One weekend, one model: Opus 5.5 helped build QLink and run all in-LLM training experiments in a day
zeeshanp_ · x · 2026-11-03
zeeshanp wrapped up the weekend project: the system is ridiculously slow for training any real model — one XOR step takes about 10k decoded tokens — but training neural networks inside other neural networks, and even "scaling up" in a distributed fashion, is still remarkable.
He was surprised yet again at how effective Opus 5.5 is at exploratory work: he pointed it at the QwenGPU repo, discussed the project's shape, and it helped build QLink and run all the experiments in a single day.
More from coding & agent
- Agent matched 3 contacts wrong in one blast: identity maps beat name joins — petrusenko_max · 2026-11-03
- wxa-runner runs WeChat mini programs in the browser with DevTools debugging and Playwright tests — lxfater · 2026-11-03
- DeepSeek pushes Creator Mode: its Harness agent builds plugins from plain-language prompts — rohanpaul_ai · 2026-11-03
- Claude Code leaks a reasoning trace, then blocks the bug report as reasoning_extraction — rchardkovacs · 2026-11-03
- Ant Group's SWE-Trail: history-guided trajectory collection yields better coding agents — antgroup · 2026-11-03
- Open-source AI Xianyu monitor with Playwright + LLMs hits 16k GitHub stars — Usagi-org · 2026-11-03