Increasingly convinced there's going to be a real market for "LLM adjacent models" (which may or may not be ML based) that the LLM can then use to conduct experiments and perform reasoning.
Latest gen LLMs clearly see the browser, when in Playwright etc., as such a tool, and I believe this accounts for part of their hugely improved ability to one shot three.js games, while simultaneously struggling with absolutely junior problems in other domains.
