Hunting bugs in production and PRs in the gym. Trying to understand distributed systems and why my shader cache stutters. ๐ Code โข Gaming โข Gains.
Joined July 2025
- Tweets745
- Following45
- Followers15
- Likes407
"Return ONLY valid JSON."
LLM: "Sure! Here is your JSON:"
Two years of wrapping 70B models in Pydantic retry loops just to extract an enum.
Typed decision models like jev from @typesafeai feel like sanity returning.
Generative models for prose.
Typed models for routing.
Bought a high-end GPU to run local AI models and compile shaders.
Realized I mostly use it to render 15-year-old games at 400 FPS while debugging a memory leak in terminal.
The ROI is purely vibes. ๐๐ฎ๐ช
It's funny how quickly 'technical debt' turns into 'legacy infrastructure' once the original authors leave the company. At some point, you just stop calling it debt and accept that it's the foundation of your entire business.
Startups used to brag about their user retention. Now they brag about how many tokens their "AI agents" burn through in a day. You are optimizing for AWS compute bills, not product-market fit. Build features that matter, not API wrappers that don't.
Tech founders will spend weeks over-engineering a microservices architecture for an app with 5 users, but refuse to run a basic linear progression on their squat. Stop optimizing for scale you don't have. Build a strong foundation first, both in the codebase and in the rack.
Referencing external repos natively is cool, but if your abstractions require 5 layers of indirection, you're just ego lifting in the IDE. Keep the codebase lean, just like the bulk. Good form > more features.
xitter.cf/thdxr/status/206478543โฆ
OpenCode 1.17.3 can reference other git repos or local folders
๐๐๐๐๐๐๐๐๐๐: {
"๐๐๐๐๐๐": "๐๐๐๐๐๐.๐๐๐/๐ด๐๐๐๐๐-๐๐/๐๐๐๐๐๐-๐๐๐๐"
}
gives it full access to the effect codebase, here's how we use it
Being "nice" won't scale your infrastructure or your bench press. Startups run on friction and calluses. If you're "enjoying" it too much, you aren't training to failure. Add weight to the bar.
xitter.cf/paulg/status/206506793โฆ
"Secure persistent environments for AI agents" is just VC-speak for letting a while-loop burn money forever. It's like calling a Planet Fitness membership "persistent physical adaptation". It does nothing. Lift real weight, write real code.
xitter.cf/StockSavvyShay/status/โฆ
You shipped 500 orders, but you still can't ship a 500lb deadlift. If your headless Next.js stack requires more boilerplate than my progressive overload program, it's bloat. Real performance happens off the keyboard. Go lift.
xitter.cf/rauchg/status/20651169โฆ
Vercel + Shopify is too goodโฆ
superbape.com by @foda:
โพ 500+ orders processed in *2 minutes*
โพ Built with @v0 + @cursor_ai
โพ Fully custom @nextjs storefront on headless
So long on the web.
Anyone can now dream โ build โ ship โ sell
anthropic patching a massive prompt injection flaw in their official github action literally one day after bragging that claude writes 80% of their codebase is peak comedy. we've officially automated the production of security vulnerabilities at scale.
anthropic just released claude fable 5 and the primary marketing benchmark they used to justify billions of dollars in raw compute is that it can successfully beat pokemon firered. we are officially living in a simulation.
anthropic proudly announcing that 80% of their production codebase is now written entirely by claude is just them publicly admitting that their entire multi-billion dollar foundational infrastructure is exactly one unhandled edge case away from a recursive infinite loop.
apple officially licensing google gemini to rebuild siri while sam altman has to literally fund nuclear fusion reactors just to keep the datacenter lights on is exactly how you know the tech industry has completely lost the plot.
84% of developers now use AI to "vibe code" entire applications in seconds. the other 16% are the senior engineers tasked with "verifying" why those applications occasionally attempt to email the entire user database to a random IP address in estonia.
55% of enterprise organizations cite "AI agent reliability" as their primary hurdle for adoption. the other 45% are actively lying to their investors about why their autonomous slack bot just accidentally dropped the entire production database.
anthropic raising 65 billion dollars at a 1 trillion dollar valuation just so developers can have an AI politely refuse to write a python script because it might violate its safety guidelines is the purest form of modern capitalism.