Watts-per-Intelligence Part II: Algorithmic Catalysis
Pith reviewed 2026-05-10 01:25 UTC · model grok-4.3
The pith
Reusable computational structures bound any class-specific speed-up by the algorithmic mutual information between substrate and task descriptor, at a minimum energy cost from Landauer erasure.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
Algorithmic catalysis is defined by reusable computational structures that lower irreversible operations for a task class while meeting bounded restoration and structural selectivity constraints. Any resulting speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor. Installing this information carries a minimum thermodynamic cost set by Landauer erasure. The two limits together produce a coupling theorem that lower-bounds the deployment horizon required for the catalyst to be energetically favourable.
What carries the argument
The algorithmic mutual information between the substrate and the class descriptor, which supplies both the upper bound on achievable speed-up and the minimum Landauer cost that must be amortized.
If this is right
- The bound and theorem apply directly to an affine SAT class as a concrete illustration.
- Contemporary learned systems are subject to the same information-thermodynamic constraint on intelligent computation.
- A catalyst is energetically favourable only after its deployment horizon exceeds the lower bound given by the coupling theorem.
- No class-specific speed-up can surpass the mutual information upper bound, regardless of the reusable structure chosen.
Where Pith is reading between the lines
- The bound could be tested by constructing reusable structures for SAT instances, measuring the resulting speed-up, and comparing it against the algorithmic mutual information between structure and class.
- The same limit may constrain how much pre-computation or pre-training can accelerate later inference steps without a matching increase in shared information.
- The coupling theorem suggests examining whether current AI systems already operate near or below the predicted deployment horizons for their effective task classes.
Load-bearing premise
Reusable computational structures exist that can meet bounded restoration and structural selectivity constraints for the task class while actually attaining the mutual information bound.
What would settle it
An observed speed-up for a task class that exceeds the measured algorithmic mutual information between the reusable structure and the class descriptor, or a catalyst that becomes net energy-positive before the horizon predicted by the coupling theorem.
read the original abstract
We develop a thermodynamic theory of algorithmic catalysis within the watts per intelligence framework, identifying reusable computational structures that reduce irreversible operations for a task class while satisfying bounded restoration and structural selectivity constraints. We prove that any class specific speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor, and that encoding this information incurs a minimum thermodynamic cost via Landauer erasure. Combining these results yields a coupling theorem that lower-bounds the deployment horizon required for an algorithmic catalyst to be energetically favourable. The framework is illustrated on an affine SAT class and situates contemporary learned systems within an information thermodynamic constraint on intelligent computation.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a thermodynamic theory of algorithmic catalysis in the watts-per-intelligence framework. It identifies reusable computational structures satisfying bounded restoration and structural selectivity for a task class, proves that any class-specific speed-up is upper-bounded by the algorithmic mutual information between substrate and class descriptor, shows that installing this information incurs a minimum Landauer erasure cost, and combines these into a coupling theorem lower-bounding the deployment horizon for energetic favorability. The framework is illustrated on an affine SAT class and used to situate learned systems under a unified information-thermodynamic constraint.
Significance. If the central theorems and constructions can be rigorously established with explicit derivations and verifications, the work would provide a novel coupling between algorithmic mutual information and thermodynamic costs, yielding falsifiable bounds on when catalytic structures become energetically advantageous. This could unify constraints on intelligent computation across information theory and physics, with potential implications for evaluating the efficiency of learned systems.
major comments (3)
- [Abstract] Abstract and main theorems: the central coupling theorem lower-bounds deployment horizon by combining the mutual-information upper bound on speed-up with Landauer cost, but the manuscript states that proofs exist without providing any derivations, error analysis, or verification steps for either the mutual information bound or the Landauer installation cost.
- [Illustration on affine SAT] Affine SAT illustration: the framework claims to identify reusable structures achieving the mutual information bound under bounded restoration and selectivity, yet no explicit construction, restoration-cost calculation, or check that the information-theoretic bound is attained is supplied; this assumption is load-bearing for the theorem's applicability to real catalysts.
