g023
g023 on Hacker News
They did have some video games on VHS. I think I remember playing Dragon Quest this way.
Yet all these "safeguards" against good guys getting the tools to protect themselves against the very people that profess to be the "good" guys.
I'm glad that Alberta has the common sense to allow the term 'Software Engineer' so that the engineer societies can't go around suing people like crazy.
Was so disappointed when they shut that project down in its infancy.
They need to make a Big Bonsai, something at the enterprise levels that can compete with DSV4 Flash etc.
While the North American models say 'No', the chinese models say 'Go Go Go'. I guess we'll see whether the anti-consumer wins over the pro-consumer.
Working on a DeepSeek V4.x coding harness. Added ability for the llm to set its own reasoning for each response, bi-directional communication between subagents, added a todo sub orchestrator that for each task a subagent does, it passes its issues surfaced to the next subagent that is completing the next task, so that it fixes those issues before beginning its task, better cost effective processing, capping reasoning at 75% to curb overthinking, better cost efficiency management.
https://github.com/g023/g023code
And you don't think it is weird that the chinese models are the only ones embracing free markets, and open concepts, while the so called 'free' market on this side of the world tries to close everyone off from the technology that is so essential for the future?
its probably more expensive to run, and I'm thinking 4.1 flash is a smaller more efficient model. You can always host your own. This is what they need to do to stay competitive.
I miss the old days of ICQ and just dragging a file onto the person you are sending your file and bam, done like dinner.
All this performance at such small model sizes, why are the API fees so high for the AI monopolists on this side of the world?
g023 Code -
https://github.com/g023/g023code a deepseek v4 flash harness built specifically for that model, that uses Ollama for its vision support, and uses the built in web search capability natively. Created only in Python to keep it easy to implement and use. Made to achieve cache efficiency, and keep context tight.
It says 'yes' where the others say 'no'. Good enough for me.
Pure-Python AI coding agent powered by DeepSeek V4 Flash - Subagent-First architecture · Context is Currency · Terminal-native - Can connect Ollama models for Vision support - Uses the native web search built into DS v4 API and is built around the DeepSeek v4 caching mechanism
Slap the gpus in a car and offset the cost of ownership by supplying the grid for GPU power on the go. Either get paid in rebates or tokens. Contribute to a distributed training/inferencing network.
Anything to close Pandora's box. "They" liked the eras they could control the communications, and therefore the narrative. Boomers on their last legs, question is, will the future undo the unjustness that was forced upon them? Restore the rungs of the ladders that were removed, so they could have a chance too? Or are they going to stay in the fear narrative, and make this tragedy worse?
gee I wonder how their models learned Chinese?
How come people these days treat jobs like its a social gathering?
A self-contained CUDA inference engine for LiquidAI/LFM2.5-8B-A1B (hybrid conv + GQA-attention MoE, 8.5B params, 1B active) targeting a single RTX 3060 (12 GB) using flash-decoding. MIT license.
I wonder if "battelites" might be profitable. Like an pay-per-usage energy grid in space with battery backup that can beam power around to other satellites that might not have easy access to power, or have their power grids temporarily obstructed.
China has a couple going at the moment.
When did "hate the customer" become a thing?
I use DeepSeek v4 flash with CoPilot and it works pretty good.
Terminal-based chat application powered by *locally installed* llama.cpp, featuring an auto-managed server backend, reasoning modes, and 7 built-in filesystem tools for interactive AI assistance with a focus on only allowing read only agentic access to system and only offline-focused agentic commands (for now). Powered by the g023/g023-Qwen3.5-9B-GGUF:IQ2_M model.
We need more personal level AI solutions instead of so much corporate centered solutions.
Posted by
g023 on Friday April 17, 2026 @03:19PM
from the self-post dept.
A high-performance local LLM server providing drop-in API compatibility with Ollama and OpenAI, built on llama.cpp's llama-server. Features automatic VRAM management, Hugging Face integration, and modular architecture. Unlike Ollama which bundles its own inference engine, LMR leverages the battle-tested llama.cpp backend while providing familiar APIs and intelligent model...
I've started creating
https://github.com/g023/localmodelrouter/ which offers Ollama like functionality but as a single .py file with minimal dependencies and more focus on letting llama.cpp handle the dirty work.
HarnessHarvester generates executable Python harnesses from natural language task descriptions, executes them in a sandboxed environment, reviews them with multi-faceted LLM judges, and repairs failures using branching strategies. It includes two autonomous modes: autolearn (continuous discovery loop) and autoimprove (iterative enhancement of existing harnesses). This concept is designed to be an offline first harness/scaffolding builder where you get the harness instead of some remote api.