The AI Briefing

Anthropic announces Claude Fable 5.1 and Claude Mythos 5.1, the day's most-voted story by a wide margin.

Newsonaut makes the case for sticking with Firefox as browser competition and default search deals come under pressure.

Google Play is rejecting AnkiDroid's Open Collective donation link, threatening one of the largest open-source Android projects' funding route.

Dan Luu checks Ed Zitron's past AI-skeptic predictions against what actually happened since he made them.

The author trains a small transformer from scratch in about 90 minutes for the ARC benchmark and reports it beating many general-purpose LLMs.

Simon Willison digs into the Codex desktop app and finds it ships an entire LibreOffice install, raising questions about desktop-app weight.

First pragmatic benchmark for Hindi, Punjabi, Tamil and Malayalam, covering deixis, speech acts, implicature, social pragmatics and coherence. Multilingual LLMs fail consistently on meanings rooted in Indic conventions, and translation quality does not reliably track pragmatic understanding. [eval, multilingual, pragmatics]

Tool Primitives swap rigid API schemas for natural-language tool interfaces, and ToolFace retrieves from a repository of 25,519 functions at inference time instead of enumerating schemas in context. The HEART planner/router/verifier harness beats SFT baselines by 10% and frontier commercial models by 6% on average while cutting API cost up to 85%. [agents, tool-use, inference]

A two-stage recipe: a 4B teacher reranker is strengthened with off-policy GRPO on LLM-judge feedback, then a 1B student is distilled on-policy over rankings it samples itself. The student reaches 0.7670 nDCG@6 on MAIR-11, beating offline listwise KD by 4.6 points, with gains concentrating under distribution shift. [distillation, reinforcement-learning, reranking]

Uses abductive reasoning, whose difficulty cannot be read off formal structure, to test whether reasoning models' thinking effort tracks human reaction times without shortcuts. Finds further alignment in reasoning cost and error patterns, and that decoding methods exploring multiple paths increase the correspondence across the three models tested. [reasoning, eval, cognitive]

Treats inter-agent messages as traces of a sender's reasoning and models 6G RAN agents as cognitive channels on a cellular sheaf, so protocol-valid reports can still carry a hallucination. Derives five design principles for resilient multi-agent systems and validates them on locally deployed 1B-parameter telecom models, where a cognitive SNR isolates a hallucinating peer that three of four neighbors agree with. [agents, multi-agent, reliability]

Decompiler Explorer runs several decompilers side by side on an uploaded binary so their output can be compared directly.

A developer ports 65,000 lines of Go to Rust with Claude Fable for roughly $400 and details where the model helped and where it did not.

A lightweight Spotify web client built to load and play fast.

Krebs on Security reports the FBI is investigating a service selling more than 153 million driver's license records.