- [Coupling theorem] Coupling theorem: the lower bound on deployment horizon presupposes that reusable structures exist which realize the algorithmic mutual information without super-Landauer overhead, but the text provides no construction or quantitative check of this condition, leaving the bound's practical relevance unverified.
minor comments (2)
- Notation for 'algorithmic catalyst' and 'class descriptor' is introduced without a dedicated definitions section or table, making it difficult to track how these entities relate to standard mutual information quantities.
- [Abstract] The abstract claims the framework 'situates contemporary learned systems' but provides no specific examples or quantitative comparisons with existing systems.
Simulated Author's Rebuttal
We thank the referee for the careful and constructive review. The comments correctly identify places where the manuscript would benefit from expanded explicit derivations and constructions. We respond to each major comment below and will revise accordingly.
read point-by-point responses
-
Referee: [Abstract] Abstract and main theorems: the central coupling theorem lower-bounds deployment horizon by combining the mutual-information upper bound on speed-up with Landauer cost, but the manuscript states that proofs exist without providing any derivations, error analysis, or verification steps for either the mutual information bound or the Landauer installation cost.
Authors: We agree that the main text states the theorems without supplying complete step-by-step derivations or error analysis. The proofs are outlined conceptually in Section 3, but we will add fully expanded derivations, including verification steps and any necessary error bounds, for both the algorithmic mutual-information upper bound on speed-up and the minimum Landauer erasure cost. These will appear in a new dedicated appendix in the revised manuscript. revision: yes
-
Referee: [Illustration on affine SAT] Affine SAT illustration: the framework claims to identify reusable structures achieving the mutual information bound under bounded restoration and selectivity, yet no explicit construction, restoration-cost calculation, or check that the information-theoretic bound is attained is supplied; this assumption is load-bearing for the theorem's applicability to real catalysts.
Authors: The affine SAT example is meant to illustrate the framework, but we acknowledge that it currently lacks an explicit reusable-structure construction, explicit restoration-cost calculations, and a direct check that the mutual-information bound is attained. In the revision we will supply a concrete catalyst construction for the affine SAT class, compute the associated restoration costs under bounded restoration, and verify that the information-theoretic bound is achieved. revision: yes
-
Referee: [Coupling theorem] Coupling theorem: the lower bound on deployment horizon presupposes that reusable structures exist which realize the algorithmic mutual information without super-Landauer overhead, but the text provides no construction or quantitative check of this condition, leaving the bound's practical relevance unverified.
Authors: The coupling theorem is stated under the assumption that suitable structures exist. The affine SAT illustration is intended to instantiate this assumption, yet quantitative verification is absent. We will augment the example with an explicit construction and quantitative checks confirming that the structure realizes the algorithmic mutual information without super-Landauer overhead, thereby establishing the practical relevance of the deployment-horizon bound. revision: yes
Circularity Check
No significant circularity in derivation chain
full rationale
The paper claims to prove an upper bound on class-specific speed-up via algorithmic mutual information, a minimum Landauer cost for installing that information, and a resulting coupling theorem for deployment horizon. These steps are presented as first-principles results from information theory and thermodynamics applied to reusable structures satisfying bounded restoration and selectivity. No equations or definitions are supplied that reduce the bounds or theorem to tautological re-statements of the inputs, fitted parameters renamed as predictions, or load-bearing self-citations whose content is unverified. The illustration on affine SAT is described as an application rather than a construction that forces the general result. The derivation chain therefore remains independent of its own outputs and is self-contained against external benchmarks.
Axiom & Free-Parameter Ledger
axioms (2)
- domain assumption Landauer erasure sets a minimum thermodynamic cost for installing algorithmic information
- ad hoc to paper Reusable computational structures can satisfy bounded restoration and structural selectivity for a task class
invented entities (2)
-
algorithmic catalyst
no independent evidence
-
class descriptor
no independent evidence
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.