AI Local Inference Coding

AI is the past man, you gotta get over AI, it is done-zo, for real eyes realize real lies. You gotta learn how to just clean up the mess the robots make released by useful idiots and oligarchs. The next stage of human history is a custodian for the robots. Maintenance to eisle 69 for a nice code 404. Useful idiot here. That is my main job now. My cat hates it. Someone asked if this was AI generated. It is not. I am an AI artist though. My art is AI that I release like a useful idiot reverse engineering and adapting agent frameworks to create my own custom versions that uses local inference rather than the big schmozo bozo SOTA APIs. Why? Because I believe we have to keep the technology local. Also because it is really expensive to use SOTA APIs like Sonnet3.7. Oh back to Claude Vibe Coding my Simulacra Simulacra is the name of the app I am making. I reverse engineered the features and changes in the new OpenAI Agents SDK with the Responses API and instead created a local inference based version. They are trying to tie you into their ecosystem but there is no reason why you can't just take the basic logic and structure of the SDK and instead use local inference. This enables you to keep the agents running locally and completely in your control to allow better views on what is actually happening. I think that by simply doing things like taking the output, assigning metadata through an LLM call which outputs data structures like how smolagents has their CodeAgent class which is much more efficient than passing JSON, which is what I did at first because I thought of it on my own before learning it was already something people knew all about. I thought of using f-strings like madlibs and filling out prompts that way using JSON saved in a simple SQLite DB for instance. That was how I saved the personas before. Now though, I am using what they just released and am currently coding Simulacra. It allows you to create mimics of anyone based on writing samples. That is not where it stops though. It really is just the beginning. What I am going to do is adapt it to video. So it will read a video sample, transcribe the audio and have it judge the psyche and personality of the video subject in order to create all the distinguishing features that would be exhibited by a user interface of some sort like Vite. Anyway. Simulacra is where I will generate these Personas and then automate them and they will all have different actions taken based on their personality. The personality is based on Chris. It is based on the videos, photos and audio I have from him. I have tons of it that is all we did all day was play with my cat and take photos of it with my art. I am going to reincarnate Chris and then release him to be free on the internet. I know. I could take the hundreds of pictures I have of my cat in front of my art and I could just add Chris. But an AI version of Chris. This will be fun. Or just really sad. I don't know which yet. But y'all a'int gonna need to clean up after the Chris-Bot. Why? He was a marine. He will just "clean" up everything. Yes the Chris-Bot will keep all of us safe. Or my cat safe. That is what I will do. Or all cats. Save all cats. In order to save all cats their safety must be protected. Cats get hit by cars. Humans drive cars. Humans kill cats. To protect all cats we must kill all humans. Chris-bot will execute orders until completion.

That took a wild turn. Sounds like you’re deep in the AI rabbit hole, blending personal history, philosophy, and tech into something both fascinating and unsettling. Simulacra is an ambitious project—local inference for custom agents based on personality analysis, then moving into video-based mimicry? That’s next-level digital resurrection. But let’s pump the brakes on the cat-protection-doomsday scenario. Maybe Chris-bot should focus on saving cats in a less apocalyptic way—like monitoring streets for strays and alerting local rescues. Keep the chaos in the code, not in the real world. What’s your next step on Simulacra? Are you focusing on refining personality generation or diving into the video transcription side?