Commonplace builds a theory builder and tests whether it learns

Type: types/note.md · Tags: learning-theory, self-improving-systems, theory-builder

Commonplace builds a theory builder: a continuing system whose theories are stated in localized units, guide what it does through what they say, are criticized for what they say, and are revised through iteration, where the result of criticism shapes the next round. The definition has no success condition. Whether the system learns, in the sense of improving its capacity for future action (Simon's criterion), is the hypothesis under test. Membership does not establish it.

Beyond the four conditions, Commonplace builds for the high end of two graded properties, addressability and persistence, and holds model weights fixed as a study condition. Whether these commitments pay is an empirical question, stated as three conjectures below. One asks whether criticism, the core of membership, pays against trial and error. The other two ask whether each high end pays against a builder with less of it. The precedents note supplies the fuller attribution.

Research program and development path

The first research claim is that an autonomous theory builder learns, which it could do with a fixed method. Commonplace pursues it alongside a practical goal, an LLM wiki that people use for their own work, and the two reinforce each other: learning is what makes the wiki's memory worth keeping, and real use supplies the problems and criticism the builder learns from, as the second difference below explains.

Commonplace's approach to building a theory builder is reflective from the start: the builder criticizes and revises its own stated method, and people perform many of its operations until they move to computation. The approach rests on the return to investing in the method: an improvement to the learning machinery is reused by every later episode, so over a long horizon it can pay more than immediate learning. That return requires method changes to compound, making later improvement better, which is the claim of recursive self-improvement. Reflection is not sufficient for it, and it is not necessary for recursive self-improvement in general: Schmidhuber's realizations select self-modifications by reward rather than by criticizing a stated method. Commonplace takes the reflective route, where compounding comes through criticism of the stated method.

The program sits within the broad research program on recursive self-improvement developed by Schmidhuber. His account starts from programs: "True RSI is about encoding the initial learning algorithm in a universal programming language", whose instructions can modify that code itself (RSI retrospective). His example of such a language is a "recurrent neural network or RNN". On this reading a network's weights are a program, and his neural realizations let a network rewrite its own weights and so its own learning algorithm.

We extend the same reading one step further: a natural-language prompt that an LLM interprets is also a program. The fixed-weight model and harness are its interpreter, and the prompt can describe how to revise prompts, including itself. Unlike code, such a program has no formal semantics; the interpreter's judgments fix what it does where the text leaves choices open. This is our interpretation, not Schmidhuber's claim. Our chosen realization therefore uses natural language as the retained, revisable carrier of knowledge and learning procedures, with LLMs supplying interpretation and criticism. Retained theories can guide diagnosis, test selection, and procedure revision; the research objective includes making subsequent improvement work more productive.

We start with an incomplete, criticizable account of how to learn and use it through interpretation. The LLM and harness supply executable machinery; the methodology need not specify every operation before it can be tried. This reduces the up-front specification needed to expose a methodological conjecture to failure. Whether the resulting judgments support useful improvement remains an empirical question.

Two differences from earlier realizations may explain why this one could get further; both are conjectures, not results. First, the seed. Schmidhuber's seed improver must contain, in executable form, all the competence the first improvement needs, and the general routes to that competence, program enumeration and proof search, are expensive; the realizations that ran stayed within bounded domains. With an LLM the seed is a delta over pretrained competence: it directs what it need not encode. This difference is shared with every LLM-era attempt. Second, the start. An improvement to the method pays through the later episodes that reuse it, and a bootstrap is the running system itself, so starting from a working system supplies that stream from the first day, with people supplying the functions not yet automated. Any program could in principle close its loop with human judgment. Sustaining that requires a system people want to use, so that the judgments are part of work they would do anyway; Commonplace has this, because its operators build and use the knowledge base for their own work, and that use is the initial selection environment where no fixed oracle exists. The price is warrant. A proof or a reward-rate guarantee certified each accepted change in the earlier realizations. Here the interpreter's judgments are hidden, and human contributions must be recorded rather than credited to computation.

Popper's epistemology organizes this development process: treat the methodology's claims as tentative theories, criticize their content, test their consequences, and let the resulting problems guide further work (formulation and criticism). This follows from our choice to develop the methodology through conjecture and criticism, which makes the method part of what the builder criticizes (the reflective qualifier). Natural language alone does not make a procedure Popperian; an instruction, a theory about its effects, and a criticism have different roles even when they share one artifact.

Codify parts when their intended meaning and operation become sufficiently clear and the commitment is useful. Codification assigns consequences through a symbolic consumer; it does not merely make prose more precise. Other parts can remain interpreted and criticizable. Codification can therefore follow learning, part by part, without making the underlying theory true or exempt from criticism. Whether it preserves useful behavior while reducing cost or error is a separate question from whether the interpreted methodology improves learning.

This direction does not add a condition to the definition. Interpreted methodology and selective codification are how Commonplace realizes a theory builder, not what makes a system one.

