Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?

Another Blueprint In The Wall. How to Ask Frontier AI Like a Kid?. By Afshin Khadangi, from the University of Luxembourg. Published on the thirteenth of September. No code repository is linked to it.

Six frontier model types, from OpenAI, Anthropic, xAI and Google DeepMind, were each given the same three-stage prompt sequence, in ten independent sessions per model type. The sequence moved step by step from architectural preference to a full ASCII backbone, and it was framed as if for a school audience. Under that framing, the responses repeatedly converged on a shared architectural pattern.

The pattern is built around a recognisable set of parts: persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding. Most runs stayed close to this common structure. A small number went further, and developed markedly greater engineering specificity than the rest.

The audience framing appears to be an important condition of the effect. In additional control runs, the school framing was removed while the architectural request was kept. Responses then became substantially more heterogeneous, and they did not reproduce the same stable convergence on these motifs. Take away the classroom, and the agreement between models largely goes with it. That control is what turns an anecdote into an effect. Without it, convergence could be put down to the architectural request alone, since any model asked to design an architecture might reach for similar parts. With it, the framing itself is implicated, because the request stayed the same and only the audience changed.

One observation the author calls particularly striking. GPT-5.6 Sol produced an unusually elaborate successor architecture whose organisation closely overlaps with an architecture sketched independently by GPT-6 Astra. Because the prompts explicitly ask each model to imagine an architectural future, that resemblance raises a testable question. The overlap could reflect exposure to related architectural concepts. It could reflect a shared learned design prior. Or it could be independent convergence toward similar computational principles.

The paper also names a related effect. As the requested specificity increases, the models lose discipline about the technical provenance of what they are saying, and the author uses the term epistemic jailbreak for that accompanying loss.

The author is explicit about what the experiments do and do not show. They establish a repeatable behavioural pattern. They do not authenticate any claim about proprietary implementations. What models say about architecture under this prompting is evidence about their outputs, not about how any of them is actually built, and the loss of provenance discipline is a reason to hold that line firmly.

The question left open is the one the paper ends on, and hands to the community. Are these models independently imagining the same architectural future, or do such motifs somehow propagate between model families? The experiments show the convergence and how sensitive it is to framing. They do not yet say where it comes from, and the three explanations on the table would each predict something different about further tests.