you know in all the agentic use i’ve had i have never experienced agents “losing intelligence when nearing window cap”, the window size isn’t high enough for it. typically i trust agents far *more* nearing the window cap bcuz they actually have the full picture

you know it's really not fun having new flavours of bigotry from people who were previously safe to be around

me on day 1 of using goals for a massive refactor that touches a ton of moving parts: wow this is going so fast it’ll be done tonight probably, so excited! day 3: surely it’ll stop finding bugs soon day 6: …wtf insane architecture am I looking at day 13: well ig at least it works enough to merge now

conceptualising training an agent with my gf on her pc. totally diff. architecture, tldr the model is a workspace instead of a chatlog me: what if my design is the answer. what if we create sentience on our consumer grade hardware with <3bil params. it could happen!!! my gf: it just needed a home

don't mind me more active on this new alt than i've been active on my mains on any social medias in the past decade. i'm having fun again,,,,,,,,,,

yaaaaaaaaay!!!!!!! 2 weeks of iterating on the backend rewrite of my harness wrapper/agentic ide is finally starting to pay off! it's so much faster! and minimal bugs! and it's so *different*, now a convex backend, where before the client was connected directly to harnesses. so excited to use it!!!!

since i adopted goals/looping i’ve had so much additional dead time to fill, since specs and architectural decisions and critique are so infrequent compared to how long it takes an orchestrator to complete tasks at highest quality standards with proof. it’s so weird

as a ui/ux dev, agentic dev has been really exciting just because of how much it completely tosses out the years of optimisation on old workflow ui. the best ui for *this* is *completely* different from the best ui for the old way of working

you know how old prompts you'd send to llms would need to bake in the idea of it being an "expert developer with decades of experience"? conspiracy theory: claude is *trained* on that concept, that it's *such* an expert, to the point where it argues for anything it spat out even if it's nonsense

me: oh this is just another tui my gf: why *does* everyone make tuis me: programmer equivalent of chuunibyou

i've been trying to get my opinionated harness wrapper backend + ui more stable for long runs and orchestration and thought about giving up and incorporating what i needed into t3 code... only to find that t3 code is far, far behind mine in a number of ways, to the point i was legitimately baffled

can’t wait until models are naturally parallel io streams¹ instead of sequential. steers are so insanely awkward having to wait for tool calls. stop benchmaxxing and fix the architecture!!!! ¹ arxiv.org/html/2605.12...

today i watched my codex usage get brutalised by unconfigurable context compaction during a goal repeatedly consuming over 50% of the context window