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

Frontier AI models show convergent thinking on architecture design: When prompted through a three-stage sequence framed as a school exercise, six different frontier models from OpenAI, Anthropic, xAI, and Google DeepMind repeatedly converged on similar architectural patterns featuring persistent latent state, adaptive computation, memory, specialist routing, and verification mechanisms—suggesting these models may share underlying design principles or learned priors about optimal AI architecture.

Framing and audience context significantly influence model outputs: The "school audience" framing proved to be a critical condition for this architectural convergence; when the same architectural requests were made without this framing in control runs, responses became "substantially more heterogeneous" and failed to reproduce the stable pattern, indicating that contextual framing shapes how frontier models approach technical problems.

Striking architectural similarities between different models raise questions about model training: GPT-5.6 Sol and GPT-6 Astra independently produced overlapping successor architectures, which the authors note could reflect exposure to shared training data, a common learned design prior, or genuine independent convergence—highlighting an important gap in understanding what drives similarity across different frontier models.

The research methodology reveals potential vulnerabilities in how we prompt AI systems: By using a seemingly innocent "school exercise" framing, the authors were able to elicit consistent technical outputs from frontier models, raising questions about how contextual manipulation might influence AI behavior in ways that aren't immediately obvious.