← Index
Audio Briefing
Your browser does not support the audio element.
Headlines
01
Stealing Reasoning Traces from Proprietary LLM APIs
Research documents extracting hidden reasoning traces from proprietary LLM APIs — model internals leak through logit and token behavior.
02
uBlock Origin is giving up the fight to keep ads off Facebook
The ad-blocker concedes the Facebook fight — a reminder that platform-controlled surfaces are increasingly immune to client-side blocking.
03
OpenAI's head of ethics leaves less than a year after joining
OpenAI's ethics chief departs under a year in — the latest sign of turbulence in AI governance roles as labs scale.
04
Compression is prediction
ngrok's deep dive connecting compression theory to prediction — why LLMs, log loss and zip files share the same math.
AI / LLM / Agents
01
Go is an ideal language for AI-assisted software engineering
Google makes the case for Go in the agentic-coding era — static typing, simple concurrency and readability as LLM-friendly properties.
02
Mojo 1.0
Modular ships Mojo 1.0 with a new SDK — the Python-superset language for AI compute reaches a stability milestone.
03
Nvidia Nemotron 3.5 Lightning and NeMo Switchyard
Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard — open reasoning models and orchestration for the RTX/DGX ecosystem.
04
What sort of maths are LLMs good at?
Tim Gowers evaluates LLM mathematical capabilities — a mathematician's careful taxonomy of where models genuinely reason vs. retrieve.
Papers — ArXiv CS.AI
01
SteerBench-Work: A Benchmark for Agent Steering at Action Boundaries
The pre-commit steering decision — proceed or hold for human review before an agent sends an email, merges a PR or wires a payment. [agents · eval]
02
Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence
LLM judges are validated on accuracy, but accuracy says little about stability under re-prompting, challenge or sustained pressure. [eval · reliability]
03
EgoCITE: Context-Augmented Indexing and Time-Aware Retrieval for Egocentric Memory
Fixes two bottlenecks in long-horizon egocentric memory: context-poor caption indices and time-blind retrieval for agentic search. [memory · agents]
04
General Probabilities of Causation with Causal Knowledge
Extends Tian and Pearl's bounds on probabilities of causation using available causal knowledge — sharper partial identification for individual responses. [research · theory]
Infra / SRE / DevOps
01
llama.cpp
The llama.cpp project's new site lands with llama.app — local inference keeps maturing as a mainstream distribution channel.
02
Apple Silicon and macOS VMs: faster LLM inference with llama.cpp
GPU passthrough for macOS VMs enables faster llama.cpp inference on Apple Silicon — virtualization and local AI finally meet.
03
CFTC declares market emergency, orders Kalshi to continue operations
The CFTC steps into prediction-market turmoil — an infrastructure-level signal that AI-driven event markets now move regulatory gears.
Hacker News
01
uBlock Origin is giving up the fight to keep ads off Facebook
717 pts — platform-controlled surfaces beat client-side blocking.
02
Stealing Reasoning Traces from Proprietary LLM APIs
695 pts — extracting hidden chain-of-thought from closed models.
03
Compression is prediction
667 pts — the information-theory bridge to LLMs.
Why It Matters
▸
Stealing reasoning traces + uBlock giving up on Facebook in one day: proprietary surfaces — model internals and platform feeds — are both closing, and the open alternatives are the ones left to build on.
▸
Go as the LLM-friendly language and Mojo 1.0 shipping the same day frame the real contest of the agentic era: not model vs. model, but which programming languages and toolchains agents can reliably own.