one researcher plus a pile of agents pretrained a competitive coding model in 3.5 months for under $10m the moat used to be headcount. now it's knowing what to run next how small can a lab get before it stops being a lab?

codex limit resets feel like a gift until you realize you're planning your workday around them i don't want surprise resets. i want limits predictable enough to build a schedule on lower but predictable limits, or bigger limits with random resets?

a 40b open coding model doing 125 tok/s on a macbook and scoring 89.7% on swe-bench pro if that holds up on real repos, the $200 plan stops being the default and becomes the fallback would you run your coding agent fully local if it was 90% of opus?

people who say ai writes bad code are just bad at reviewing it

i'm 30, a dad, shipping Crescitaly.com, crescicode, sendalto and a few more, mostly at early morning looking to #connect with builders active on x drop a "hi" if that's you 🤝

Hot take: if you're a solo builder, Figma is dead weight now by the time the mockup is pixel perfect, claude code could've shipped a working version you can click Figma is turning into a tool for teams to align, not for founders to build what's the last thing you designed before coding it?

deepseek just raised ~$15b. that's not a chatbot budget every time open models catch up, the price of frontier coding drops for everyone if an open model hits opus level next year, are you still paying for max?

currently building: Crescicode, Crescitaly, Crescipulse, Journey, Electrolyte brand and clothing brand. Probably too many. shipping anyway

claude code compacting mid task is where most agent sessions quietly die the summary keeps the plan and drops which files it already touched free fix: have the agent keep a progress.md it updates after every edit, and tell it to reread it right after compaction what's your trick for this?

85% of public claude code configs have zero deny rules we give agents the keys to the repo, the terminal and the .env and just hope 5 minutes, add deny rules for rm -rf, force push and reading secrets. do it before the agent does something you can't undo do you run yours locked down or full yolo?

claude duet is basically "an agent that works while you sleep and asks before doing anything risky" the hard part was never the model. it's trusting it with your inbox and calendar would you give a proactive agent write access on day one, or read-only for a month first?

openai models + codex are now in google cloud. that's all 3 big clouds the model war is turning into a procurement war. most companies won't pick the best model, they'll pick whatever burns down their cloud commit benchmarks don't matter if finance picks the model

honest question for solo builders: do you still open figma before you build? with claude code a working prototype now takes less time than a clean mockup. the mockup is starting to feel like the slow draft

sonnet 5.5 is beating opus 5.5 on the coding agent index. 68.4 vs 66 the cheaper, faster model wins and burns way less of your weekly limit who's actually switching their default model this week?

Only us?

Mamma mia lovely competition

The answer to Dot, Grok Bot is here

I love talking with Argon

Fable 5.5 Here we are letss gooo, you should have it too

Crazy