fixes the deployment domain by analysing dynamics rather than by testing. The essay also corroborates, from ML research outside Commonplace, the fit-versus-explanation polarity in first-principles reasoning selects for explanatory-reach over adaptive fit. It does not support that note's claims about KB methodology.

Among sources, Einstein on Brownian motion (1905) is the sharpest yardstick. It is the same micro-to-macro programme, but Einstein's derivation produced a macroscopic quantity a test could refute; Pathway's has not yet produced one. Beyond Transformers: Sudoku Bench is the same vendor and model. That post argued from a benchmark result, while this essay argues that benchmarks cannot establish behaviour outside tested regimes, and the Sudoku ingest's commercial-incentive caveats carry over. The essay also partly answers that ingest's complaint that the BDH architecture was undisclosed, at the level of prose. Nested Learning shares the view of memory as timescale-separated update dynamics inside the model rather than a separate module.

Learning Claims (our opinion)

The mechanism on its own terms. BDH has two timescales of change. Training sets nonnegative edge amplitudes on a fixed set of neuron graphs. During inference, a fast synaptic state on the edges is read (activity propagates as weighted implication: more confidence in X raises confidence in Y) and written (a Hebbian outer product strengthens edges between co-active neurons). The essay calls this continual, online adaptation and says memory is "part of the temporal fabric of computation" rather than a module. The adaptation it describes is in-weight state change during a run. It is not training-time learning, and the essay does not claim the fast state persists across runs.

Mapping onto Commonplace concepts. The fast synaptic state is distributed-parametric content in the KB's representational form sense. The essay claims it is more localized than Transformer state, because edges carry implication strengths between specific neurons and only a sparse set is active at each step. Whether a given edge "says something" that could be criticized is not argued. It is a candidate case for the theory-builder definition's "content located in weights" boundary, and the essay supplies no interpretability evidence to settle it. It also bears on opacity is a scale threshold as a conjectured structural affordance: sparse, local dynamics might raise the scale at which parametric state becomes opaque. The essay asserts this and shows only one sparsity figure.

Against the theory-builder conditions, judged for BDH as a model:

  • Localized content: open, weakly argued. Edges as implications are closer to localized units than dense attention weights, but the essay gives no evidence that individual edges carry stateable content.
  • Consumption: met within the model. The fast state directly conditions the next steps of inference.
  • Criticism: not met. The Hebbian write is correlation-driven. Nothing aims attempted refutation at what an edge says; the "modus ponens" read propagates confidence and does not test it. The update resembles the definition's weight-adaptation exclusion more than criticism.
  • Iteration: met in a mechanical sense. Each round's state change carries into the next round.
  • Persistence: within an inference run for the fast state; training-time for edge amplitudes. The essay does not describe any cross-run persistence.

BDH is therefore at most a component a theory builder could use, like any model. The essay's own programme is the part that resembles theory building: its stated axioms are to be tested by deriving macroscopic consequences. That programme has not yet produced a consequence a test could refute.

Fixed decomposition. Under learning inside a fixed decomposition inherits its mistakes, the local rules are the fixed decomposition. Training and Hebbian updates change amplitudes and state inside it. The simulation proof shows the decomposition is expressive enough to reproduce attention within comparable bounds. That is evidence the decomposition does not lose expressiveness relative to attention. It is not evidence that this decomposition is the right one for reasoning, or that improvements inside it transfer.

What it adds to our account. It adds little to Commonplace's account of learning. The in-context Hebbian state is a plausible mechanism with no reported evidence, in this essay, that it improves capacity for future action. Its main value is as an outside instance of the pattern in which a formal equivalence proof is presented as licensing behavioural extrapolation. It does not put the KB's current definitions in question.

Extractable Value

  1. A worked case of the proof-route formalization boundary -- the essay proves (via the BDH paper) that attention is simulable by local dynamics, then claims extrapolation and safety bounds that the proof does not cover. This is a clean, contemporary illustration for the formal-systems reach note: a simulation or expressiveness proof establishes representational equivalence, not behavioural bounds. High reach, because the pattern recurs wherever formal equivalence is sold as predictive theory. [quick-win]
  2. Einstein contrast as a test for axiomatics claims -- set beside the Brownian-motion ingest, the essay sharpens a criterion: a micro-to-macro axiomatics earns explanatory-reach only when the derivation yields a macroscopic quantity a test could refute. This could fold into the formal-systems note as a bounded qualification. [quick-win]
  3. Inherited-then-proved axiom warrant -- the axioms get starting warrant from neuroscience inheritance and scope from proof over a narrow formal domain. It is an example for the derivation-and-inheritance note of the two warrant sources acting on one axiom set with different scopes. [just-a-reference]
  4. Outside corroboration of the fit-versus-explanation polarity -- an ML research group independently frames benchmarks and scaling laws as "phenomenological" and asks for laws that predict beyond tested regimes. It is useful as an external citation that the polarity is live in model research, not only in KB methodology. [just-a-reference]
  5. Vendor evidence posture -- the same vendor argues from benchmarks in the Sudoku post and against benchmark sufficiency here. It is a small data point for reading vendor claims: the evidential mode shifts with the claim being sold. [just-a-reference]

Limitations (our opinion)

  • Naming is not deriving. The essay's central move is to call the BDH update rules "axioms" and invoke statistical mechanics, Carnot, and Hilbert. The analogy asserts that macroscopic laws will follow but derives none. Carnot's bound and Einstein's displacement formula were results; here the result is promised. The claim is easy to vary: almost any architecture expressed as local rules could make the same pitch.
  • The proof's scope is narrower than the prose suggests. The simulation chain shows expressive equivalence within complexity bounds. It does not show that analysing the local rules predicts deployed behaviour, that sparsity stays stable at scale, or that errors stay bounded. The essay also claims that Transformers lack uniform description across scale, but a Transformer is also the same rules applied to more parameters; the contrast is not argued.
  • Secondhand formal content. All formal claims are summaries of the BDH paper, and the captured text lost every equation symbol. Any claim about exactly what is proved should be checked against the arXiv paper, not this essay.
  • Safety claims are the weakest part. "Restrict deployment to regimes where these quantities remain bounded" presupposes bounds nobody has derived and a link from those bounds to misaligned behaviour of the kind the essay cites (OpenAI, Claude Opus incidents). The paperclip example concerns objective specification, which local dynamics do not address.
  • Brain inspiration as warrant. The essay itself says it does not claim AI must reproduce the brain. The brain works as an existence proof that sparse local dynamics can support cognition, not as evidence that BDH's particular rules do.
  • Commercial interest. The author sells the architecture. The essay reports no failure modes, no comparison with other local-rule or state-space architectures, and no test of the claimed extrapolation.

Update Causal and proof obligations are two formal routes to assessing explanatory-reach with a bounded qualification (under 200 words) citing this ingest and the Einstein Brownian-motion ingest: a micro-to-macro axiomatics earns explanatory-reach only when its derivation yields a refutable macroscopic quantity, and a simulation or expressiveness proof establishes representational equivalence, not behavioural bounds.