@amasad I think model/harness oscillations tend to be downstream of our evolving obsessions with what “matters” in an AI world. The Anthropic swell followed the “taste is what will matter” cycle, and both the company and its models are indeed more tasteful.
@mathemagic1an And also that the central value prop was a slightly better browser use tool, which still fundamentally sucked! It’s *good* to have your agent work with you in the browser - but it should be agent native/programmatic use, not the weird (and token hungry) mechanical tools.
@dieworkwear It’s increasingly disincentivized. If you’re smart, motivated, and principled it’s hard to think of a job less attractive than (federal) public office right now.
@EndrewBiz@nia_thinks Raw model + basic harness maybe, but add in a well developed /goal and a mountain of tests and you get top quintile output consistently.
@jaigerbombs@amasad We’re definitely at/close to the point where I don’t feel a huge need to chase the latest frontier models. Fable is unbelievable - and also, at this stage, only a tiny, tiny volume of my work is Fable worthy.
@amasad Great piece, Amjad. Wondering - At what point does a continually improving harness start to matter more than the underlying frontier model? Could a world-class harness consistently outperform a mediocre harness running a model that’s one generation ahead?