Another Blueprint In The Wall: How to Ask Frontier AI Like a Kid?
| Source: arXiv AI
Tags: frontier models, GPT-5.6, GPT-6, Anthropic, prompt engineering, LLM architecture
When prompted with a school-audience framing, GPT-5.6 Sol, GPT-6 Astra, Claude, and models from xAI and Google DeepMind converged on the same architectural pattern in 10 independent sessions each — persistent latent state, adaptive compute, memory, specialist routing — while removing that framing eliminated the convergence.
Details
Khadangi runs a repeatable behavioral experiment across six frontier model families (OpenAI, Anthropic, xAI, Google DeepMind) probing whether models share latent architectural intuitions about AI design. A three-stage prompt progressively asks for architectural preferences then a full ASCII backbone, framed as explaining to a school audience. Across 10 independent sessions per model type, responses converged on a shared pattern: persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed decoding. Removing the school-audience framing while keeping the architectural request caused responses to become substantially more heterogeneous — the framing is a key condition for convergence. A striking specific finding: GPT-5.6 Sol produced an elaborate successor architecture whose structure closely overlaps with one independently sketched by GPT-6 Astra. The paper introduces the term epistemic jailbreak for the loss of technical discipline as requested specificity increases. The paper does not authenticate proprietary implementation claims and is explicitly exploratory. The open question it raises — do models independently converge toward the same computational principles, or do architectural motifs propagate between families? — is a genuine empirical puzzle, but answering it would require access to training data that is not public.