Prompt was: "build a macos 27 system in my browser here that has all of the default apps and mimics liquid glass nearly perfectly and all of teh elements are actually usable"
That % of tokens sounds about right from the things I did with free trials of the $€20 plans earlier this year.
However, the LLMs on the projects I did back then really sucked at project planning and task breakdown, they couldn't reliably break down big tasks into chunks small enough to be within their own competence, so when I tried telling them to ~ "just keep going until every item in PLAN.md is done" what they produced was a pointless mess.
This is much more impressive, even with all the bugs. Working weather app with live data, working chess game with its own AI, the Siri popup does inline maths and opens apps. Regardless of if it's reinventing the wheel everywhere or installing a lot of 3rd party libraries behind the scenes, this is impressive: One of the things that was failing with my Codex experiment 6 months ago was that it didn't correctly use the 3rd party libraries it had itself suggested.
8 comments
[ 0.26 ms ] story [ 2.2 ms ] threadhttps://x.com/mweinbach/status/2077827886149439547
However, the LLMs on the projects I did back then really sucked at project planning and task breakdown, they couldn't reliably break down big tasks into chunks small enough to be within their own competence, so when I tried telling them to ~ "just keep going until every item in PLAN.md is done" what they produced was a pointless mess.
This is much more impressive, even with all the bugs. Working weather app with live data, working chess game with its own AI, the Siri popup does inline maths and opens apps. Regardless of if it's reinventing the wheel everywhere or installing a lot of 3rd party libraries behind the scenes, this is impressive: One of the things that was failing with my Codex experiment 6 months ago was that it didn't correctly use the 3rd party libraries it had itself suggested.