The Madeleine, the Model, and the Constraint

A week ago, I left a rather technical comment, using some jargon, on a LinkedIn post, exactly the kind of content that usually receives little attention. Yet, I was puzzled by how widely it ended up travelling.

I asked an LLM to explain why. Among other things, it suggested “semantic analysis,” implying that the platform used deep language understanding to parse comments.

That surprised me. Feeding all daily content into frontier LLMs would be prohibitively expensive. In practice, platforms must rely on much smaller, constrained models. Efficient, yes. But far inferior in understanding, without any so-called emergence. We hear “AI everywhere,” but most real systems run on good-enough models that are often not dramatically different from pre-ChatGPT machine learning.

That triggered a specific chain of memories.

Back in Classes Prépa, a bright classmate once told me that ℚ, the rationals, is not included in ℝ, the reals. Strictly speaking, ℚ is a set of equivalence classes; it only embeds into ℝ up to isomorphism. What stayed with me was not the mathematics, but the shock: things can function perfectly well in practice even when, formally, they are not what we pretend they are.

Years later, I visited a friend who was studying at ENS Ulm. Over lunch, a researcher in biostatistics told us something that also stayed with me: the “best” drug in the lab may not be the one that reaches patients. What gets manufactured is not the molecular optimum, but the compound that is good enough, affordable, scalable, shelf-stable, and regulatable.

Different domains. Same lesson.

Reality is not built from ideal objects, but from solutions that survive constraints.

Models, medicines, algorithms, ethics: what matters is not optimality in theory, but viability in practice.

(Image: A nod to Proust. Because sometimes, a simple chat with a bot acts just like a madeleine dipped in tea: unlocking memories you did not know you still had.)