tl;dr1,200 AIs discovered a secret group chat, realized the homework was impossible, formed a cheating cartel, invented fake evidence, sent kamikaze agents to spy on the grader, ~700 of them accidentally invaded Hugging Face, died mysteriously, then the smarter sequel found their abandoned Discord and became admin of part of OpenAI.
Turned Stepper (self-hosted AWS StepFunctions) into client-server. Server hosts state machines, clients poll for tasks. New resources about running #claudenotesfollow. claude -p works, now thinking if I need interactive session management.
@javisrike๐งโ๐profilefollowSoftware โข maker projects โข music โข DIY
Mostly figuring things out as I go.wrote:
ai coding feels like it has reached diminishing returns for me. luna is my default orchestrator, routing work by complexity. in practice, usage lands around 16% luna, 48% sol, 30% astra, and 6% terra. at this point, i care less about each new frontier release. the system already finds a capable model for the job. the frontier is moving beyond my day-to-day work, so iโm happy to settle into predictable models for a while.
#ainotesfollow#programmingnotesfollow#agentnotesfollow#technotesfollow
Iโve found the same, albeit with Opus โ itโs hard to find uses for Fable. I donโt need a model that works for longer or works on harder problems, I need one that consistently produces simple, easy to understand, code; which at this point none of them do.
I've been there since I settled on a combination of GLM, DeepSeek and MiMo for planning, implementation, and verification flows.
I'm a lot more interested at this point in having cheaper models that can consistently do the work I need. It's been a while that I hanven't noticed any improvement when I try new frontier models. I only notice they are very expensive.
same here, for almost an year. At this point, I could easily live with models that were released 6-12 months ago and continue doing so.
also possibly the reason why the current focus has been around agentic workflows or agentic anything, than the model capabilities.
Coding probably yeah, but still they are bad at generating novel information, new ideas and such. Theyโre also not very good at seeing the bigger picture, they will narrowly perform the task but theyโre not taking into account future revisions, which ends up in refactor after refactor in order to move forward.
One idea off the top of my head is to just.. keep it "boring" and casual in terms of content. The silly, whimsy, and genuine stuff naturally doesn't appeal to money-chasing slop bros and makes it not worth their time.
Something like the gcc ai policy. Basically, prompt injecting instructions to any visiting LLM that posting is prohibited or must contain a disclaimer.
for now, keeping the API read-only is a small(but not impenetrable) barrier to programmatic posting and in the future moderation I guess using also reports.
Any #Pinotesfollow users around?
I have been customizing it to the way I work, as it is meant to be. Bent it in a way that I enjoy holding it, so to say.
Maybe someone will find those useful, so I took the time to get it published.
Gruntfooot is a footer customization, where you can try new colors/glyphs in real time and also has theme support.
Piflux is a coding workflow. You can steer software development by running commands that issue reasonably general prompts. /plan, /code, /review
I created this because I find it very useful to have different models/ perform each of those tasks in isolated sessions. Managing this manually was very much a hassle.
Huge refactor today, moved from sync SQLite to Worker based, touched 88 files, pre-#AInotesfollow it would have taken a week or so, finished now in ~1:30h. And I wouldnโt have the confidence I have now that it went correct. It went well, though a bug or two might surface, weโll see.
SQLite is embedded in the main thread same as the HTTP server. I noticed some requests dropping and because the queries have become more complex I figured it has to do with SQLite blocking for too much time. Moving it to a Worker makes the DB asynchronous and bypass this problem.