Show HN: I successfully failed at one-shot-ing a video codec like h.264

The experiment

A developer published libsinter, an experimental video codec built from scratch as a learning exercise. It has been reported that the work was produced by "one-shot" Claude Code agent teams that simulated an expert group to design and implement the codec. The goal wasn't to ship a production competitor to H.264 — it was to see how far a simulated expert team could push a novel, patent-free architecture in a single pass. Spoiler: impressive engineering, humble results.

The results

At comparable luma quality (~49 dB) libsinter ended up producing files about 18.6× larger than an H.264 baseline — for example, 136 KB vs. 33 KB at QP 4 on a 256×256, 30-frame testsrc. It has been reported that the repo documents PSNR and BD-rate data and includes ~5,000 lines of C, lapped transforms, hybrid PVQ/scalar quantization, dual-interleaved rANS, P-frame inter prediction, and a small CLI (sinterenc/sinterdec). The code is patent-free by design and comes with 32 passing tests, but perceptual quality wins did not translate to compression parity.

Why the gap exists

The team concluded the gap is multiplicative — missing sub-pixel motion compensation, no B-frames, PVQ verbosity, and far fewer entropy contexts each cost a factor. Realistic, patent-safe tuning might close some of that to a 4–6× disadvantage, but matching H.264 would require re-implementing many of its contested—but extremely effective—tools. In short: you can invent interesting alternatives, but the incumbents have years of tricks under the hood.

So what now?

This is a classic proud failure. Did it "fail"? Depends on the metric. As an experiment about agent-led workflows and pedagogical architecture exploration, it succeeded. As a drop-in H.264 challenger, not even close. The repo and write-up are on GitHub for anyone who wants to poke at lapped transforms, PVQ, and rANS in a compact C codebase — useful reading for codec researchers, AI tooling skeptics, and anyone curious whether AI can bootstrap deep-domain engineering in one shot.

Sources: github.com/dheerg, Hacker News