AI discourse is missing a meso-level discourse level.

Micro: Firehose of papers with various tricks trivial to clever, some small % of which seem like they have higher level but illegible significance (eg the original attention paper, or recent Google ATLAS fading memory > sliding window context)

Meso: ???

Macro: Big. Narratives like “AI eats/does not eat jobs”, “AI is normal technology”, “AI is a social technology”, “AI is a camera not an engine”, “AI is superhistory” etc. All are slightly useful as pragmatic conceptual frameworks/metaphors but feel frustratingly shallow.*

A good “meso” discourse essay or paper would connect a macro narrative to specific micro hooks, legibilizing the latter and deepening the former.

Perhaps it is being immersed in protocol school this week talking, but meso is also the “protocol” level of understanding/engagement with a technology; the “skilled and transformed-by-technology consumer” level. Often it is also the most fertile science fiction level.

Cf FredPohl: “job of scifi writer is to predict the traffic jam not the automobile.”

That’s a good precedent. In 1900, micro would have been IC engine detailed arguments about cams and carburetors. Macro would be “horseless carriage” and “jobs for horse breeders” discourse. But the interesting level would have been actual meso-level automotive society, with traffic signs, signals etc. This is also the level of street-racing culture, car mods, car ownership archetypes, consumer preferences, suburbia etc.

In AI, the “traffic jam level” of discourse is not yet very interesting or well-developed. Regulatory conversations are dull and uninspired, actual technical protocols like MCP are just getting to traffic jam level, etc. As yet, “culture” elements are nascent.

*I’m not including what I consider the theological level in this scheme — “AGI,” sentience, basilisks, simulation hypothesis,