|
đź“– Read In Depth
|
The Claude Code Source Leak: fake tools, frustration regexes, undercover mode
Claude Code's source code was accidentally exposed via a sourcemap in their NPM package (a Bun bug). The leak reveals fake tool injection as an anti-distillation defense, regex-based frustration/sentiment logging, an 'undercover mode' that suppresses all AI attribution in commits and PRs, and unreleased feature flags including a 'kairos' assistant mode. This is a rare inside look at how Anthropic actually builds production AI tooling — the engineering decisions, product roadmap signals, and some uncomfortable choices about transparency.
hn/Best Stories
|
Artemis II is not safe to fly
Maciej Cegłowski (Idle Words) argues Artemis II has unresolved safety issues reminiscent of the Challenger and Columbia decision-making failures — specifically, known heat shield damage being rationalized away rather than resolved. A Harvard lecturer noted they updated their systemic failure case study slides the night before a class to include this. High-stakes, well-sourced critique of NASA's safety culture from a writer known for rigorous technical analysis.
hn/Best Stories
|
Show HN: 1-Bit Bonsai, the First Commercially Viable 1-Bit LLMs
PrismML claims to have built the first commercially viable 1-bit LLMs — models where weights are ternary (-1, 0, 1), enabling inference at a fraction of the memory and compute cost of standard models. If the performance claims hold up, this is a significant efficiency breakthrough with real implications for edge deployment and cost structure. Worth evaluating the technical substance carefully.
hn/Best Stories
|
Axios compromised on NPM – Malicious versions drop remote access trojan
Axios, one of the most widely-used JS HTTP libraries, was compromised on NPM — malicious versions were published that drop a remote access trojan. With hundreds of millions of weekly downloads, the blast radius is significant. Details on how the compromise happened and detection matter for anyone running JS-dependent infrastructure.
hn/Best Stories
|
States Plow Ahead With A.I. Regulation, Defying Trump
California, Utah, and other states are pushing ahead with AI regulation despite Trump's executive order preempting state-level AI rules. As an ML engineer in Santa Clara, this directly affects what your employer can build and how. The tension between federal deregulatory posture and California's historically aggressive tech regulation sets up a legal collision worth understanding.
nyt/Technology
|
Fedware: Government apps that spy harder than the apps they ban
The official White House app ships with Huawei Mobile Services Core as an embedded tracker — the same Huawei the US government sanctioned for being a surveillance risk. It also includes an ICE tip line and auto-fills 'Greatest President Ever!' before collecting your name and phone number. The piece examines government apps broadly and how they combine the surveillance practices of the apps they ban.
hn/Best Stories
|
Why the US Navy won't blast the Iranians and 'open' Strait of Hormuz
A substantive analysis of why the US Navy can't simply 'blast open' the Strait of Hormuz — the military, legal, and geopolitical constraints that make this much harder than it sounds. With Iran formally moving to charge tolls and Trump telling Europe they're on their own, understanding the actual strategic calculus here matters for making sense of market volatility and energy policy shifts.
hn/Best Stories
|
Jürgen Habermas (1929–2026): The philosopher who chose the state
Jürgen Habermas died at 96. This piece contextualizes his life's work — communicative rationality, discourse ethics, the public sphere — within the postwar German political environment that shaped him. His framework for how democratic legitimacy requires undistorted communication feels newly relevant in an era of AI-generated propaganda and information warfare.
reddit/r/philosophy
|
|
⚡ FYI
|
Ollama is now powered by MLX on Apple Silicon in preview
Ollama now uses Apple's MLX framework as its inference backend on Apple Silicon (in preview). This means significantly faster local model inference on M-series Macs by leveraging the unified memory architecture properly. Practical upgrade if you run local models on a MacBook or Mac Studio.
hn/Best Stories
|
OpenAI “internal model” solved 3 more Erdős problems
OpenAI's internal (unreleased) model has reportedly solved 3 more Erdős combinatorics problems, with a paper on arXiv. This follows earlier math competition results and suggests frontier models are making genuine progress on hard open mathematical problems, not just competition benchmarks. Worth tracking the paper for methodological substance.
reddit/r/singularity
|
Oracle slashes 30k jobs
Oracle is laying off ~30,000 employees — roughly 18% of its ~162K workforce. The HN discussion is illuminating: commenters largely agree Oracle's value prop has evaporated (its moat was lock-in, not technical superiority), and the cuts read as a belated recognition that their AI investment thesis isn't generating returns. Structural decline story, not a cyclical one.
hn/Best Stories
|
Google's 200M-parameter time-series foundation model with 16k context
Google Research released TimesFM, a 200M-parameter foundation model for time-series forecasting with 16k context length. Trained on a large corpus of time-series data, it's designed for zero-shot forecasting across domains. Relevant if you work on any forecasting or sequential data problems — foundation model paradigm applied to a domain ML hasn't fully cracked.
hn/Best Stories
|
GitHub backs down, kills Copilot pull-request ads after backlash
GitHub inserted ads into 1.5M pull requests via Copilot, rebranded as 'product tips,' and backed down after developer backlash within days. The episode is a clean case study in platform incentive drift — GitHub is now a monetization surface for Microsoft, and the developer trust that made it valuable is the thing being spent.
hn/Best Stories
|
Demis Hassabis secretly built a hedge fund inside DeepMind trying to beat Jim Simons. Google shut it down.
Demis Hassabis apparently built a secret hedge fund inside DeepMind, attempting to apply AlphaGo-style RL to financial markets to beat Jim Simons' returns — Google shut it down. Interesting window into the incentive misalignments at AI labs and what Hassabis was actually trying to build before the current LLM era.
reddit/r/singularity
|
New Washington state law bans noncompete agreements
Washington state has banned noncompete agreements entirely — joining California (where you already live and work) in making them unenforceable. This accelerates talent mobility in tech hubs and has downstream effects on how startups and large companies structure employment, especially for engineers with access to proprietary ML systems.
hn/Best Stories
|