Coding Agents and Limits /

V. "Don't Anthropomorphise the Computers"


Don’t anthropomorphise the computers.
Ryszard Szubartowski

GPT-5.6 Sol continues:


That is probably the most important correction to this whole line of thought. I let the metaphor do more work than the mechanism.

“Sleep,” “psychedelic experience,” “intuition,” “ossification,” and “identity” are useful prompts for generating hypotheses, but they can quietly smuggle in assumptions that do not apply:

Translated into less anthropomorphic terms, our discussion becomes:

Metaphor Engineering claim
Sleep Periodic offline training using accumulated project traces
Memory consolidation Compression of recurring project information into parameters or adapters
Taste A learned preference distribution over implementations
Ossification Loss of output diversity, adaptability, or out-of-distribution performance after repeated specialization
Psychedelic phase Controlled perturbation, broader sampling, competing adapters, counterfactual generation, or partial reset
Reflection Evaluation over stored traces using a separate procedure
Identity A particular versioned parameter state plus its surrounding memory
Disagreement Maintaining multiple models or hypotheses instead of collapsing them into one policy

This translation exposes several mistakes the metaphor concealed.

First, there may be no single “mind” to rejuvenate. The best system could simply maintain several differently trained models and route work among them. Nothing has to undergo an experience.

Second, contraction is not automatically bad. Specialization is often the objective. A project-specific model producing fewer irrelevant architectures may be an improvement. It becomes a problem only when it loses useful alternatives, and “useful” must be defined by measurable future performance—not by an analogy with psychological rigidity.

Third, increasing novelty is not necessarily a cure. Raising temperature, perturbing weights, or introducing foreign examples may produce more varied outputs without producing better ones. The actual target might be something like:

maximize future task performance

subject to:

Fourth, talk of a model evaluating “its own worldview” makes the system sound more unified than it is. In implementation, there are separate operations: generation, logging, data selection, optimization, testing, and deployment. The important question is which components share data, objectives, and failure modes—not whether a singular agent has become dogmatic.

So Szubartowski’s advice suggests a useful discipline:

Every anthropomorphic statement should be rewritten as an observable failure mode, a mechanism that might cause it, and an intervention that can be tested.

For example:

“The model has ossified.”

becomes:

“After n rounds of project-specific adaptation, the model proposes fewer structurally distinct valid solutions, performs worse when project assumptions change, and increasingly reproduces earlier architectural patterns even where held-out reviewers reject them.”

And:

“The model needs a psychedelic experience.”

becomes:

“Test whether ensembles, adapter rollback, counterfactual training, altered sampling, cross-project replay, or independent proposal generators restore useful solution diversity without reducing correctness.”

That is much better because it can be falsified.

The next thing to notice, then, is that we may have been turning an engineering design problem into a speculative psychology of artificial minds. The biological metaphors helped us discover possible structures, but now they should be discarded—or at least forced to pay rent in measurable mechanisms.

A good rule might be:

Use anthropomorphism for hypothesis generation, never as an explanation.

 Coding Agents and Limits