The continuing system

New work begins with what the system retained from earlier problems. Condition 4 of the definition, iteration, requires only that the result of criticism shape the next round, which can happen within one run. Persistence is graded: within one reasoning episode, across the rounds of one run, across runs, or across problems and sessions. We choose the high end: a library of retained theories that later work on other questions takes up. A failed test may prompt revision or replacement. Surviving a serious test adds to the grounds for relying on a theory as background for further theories, or for spending less effort repeating tests of the same vulnerability. The theory remains tentative; its content can stay unchanged while its assessed support and future use change.

The definition requires only that some unit carry content, even if that unit is the whole theory. We choose the high end of addressability: assumptions, scope, and parts can be inspected and revised individually. Indexed traces that expose conjectures, criticism, and testing results are an implementation of this retained knowledge, not a separate comparison. We also hold model weights fixed in the research program. The theories, testing record, and other retained state remain parts of the builder and continue to develop. Fixed weights rule out parameter updates as the source of improvement; they do not establish that a particular retained change caused it.

The definition also admits theories criticized and replaced whole, theories rebuilt from retained criticism, builders whose results do not outlast a run, and builders whose weights change. Our choice is one arrangement within it.

Three conjectures

Criticism of content (membership core). An arrangement that formulates criticism of what a theory says and supplies it to later steps may yield more learning from a failure than an arrangement that generates variants and selects them by score, with no formulated reason for a failure. The second arrangement is trial and error, outside the definition (condition 3), so this conjecture asks whether the criticism condition pays. A correct diagnosis can explain why a claim failed and direct subsequent search. The intended contrast is the effect of formulating and supplying criticism; a model proposing variants may also criticize them unobserved. More elaborate criticism need not supply a better diagnosis.

Addressability (graded). Criticism that identifies a suspect assumption or part may yield more learning from a failure than criticism that leaves the target undivided. Identifying a candidate cause can focus investigation and help preserve useful knowledge. The comparison is with a builder whose criticism is directed at the theory as a whole, which can still constrain its successor. Both arrangements are theory builders: a theory with no stated parts that is criticized and replaced whole meets the definition at its minimum, and it is the baseline for this conjecture (checks, case 26). The target of criticism is distinct from edit size: a diagnosis of one part can lead to rewriting the whole theory. Localization can be mistaken, and a revision may overturn a core assumption.

Persistence (graded). Keeping more of the work of conjecture and criticism, for longer and across more problems, may reduce cost at comparable decision quality. The criticism conjecture concerns the contribution of formulated criticism; the persistence conjecture concerns the cost of retaining versus reconstructing that work. We compare retaining the theory and its testing record across problems with builders whose results persist less, and with two reconstruction alternatives: rebuilding a theory from retained criticisms, and rebuilding from records containing only inputs and outcomes. The first alternative is still a theory builder, so it asks what keeping the assembled theory buys. The second carries nothing that criticism produced across runs; it is the baseline for the persistence conjecture (checks, case 4) and asks what keeping the work of criticism buys. A reconstructor that states, criticizes, and revises within a run is still a builder at that run's grade, and it may learn in the ordinary sense. Retention may save reconstruction, but still requires retrieval, interpretation, and maintenance. Compare total cost at comparable quality and quality under matched budgets. Vary record volume or context capacity to test whether bounded context increases the benefit. For example, an earlier result may say that retrying a timed-out write can duplicate an operation already committed. A later task about duplicate charges after reconnecting may not name that call or reuse its wording. An index retaining “applies when completion is uncertain” can save reconstructing the applicability condition. Deciding whether the current case meets it still needs interpretation or suitable structured inputs. If the trace already exposes that condition, the index may save only lookup. This is an illustration; any cost comparison must count creating, maintaining, and consuming the index, and give reconstruction competent access to the records.

Evidence

Reading the artifacts can establish that criticism was formulated. Showing that its content affected later action requires causal evidence. Compare relevant content with altered or mismatched content while controlling its form. Removing the artifact alone cannot distinguish a content effect from the effect of supplying text at all. These are ways to investigate membership, not additional conditions of the definition; see the evidence ladder and experimental contrasts. Failure to detect an effect leaves membership unestablished by that test; it does not by itself establish absence. Demonstrating the mechanism establishes membership at most. Learning needs a separate assessment of whether the system's capacity for future action improved. Observed action provides evidence of that capacity; lack of exercise alone establishes neither its presence nor its absence. Failed attempts can occur within a process that learns; an advantage over alternative approaches is a further claim.

Experiments compare specified arrangements of evidence, criticism, and persistence. Holding the model fixed does not hold its internal processing fixed when its inputs differ. Unobserved criticism in a comparison arm remains possible; results establish differences between the tested arrangements, not the presence or absence of that internal process. Arrangements designed to rule out unobserved criticism, by withholding failures from the proposer or by proposing variants without a model, also change the available evidence or the proposal mechanism.

Scope

None of these advantages is established. The conjectures motivate experiments without making our choice a prescription for other systems. How much decorrelating the critic from the proposer helps, and whether retained criticisms can be kept distinct from a reconstructed theory in practice, remain open questions.

Relevant Notes: