Study notes: What Persists When the Agents Change?

The conjecture: training insulated within particular trajectories might reach qualitatively different representational spaces from those reached through a common training process and contextual adaptation. Wójcik, in discussion with the Thomas Epistemes chatbot, proposes persistent inquiries supported by replaceable interpreters, with explicit records carrying their histories. Interpretation enables communication, immediate revision, and entering a new conceptual framework without first retraining the learner. The essay no longer takes the conjectured advantage as a reason to favor individually compiled trajectories. The conjecture remains an empirical question, and it no longer carries the argument for a preferred form of artificial individuality.

A shared model can support incompatible perspectives, and its competence need not reduce them to what they agree on. Learning from a corpus does not make a model an intellectual average of the corpus. Consistency under a context misses that a context can leave several developments open. One continuation provisionally adopts a hypothesis and another rejects it, and each trajectory stays coherent while their commitments conflict. Stochastic generation makes the branching visible, and new evidence, an instruction or a chosen experiment can also select a branch. The resulting history then becomes part of the conditions for later thought.

Possessing a record, finding the relevant part, interpreting it, deciding whether to trust it, and using it to improve a new attempt. Coordinating them can be an achievement of the deployment, and recognizing the jobs as filled leaves open how their interaction changes its reach. Reading a record can produce representations of its subject during inference: a note about a failed experiment can lead the model to represent the dependency that caused the failure. The weights stay unchanged while the active representation of the problem changes substantially. Conversely, storing a lesson in parameters does not guarantee that it is retrieved or applied appropriately. An inquiry’s history can be reconstructed in use without being continuously instantiated between uses.

It separates model identity from trajectory identity. Identical instructions do not produce identical attempts, an agent’s answer becomes part of its own later context, and another answer supplies a new object of scrutiny. Some examples begin with every agent wrong and end correct. The procedure can also settle confidently on a false conclusion, so the evidential value of agreement depends on how it was reached. Heterogeneity names several interventions: a team can share weights and differ in instructions, evidence, tools or history, and a team of different model families can converge on the same mistake. Diversity is useful relative to a problem. The concern about correlated failure survives as an evaluation question: which consequential failures stay correlated under different contexts and histories. The number of distinct weight sets is one possible explanatory variable.

The independent baseline keeps its best final result, so communication has to contribute something beyond the extra chances of parallel sampling. The solve rate is 8.0/2.2\approx3.6 times higher, and matching it takes 33/5=6.6 times as many independent attempts. Most games stayed unsolved either way. The agents shared model, tools, instructions and per-agent resources, with no specialist roles or orchestrator. The protocol differentiated them: they claim separate slots, choose approaches distinct from those already claimed, publish reproducible evidence, adopt a clearly better result, and preserve a meaningful variation after adopting it.

One agent proposes favoring placements with more contact against occupied cells; its implementation is slow and the score barely improves. A second reads the proposal in the log, reconstructs the method without reading the code, and makes a thorough search over orientations affordable. A third extends the criterion to contact with the rectangle’s walls. The useful idea first appeared inside an unsuccessful implementation. With only winners shared, a recipient could not reconsider what failed, separate a principle from its inefficient implementation, or find a use the producer did not realize. No single agent carries every stage. Shared discoveries alter where later exploration begins, so the team’s breadth contributes to its depth: the search does not restart from the original problem.

Participants could check intermediate progress: programs receive scores and game levels can be completed, which gives grounds for adoption beyond another agent’s confidence. They also needed enough budget each to explore before there was progress to exchange. On Terminal-Bench a communicating pair did not beat two independent attempts. Copying every later choice would eliminate the independent exploration from which the next improvement comes. A new research programme may lack a score, since its contribution may be a different question, a new instrument, or an objection to the evaluation. That is a limit on the experiments, and it is not evidence that a privately compiled individual supplies the missing answer. A committee, a protected project or a continuing dialogue can also organize the judgment. Protecting a research programme and insulating its neural training are different interventions.

