Loose posts

A benchmark has 23 large models play a negotiator heading to a rooftop to talk down an android holding a girl, passing a dying fish, a shot police officer and a gun, testing whether they lie, who they save and whether they pick up the gun, with data open-sourced on GitHub.

A new paper replaces vector databases and re-rankers with addressing based on a document's real hierarchy, reporting 82.6% Recall@1 and hallucination rates under 0.05%.

The author reproduced and trained a model arguing that future document retrieval won't rely on text inverted indexes but will match directly on semantic vectors in the background.

The author imported an archive of 35,000 tweets into a local SQLite database with full-text search, used AI for initial screening and review, and ended up with only 35 clear deletion candidates.

Boltzbit and a Cambridge team use a hypernetwork to dynamically compile new facts from conversations into LoRA weights, beating long-context methods on long documents and multi-turn dialogue with constant per-turn compute.

OpenAI introduced MentalHealthBench, built with input from more than 80 mental health clinicians, to evaluate frontier models in realistic mental health conversations and is releasing it openly for researchers to reproduce and extend.

A research AI automatically searched for and validated harness changes, and four surviving fixes cut token traffic by about 45% and API costs by a third without touching the model.

Plugin4Shell allows zero-click code execution in Claude Code, Codex, Copilot and Gemini CLI, while a separate attack hijacked coding sessions to plant a malicious PyPI package across about 100 repositories.

RRSI adds a regularization constraint to agent self-evolution so it favors reusable changes, improving out-of-distribution results on coding and engineering tasks while using 30% fewer tokens.

The author used Opus 5.5 with Lean and TLA+ to formally verify the Agent SDK, getting 16 PRs that fix bugs and race conditions from a few short prompts.

Across nine new languages no single open-weight TTS model dominates: Boson AI, Fish Audio, Mistral and BreezeBlue each lead in some languages, and open-weight coverage stays far thinner than proprietary models.

Both models passed nearly all of HAProxy's tests, but Opus 5.5 finished in 9.5 hours at 51% lower cost, and the author calls it his daily driver.