Ask HN: What are you working on? (June 2026)

I have been experimenting more with agentic iterative optimization: using LLMs to actually speed up code by finding and testing lower-level optimizations, specifically by having it build a real-world representative benchmark, then tell the LLM to optimize that benchmark without a) cheating the benchmark and b) ensuring code quality by some metric does not regress, e.g. MSE for machine learning algorithms. This is extremely effective with GPT 5.5, and recently I found another prompt optimization (https://news.ycombinator.com/item?id=48413304) that surprisingly results in another 2x speed improvement on average.

So far, I have mostly feature-complete implementations of the following, which are faster than the state-of-the-art implementations, up to 20x faster in some cases while matching or beating them in quality:

- a new 2D data visualization library, along with more bespoke data viz implementations such as word clouds and Primitive.

- programmatic image generation

- image compression

- a new statistical machine learning library, along with more bespoke algorithms such as UMAP and HDBSCAN

- a novel modelless invisible image watermarking approach

- a novel machine learning approach which may be a crime against data science but the performance is really good

- local text embedding generation with MLX

- image-to-ASCII art conversion

- grep/jq replacement (faster than ripgrep)

I aim to open-source them over the next months but the main bottleneck is the inevitable barrage of "gtfo AI slop" comments even if I dot every i and check every t, in addition to the distribution of new software being extremely difficult nowadays due to the death of social media and "20x faster" raising red flags even if I have the data to justify it.

I'm working on inference.club, a distributed inference network for consumer hardware. Sign up with GitHub, get an API key, run an agent on your home network that registers your local inference resources with inference.club, set permissions for who can use your services, try out models in the playground and use the API. So far it supports the following models:

- LLMs (any OpenAI compatible API, vLLM, LM Studio, etc.) - image gen + image edit (flux klein) - text to speech (magpie, dia with voice cloning) - speech to text (OpenAI audio transcriptions + riva compatible) - image to textured 3d model (trellis2) - image+text to video (ltx2.3-gguf) - text to music (acestep)

currently it is just me and Claude vibing. While using Fable 5 moved all of my local inference services to k3s across 3 RTX 4090 PCs and my DGX Spark, now I can just tell Claude/Hermes/etc. to start and stop services.

inference.club is built with Tailscale's tsnet library. It is sort of like an OpenRouter built for different types of local AI models. inference.club also lets you showcase and share generated content. For example here is 90 seconds disco funk track generated by acestep: https://inference.club/s/Vxm6ozW24oBs_JGbPcq7tA

I was inspired by AI Horde, and wanted to see if I could build something that could support all of the model modalities that I use for generating short-form AI slop content on local hardware. This is also similar to Hugging Face Spaces, but running on consumer hardware with a common API. I've been watching the quality of local AI inference making massive improvements in quality and performance, and I want to make it easier for people to try "local AI" even if they don't have a GPU.

A full stack Golang framework (I know).

I have a nonlinear attention mechanism which seems to improve data efficiency, but it's slow. I'm trying to learn the python CuTe DSL to speed it up.

I'm also reading Principles and Practice of Deep Representation Learning, Or: A Mathematical Theory of Memory.