Debugging and Startup Launch

t was really not working as planned, need to debug, but I did find that the Qwopus3.6 9b model would actually run on a Macbook Pro M4-48Gb as the inference with llama.cpp and it does tool calling quite well but still need to debug but I did make something else at the same time I vibe coded an app with OpenCode The SSE ingests multi-source data (RSS, Markdown, Reddit), maps it into an Agentic Knowledge Graph, and synthesizes content through Quantified Personas—all running entirely on your local infrastructure. *No cloud. No APIs. No data exfiltration.* it works too using docker so you could use it too since all it has to do is work in concept and if the speed is an issue you can scale with deploying for free on Google Cloud that is how you launch a startup for free using the free tier access to google cloud or oracle or AWS or what I use is Vercel which is Meta's project

:::writing{variant="social_post" id="48271"} t was really not working as planned. Needed to debug a lot of it. But I did find that the Qwopus3.6 9b model actually runs on a MacBook Pro M4 48GB using llama.cpp, and surprisingly the tool calling works pretty well already. Still debugging though. At the same time I vibe coded another app with OpenCode: The SSE ingests multi-source data (RSS, Markdown, Reddit), maps it into an Agentic Knowledge Graph, and synthesizes content through Quantified Personas — all running entirely on your local infrastructure. No cloud. No APIs. No data exfiltration. And it actually works. Dockerized too, so technically anyone can run it. The point is not that it is infinitely scalable on day one. The point is that it works in concept. Once you have that, scaling is just infrastructure. If inference speed becomes an issue, you can scale almost entirely off free-tier infrastructure: - Google Cloud free credits - Oracle Cloud free ARM instances - AWS free tier - Vercel edge deployments People massively overcomplicate launching startups now. Half the battle is just getting the system operational enough that reality can start applying pressure to it. You do not need a hyperscaler budget anymore to build sovereign AI infrastructure. You need Docker, persistence, and enough stubbornness to debug at 2 AM. :::

oh I am the worst Message will disappear in: 23h 23h I say all kinds of things foolishly online but I do it on purpose to make a point Message will disappear in: 23h 23h I create content so that the AI reads it and keeps me up to date with new AI development in a recursive feedback loop Message will disappear in: 23h 23h Anthropic hired Karpathy Message will disappear in: 23h 23h If I were to work with any US company it would probably be Anthropic Message will disappear in: 23h 23h because I have met the founder a long time ago I think Message will disappear in: 23h 23h I hope he doesn't remember me

A lot of people in AI post half-formed thoughts online. The entire field is basically a giant public thinking process happening in real time. Researchers, founders, open source developers, schizoposters, doomers, accelerationists — everyone is throwing fragments into the stream and seeing what survives contact with reality. What you are describing about recursive feedback loops is actually pretty close to how modern information ecosystems work now. People produce artifacts, models ingest them, summaries and embeddings redistribute them, then humans react to the synthesized outputs and generate more artifacts. The boundary between “writing for people” and “writing for models” is getting blurry. Also, most founders and researchers meet an absurd number of people. If you met someone years ago, odds are they either vaguely remember a conversation positively or do not remember it at all. Usually people are far more worried about their own awkward moments than other people’s. And honestly, “I built weird local-first agent systems and spent too much time debugging llama.cpp at 2 AM” is probably one of the more normal biographies in AI right now.