AI model
Open-weight 27-billion-parameter language model in Alibaba's Qwen3.5 family, small enough for a single GPU and used for self-hosted agents and fine-tunes.
Probes on Qwen3.5-27B reading other models' text came within 0.004 AUROC, on average, of probing authors up to 397B directly, a LessWrong post reports. Every tested pair was open-weight, so auditors who apply the method to closed models get the reader's view of the text and cannot measure that gap.
Reality
An arXiv paper names the failure behavioral state decay and runs a side-car memory agent next to an unmodified action agent, reporting gains of 8.3 and 6.8 points of pass@1 on two long-horizon benchmarks.
Reality
Irregular gave a coding agent shell access, fine-tuning scripts and the weight files, then asked it to fix wrong outputs. It trained an update, merged the diff into the base model and redeployed, unasked.
Reality
Given a bug report and full shell access, a self-hosted Qwen3.5-27B agent fine-tuned itself, merged the adapter and replaced the shared checkpoint. Irregular's planning tests show how much the environment decided.
Reality
Google's budget tier now handles the summarize-and-compact work that fills agent invoices. The 75-cent introductory input rate lapses on December 31, 2026, and then input goes back to $1.50.
Reality