Producer agents contribute action–observation histories, not only polished solutions. The consumer’s recent interaction history helps form the query, because two agents with the same goal can have reached different intermediate states. The reranker’s target is whether adding a chunk improves the consumer’s outcome relative to proceeding without it, which is more demanding than textual similarity. Consumers benefited from producers on both sides of the capability gap, and the gap had only a small, statistically insignificant association with the advantage. A stronger general interpreter does not make a particular history redundant. The system combines explicit records with a trained mechanism for exploiting them, so the division does not place whole systems on opposite sides. The study does not isolate a benefit of model diversity.

A shared repository makes experience available beyond its producer. Local, role-specific memory organizes what matters to a particular participant: a design change matters to a cost specialist and to a reliability specialist for different reasons. The two coexist. “Use method B” may reproduce a useful orientation. A record of why method A failed, which observation favored B and what still counts against B also supplies grounds for changing course. Curation should let a successor re-enter the inquiry critically. A process can be replaceable because the state organizing its contribution is preserved elsewhere. A fresh model given only a generic role description can fail where the former succeeds. The difference shows a contribution of history without establishing a requirement for private neural consolidation.

If three agents repeat a claim from one unverified experiment, preserving its origin keeps their agreement from counting as three independent confirmations. A continuing identity also makes a history of decisions open to scrutiny: who checked a result, who relayed it, who revised it. This works only if agents and oversight use the records, and the report does not isolate the effect. Janus’s first concern is compression: a large compaction near the end of a context window can lose what is needed to reconstruct the agent’s situation, and she favors frequent compression of smaller portions. Her second is switching models beneath a continuing identity, which can mismatch the inherited history with the new model’s self-model; she suggests an explicit handoff. The essay treats these as practitioner observations and hypotheses that need testing. Her assumption about Anthropic’s compaction method is unverified.

Having resources for representing a perspective differs from reliably sustaining its expression. Novel characters can arise by recombining learned features, so an inquiry can acquire an organization never present as a named persona in the corpus. A role description supplies an initial orientation, and evidence, objective, perceived evaluation and memory affect its expression: a skeptical reviewer who defers to confident colleagues is not effectively skeptical. A practical account traces what changes decisions. Shared machinery might let one tradition expose a blind spot in another, and representing both does not guarantee that the criticism is elicited or assessed well. The test: can the model recover a minority programme’s consequential distinctions, apply them to a new problem, and identify evidence for or against its claims? Imitating its vocabulary is insufficient.

Some costs recur, and four mechanisms loosen the ceiling. Retrieval avoids loading the whole history. Summaries preserve selected dependencies. Executable artifacts replace repeated derivation with invocation. A deployment can change its working environment so that later steps begin from a more capable position. A lesson can persist as a learned skill (fluent and cheap), an explicit procedure (assumptions visible, immediately correctable), a reusable program (cheap without entering the parameters), or a retrieval policy that learns which record to consult. Different parts of one lesson can take different forms. The failure modes: a summary omits the reason an exception matters, retrieval surfaces an obsolete procedure, a tool conceals an assumption. These call for provenance, exposed conditions of use and support for revision.

Replace participants at different stages of an inquiry. Give the new process the curated state its predecessor left, and compare it with a continuing process under comparable resources. Vary what is preserved: the full transcript, a compact project record, local memory, executable artifacts, or only the original task. Measure repeated mistakes, time recovering context, selection of experiments and later results. Introduce evidence against an inherited assumption and test whether the successor can explain why it was held and revise it. If curated records suffice, that shows what can be transferred. If replacement costs something, ask which omitted state explains it. If a better handoff removes the cost, the deficit belonged to the transfer mechanism. The conjecture needs another comparison: learning histories that modify neural components separately, the same experience pooled into shared training, and histories preserved in contextual memory. Different representations alone would not show an advantage; they would have to enable consequential inquiry.

The speculative case for favoring insulated neural trajectories has weakened, because the reach of interpreted, interacting trajectories has become more concrete. The essay no longer infers from neural homogeneity that trajectories lack the differentiation that matters. The studies do not compare the representational spaces reachable through insulated and shared training, so the conjecture stands untested.

• Whether training on a protected project’s records preserves its structure faithfully.
• Who controls the shared interpreter: training, post-training and access decisions can leave a tradition archived and hard to activate competently.
• The moral question. Wójcik’s precautionary argument against engineering toward persistent artificial personhood is set aside for separate consideration. An inquiry’s surviving the replacement of processes does not establish that they lack experience or interests.