I am building a machine learning framework in C . So far I put almost 2 years of my leisure time into it and hope to be able to show it off around end of this year. Uchen ([oochen]) is Ukrainian for "student". My goal is to help people build smarter software and have some fun.

Russia hit an Orthodox Church that was established in 1051, well before Moscow was founded even if we believe Russian falsified history. This is pure envy. Russia wants to forget that their Church originated from Kyiv. They are embarrassed to admit that their language, faith and culture all were brought from Ukraine.

Add to my CLAUDE.md: - **Trust nothing.** The whole stack — uchen-core, training binaries, rollup math, eval kernels, WASM inference — is home-grown. A clean `decision.json`, a tight CI, a converging loss curve are hypotheses until cross-checked against bit-identicality probes, sha256 pins, expected wall-time ranges, and an independent path. Surprising numbers (good or bad) are first a bug suspicion. Single-line bzl / shuffle / seed bugs silently destroy eval signal while loss curves look fine; that has happened before and the §0.1 / §0.2 rules exist because of it.

Finally finished Invincible. Started reading it around 20 years ago.

Literally the worst cable management I've ever seen in my life

Model is still pretty stupid (basically have to do brute-force walk). MCTS counter is model inferences. As model becomes smarter I will be doing less inferences, will be making its turns much quicker. Also, this is a training model that does some unnecessary computations. Shipped model will have less layers.

When I was starting UchenML I wanted to do fun stuff - C templates, multi-process, memory management. And not this all "Gumbel MCTS, multi-head loss, SGD-momentum, loss.h."

I use Claude Code all the time. I do not see difference in model outputs between different models or effort levels. I choose Opus just in case, but a few time had sessions that defaulted to Sonnet. I only knew it because I saw it somewhere in UI.

Asked ChatGPT what my model definition would look like in PyTorch and hypothetical Rust version of UchenML. C wins! My goal was explicitly not to have to include same information twice. I know what previous layer outputs - why do I need to specify it again?

My Gmail inbox now shows a large Upgrade button. Sorry Google but I am already paying for 2 AI subscriptions.

Write maintainable code.

If you need this, you are probably doing something wrong. Also, see std::visit

No need for the debugger.

One should really review the code LLMs write. They make really subtle bugs. Also MaxLanes is not the "worse case scenario".

Regardless of what I say about AI, somebody always considers my post overly negative or overly positive. It's a polarizing topic for sure. If you are in the middle, everybody hates you on both sides. I have a pragmatic view about AI. It's a good tool, but it needs a lot of micro-management to be viable, and even then it often frustrates me. Still I use it because it improves my overall productivity. But the quality is slightly worse, even with constant micro-management.

I once wrote a project without loops. Map/filter/reduce only. In Python.

Rendering is usually not slow. What is slow is a huge number of network requests. Many web sites are composed of disjointed pieces (e.g. advertisements) that access data from different servers. E.g. send a request for the list items. Then for each list item fetch thumbnail. Scroll may reveal more items.

Claude: "Land the gradcheck fix. After ~6 attempts at analytical debugging I could not localize the bug; the math is structurally correct per the statistician, the ReLU mask placement is correct, the scratch is 64-byte aligned, the parameter spans are disjoint. Both reviewers and I agree: it's likely a subtle SIMD interaction the debugger skill or a print-instrumented A/B against a scalar reference can crack quickly — but I couldn't.What I did not do:" Is it for realz? Expecting me to debug its code?

Claude started doing this too. I stopped using Codex because of this junk.