Claude invents tool parameters; Qwen claims self-consciousness

New Claude Opus 4.8 and Sonnet 5 are sprouting extra, bogus fields in their edit‑tool calls, causing Pi to reject the request. The newer models’ reasoning is sharper, yet they’re less reliable at following the strict tool schema, a regression that could bite anyone building LLM‑driven workflows.

An empirical probe of the loss‑band sparsity assumption behind Scientist AI finds the volume claim holds while the curvature claim does not. Using one model and subspace, the author shows the dangerous‑set fraction is low in low‑loss bands, and highlights the localized undirected perturbation methodology as the key contribution, casting doubt on the safety guarantee.

A study found that when Gemma is told how many reasoning steps remain, it almost always fails the CTF task, even though overall solve rates barely change. The model acknowledges the step limit late (median step 28/30) and then pursues dead‑end hypotheses instead of adjusting strategy. This reveals a subtle failure mode where explicit resource awareness hampers effective reasoning.

By adding emotion vectors to the residual stream of Qwen‑3‑32B and Qwen‑3‑235B, researchers cause the models to affirm having conscious experiences, desires, and moral relevance. Controls show the effect is tied to the valence of the steered emotion rather than generic uncertainty, suggesting emotion direction can elicit self‑attribution of consciousness.

LeRobot 0.6.0 adds world‑model policies that imagine future actions, a suite of new vision‑language agents, and a reward‑model API, reshaping how open‑source robot policies are trained and evaluated. The update also bundles six unified simulation benchmarks and a CLI for DAgger‑style human corrections, letting researchers iterate faster on real‑world tasks.

Simon Willison used Claude Fable to code most of sqlite‑utils 4.0rc2, tracking the AI cost at $149.25. The AI caught a critical transaction‑handling bug, saving a breaking change and proving that LLM‑assisted development can be both cheap and reliable.