@teortaxesTex They probably hired a ton of digital artists to do this. At first it’s mind blowing but when you think about how all digital artists has stroke level diffs it makes a lot more sense and is a lot less mindblowing.
@nickbaumann_ Can you folks please fix the codex seatbelt permissons to allow for modification of ical? Rigjt now it fails at managing the Mac local calendar and its all because of a permissions bug or lack thereof for that one framework
@pvncher Dots can’t answer the “hey are you able to stop using YouTube to dj stuff now while I run computer use” in other threads prompts can they lol.
@tonygaorx Ya lol Astra and OpenAI models are absolute shit at vibe hill climbing. It spawns subagents to do the dumbest of work and burn compute while gettin nothing done. Claude is a lot better at give hill climbing but my conclusion was that yeah, guidance and scaffolding is not dead
@PatrickToulme@harris_p10 Funniest thing is @harris_p10 best way to actualitize your will and gain mini fame is to make a testslop bench and do what simonw does and make sure everyone in hacker news knows how well models perform with it. labs don’t move unless you bench it or complain on x w/ a swarm lol
@harris_p10@PatrickToulme Benchmarks don’t ding you for excessive testing and cruft. Feels like a lot of perf gain at least for the current crop of models comes from excessive validation. Idk, test slop is fixable but that cant be such a hi priority in labs over training new functions
@BogdanVko I have a lot of high expectations, I love google’s pretrains and ability to write with weight. The only tragic thing about it is that while Claude was 2dmaxxed, astra was 3dmaxxed, prose written by Gemini won’t go viral because while, to me, it’s the most important axis,