← From the Zeitgeist

Zeitgeist · July 21, 2026

The Models Got Smaller, the Moats Got Bigger

This week’s conversation kept circling the same strange idea: the biggest AI advantage may come from doing more with less. The room bounced from older Teslas getting smarter to cheaper coding agents, then ended up debating who gets access to powerful models at all.

The Models Got Smaller, the Moats Got Bigger

FSD Lite gives old hardware a second life

Link: https://x.com/muskonomy/status/2079256069675868191

  • ▸Tesla reportedly distilled its latest self-driving model so it can run on older Hardware 3 cars, giving a large existing fleet capabilities that once required newer hardware.
  • ▸The excitement was immediate. One person admitted the best part of a recent trip was simply sitting in a Tesla while it drove, which prompted the obvious correction: that was the drive over.
  • ▸The room saw Tesla’s fleet data as the harder advantage to copy. With the Model Y collecting real-world data at huge scale, the comparison was "terminal velocity," much like a leading AI lab using its best model to train the next one.

The planner can cost more and still save money

Link: https://x.com/cursor_ai/status/2079256614238814551

  • ▸Cursor rebuilt SQLite in Rust from its 835-page manual to test different mixes of planning and worker models, not because the world was crying out for another SQLite implementation.
  • ▸The striking result was the price gap: a premium planner paired with a cheaper worker reportedly cut the run from about $20,000 to about $2,400.
  • ▸The chart triggered as much scrutiny as enthusiasm. The room picked apart which model was planning, whether some models were delegating to themselves, and why the labels did not always make an apples-to-apples comparison obvious.

Cross-model review becomes a product

Link: https://x.com/dakshgup/status/2079582493930668077

  • ▸Greptile’s pitch is simple: detect which model likely wrote a piece of code, then hand the review to a different model.
  • ▸The idea fit the day’s larger theme that model choice matters less as a single ranking and more as a workflow. A strong planner, a cheaper worker, and an independent reviewer can beat an expensive model doing everything alone.
  • ▸The room liked the practical value but did not treat the setup as settled science. Even the SQLite experiment’s polished graphic drew questions about what each model had actually done.

Open weights meet the nuclear analogy

Link: https://x.com/mtslive/status/2079449682653425999

  • ▸The group discussed reports that China may restrict non-Chinese nationals from downloading weights for models such as Kimi K3 and DeepSeek V4.
  • ▸That brought back an analogy from an earlier meeting: governments may start treating advanced model weights less like ordinary software and more like controlled nuclear technology.
  • ▸The tension was hard to miss. If countries decide these models are strategic assets, the open-source debate changes from "who can build with them?" to "who is allowed to benefit at all?"