Top Stories

Stealing Reasoning Traces from Proprietary LLM APIs

629 points · stolen-thoughts.com

A new writeup demonstrates that the “hidden” chain-of-thought reasoning that labs like OpenAI and Anthropic deliberately keep behind their APIs can be partially reconstructed by an attacker with only ordinary API access. By exploiting timing signals, token-level metadata, and summary leakage, the researchers reassemble the internal reasoning traces the providers charge for but never intended to expose.

HN is fixated on this because reasoning traces are both a competitive moat and a safety artifact — labs hide them precisely so competitors can’t distill them into cheaper models. If they leak this easily, the whole “closed reasoning” business model is on shakier ground, and the distillation arms race gets a lot messier.

Compression Is Prediction

548 points · ngrok.com

An ngrok engineering post revisits the deep equivalence between compression and prediction — the idea that a model good at predicting the next token is, formally, a good compressor of the data it was trained on. The piece walks through the intuition with concrete examples rather than hand-waving at Shannon.

The thread is classic HN catnip: half the comments are people who find the compression-equals-intelligence framing genuinely clarifying for why LLMs work, and the other half are pushing back on how far the analogy actually stretches. Either way it’s the kind of first-principles explainer the community loves to argue over.

OpenAI’s Head of Ethics Leaves Less Than a Year After Joining

451 points · ft.com

The Financial Times reports another high-profile safety and ethics departure at OpenAI, this time the head of ethics exiting under a year into the role. It’s the latest in a long string of governance-side exits that have followed the company through its various restructurings.

HN’s read is predictably cynical: every time a safety or ethics leader walks, commenters treat it as a signal about where real decision-making power sits inside the company. The debate is less about this one person and more about whether OpenAI’s ethics function has any teeth at all.

Mojo 1.0

392 points · modular.com

Modular has shipped Mojo 1.0, marking the first stable release of Chris Lattner’s Python-superset language aimed at fusing Python’s ergonomics with systems-level performance. The 1.0 milestone signals API stability and a commitment developers can finally build on without expecting the language to shift under them.

For HN this is a long-awaited moment — Mojo has been “promising but pre-1.0” for years, and the crowd is split between excitement about a genuinely fast Python-like language for ML kernels and skepticism about lock-in to Modular’s ecosystem and how open the language really is.

Go Is an Ideal Language for AI-Assisted Software Engineering

386 points · developers.googleblog.com

Google’s developer blog argues that Go’s simplicity, small feature surface, fast compilation, and strong standard library make it especially well-suited to a world where AI agents write and iterate on code. The claim is essentially that a language with fewer ways to do things gives an LLM fewer ways to go wrong.

The comments are a lively language war: Go fans nodding along, Rust and functional-programming folks arguing that “simple” often just means “verbose,” and a contingent noting the piece is conveniently self-serving coming from Go’s own steward.

London Underground Begins Scanning Passengers’ Faces

351 points · btp.police.uk

British Transport Police are expanding a live facial recognition trial into London Underground stations, using real-time camera matching against watchlists as passengers move through the network. The rollout extends surveillance tech that’s already been controversial above ground into the tube.

Privacy is the dominant thread on HN, with commenters debating false-positive rates, function creep, and how quickly “trial” becomes permanent infrastructure. Several point to it as a preview of where transit surveillance is heading in other major cities.

Grok Bot

299 points · x.ai

xAI has launched Grok Bot, a new agentic product from Elon Musk’s AI outfit, drawing the usual outsized attention that anything Grok-branded gets on HN. The launch page pitches an always-available assistant plugged into the X ecosystem.

Discussion splits along familiar lines — genuine curiosity about the technical capabilities versus deep skepticism about moderation, data sourcing, and the broader X integration. As always with xAI, the product and the politics are hard to disentangle in the comments.

llama.cpp

249 points · llama.app

The project behind the wildly popular llama.cpp inference engine gets a fresh spotlight, underscoring how central Georgi Gerganov’s C/C++ runtime has become for running open-weight models locally on consumer hardware. It remains the backbone of a huge slice of the local-LLM ecosystem.

HN loves llama.cpp as the poster child for lean, dependency-light engineering — the ability to run capable models on a laptop with no cloud dependency is exactly the kind of thing the community rallies behind, and the thread is full of people sharing their local setups.

Nvidia Nemotron 3.5 Lightning and NeMo Switchyard

234 points · blogs.nvidia.com

Nvidia announced Nemotron 3.5 Lightning, a new efficiency-focused model family, alongside NeMo Switchyard tooling aimed at routing and serving models across RTX and DGX hardware. The pitch is faster, cheaper inference tightly optimized for Nvidia’s own stack.

The HN angle is partly technical interest in the Lightning models’ latency claims and partly the recurring debate about Nvidia’s vertical integration — commenters note how effectively the company keeps pulling the whole AI stack toward its silicon.