The AI Briefing

Meta Research publishes Muse Spark Open — the day's top story on Hacker News at 747 points, with the discussion still running hot.

Goodfire, the interpretability company, publishes an interactive piece arguing models are far more legible than the field assumes — no rewiring required. 615 points.

Citizen Lab details how Discover Tech converted targeted advertising into a delivery channel for mercenary spyware — no click required from the victim. 590 comments; the ad-tech-to-spyware supply chain is the story.

From 30 October, Meta's commercial terms will allow scraping of websites behind business accounts — legal experts read that as covering news sites, no opt-out in sight.

Krea releases Krea 3, positioning editing and iteration — not one-shot generation — as the primary workload.

Christian Schaller on the GNOME blog walks through applying models to reverse engineer undocumented hardware — with obvious consequences for driver and firmware work.

The Simons Foundation launches a $15M programme to help mathematicians work with — and against — frontier models.

Derek Thompson's essay on what frontier models mean for pure mathematics — the emotional counterpart to the Simons grant. 588 comments.

A follow-up digging further into the resemblance between Anthropic's Scion output and Princess Mononoke — and what was actually lifted.

Agent harnesses call a large LLM at every step, but most of those calls are just multiple-choice decisions — which tool next. This paper builds AgentDecisionBench to test delegating those decisions to cheap decision models, and finds the best isolated-question model still solved only 69.5% of full agent tasks; without a fitted confidence threshold, delegation cut success by 19.0 points. The gap between answering questions and running an agent is now measured, not assumed. [agent harnesses, decision models, benchmarking]

Coding-agent benchmarks that fix one accelerator or harness cannot tell you whether a tool or skill actually pays for itself. HPCAgent-Bench specifies 689 kernels once in NumPy, derives C and Fortran references, and runs submissions through a common ABI on CPUs or GPUs — with an external judge checking correctness on hidden fuzzed inputs to close reward-hacking paths. Paired-run metrics isolate each intervention's contribution. [code optimisation, benchmarking, hpc]

Repository-level benchmarks treat one issue or PR as the task, but real software work unfolds across coordinated multi-PR changes. WEAVE converts verified PR histories into executable tasks organised as bounded DAGs, and introduces Graph Credit — a deterministic, test-grounded metric for measuring contribution across dependent changes. That gives long-horizon code agents both a realistic evaluation and a usable learning signal. [code agents, benchmarking, long-horizon tasks]

Self-evolving agent memory systems adapt from coarse feedback — evaluation scores, compressed summaries — which cannot localise failure, so updates end up rewriting architectures that were mostly fine. MemScope grounds evolution in evidence, using tool-use traces to pinpoint the failing component before any code change. Localised repair instead of wholesale rewrite. [agent memory, self-evolution, failure localisation]

An agent that answers correctly on clean inputs can still fail at the actual job: escalating when evidence is missing, stale, or conflicting. Across 400 perturbation-controlled healthcare workflow episodes, five open-weight models hit 0.97 clean accuracy yet missed 32–97% of required escalations, and the strongest 9.7B configuration dropped to 0.78 under conflict. Verbalised confidence did not close the gap. [safety, evaluation, healthcare agents]

As agent registries grow, dumping every Agent Card into the orchestrator's context inflates cost and latency until the window breaks. SSR is a deterministic, training-free retriever that weights lexical signals from capability-bearing card fields with IDF and scores candidates through an inverted index — keeping the candidate set compact without a learned model in the loop. [multi-agent systems, agent discovery, retrieval]

Trajectory-level rewards hide which operations in an agentic knowledge-graph QA pipeline actually helped. OpCredit scores individual operations against downstream outcomes — useful if you're doing RL over graph tool calls.

A natural-language-to-SPARQL framework built around the vocabulary problem: questions word concepts differently from the target schema, and limited context demands selective schema access.

Class hierarchies carry subtle subsumption errors; the authors use embeddings plus LLM review to refine the CASTON casting-domain ontology — a lightweight, scalable repair loop.

The General Conference resolves to end leap-second insertions — closing out a 54-year patch on Earth's irregular rotation and, eventually, a whole class of outage postmortems.

The libre BIOS/UEFI replacement adds UEFI boot and Chromebook targets in the 'Ominous Olive' release.

One engineer's full self-hosting inventory — 396 points of config envy and stack comparisons.

LWN compares how the two organisations actually develop Chromium — commit flow, review latency, and who ships what upstream.

Seven decades on, the compilers still run the world's weather models and HPC benchmarks.

A long-documented voltage-glitching campaign against the Xbox 360 — a masterclass in hardware persistence.

From fetching a firmware image to root on the EAP770 access point — the full exploit chain written up.

Zimmermann's classic essay — the prosecution threat, cryptography as protected speech — still worth ten minutes in 2026.