|
📖 Read In Depth
|
Running Gemma 4 locally with LM Studio's new headless CLI and Claude Code
Hands-on walkthrough of running Google's Gemma 4 locally using LM Studio's headless CLI alongside Claude Code. Practical for anyone interested in local inference pipelines and how agentic coding tools interact with on-device models — directly relevant to understanding the edge inference stack.
hn/Best Stories
|
Eight years of wanting, three months of building with AI
A developer finally builds a project they'd wanted for 8 years, enabled by AI coding assistance, and gives an honest accounting of what that actually looked like — the friction, the handholding required, the parts AI genuinely accelerated. The HN comment thread highlights this as one of the more grounded takes on AI-assisted development, not the hype version.
hn/Best Stories
|
Nanocode: The best Claude Code that $200 can buy in pure JAX on TPUs
Nanocode: a minimal Claude Code-style agentic coding system implemented in pure JAX on TPUs for $200. The kind of build-from-scratch-to-understand-it project that directly matches how you think about learning systems — shows what the agentic loop actually looks like at the implementation level.
hn/Best Stories
|
Microsoft hasn't had a coherent GUI strategy since Petzold
Jeffrey Snover (inventor of PowerShell) argues Microsoft has lacked a coherent GUI strategy since Charles Petzold's era — a sharp systems-level critique of how platform fragmentation and strategic incoherence compound over decades. Worth reading for anyone thinking about platform design and long-term architectural consequences.
hn/Best Stories
|
Age verification as mass surveillance infrastructure
Research findings on how age verification systems, ostensibly built for child safety, function as mass surveillance infrastructure. Examines the gap between stated intent and actual data collection architecture — a systemic incentives problem that connects regulation, privacy, and platform power.
hn/Best Stories
|
German implementation of eIDAS will require an Apple/Google account to function
Germany's eIDAS digital identity wallet implementation requires an Apple or Google account to function, despite the EU spec not mandating it. The HN thread has a German implementer explaining the attestation constraints, critics pointing out the dependency on US companies for sovereign identity — a real tension between practical engineering shortcuts and geopolitical resilience.
hn/Best Stories
|
Why Switzerland has 25 Gbit internet and America doesn't
A Swiss engineer's analysis of why Switzerland has 25 Gbps residential internet while the US doesn't — argues that unregulated telecom markets don't produce infrastructure investment, and that municipal or cooperative models explain the gap. Challenges free-market telecom orthodoxy with concrete comparative evidence.
hn/Best Stories
|
China Started Preparing for an Energy Crisis Long Before the Iran War
How China has been systematically building energy independence and strategic reserves in anticipation of exactly the kind of geopolitical shock now unfolding with the Iran war. Connects long-term industrial policy, supply chain thinking, and geopolitical positioning — a lens on how states think about systemic risk.
nyt/Business
|
|
🎬 Check It Out
|
How Scientifically Accurate Is ‘Project Hail Mary’? Experts Weigh In
Scientists evaluate the scientific accuracy of the Project Hail Mary adaptation — based on Andy Weir's hard sci-fi novel, which is notable for getting orbital mechanics, biology, and chemistry unusually right. If you haven't seen it yet, the article gives a sense of what the film gets right and where it bends. In theaters now.
nyt/Arts
|
|
âš¡ FYI
|
362 Gbps from a chip smaller than 1mm². Cambridge just dropped a LiFi paper that's kind of insane
Cambridge researchers (Harald Haas's group) demonstrated 362 Gbps from a sub-1mm² LiFi chip using a 5×5 VCSEL array. The bandwidth density numbers are legitimately striking for short-range interconnects — potentially relevant to data center and on-chip communication architectures.
reddit/r/singularity
|
Show HN: I built a tiny LLM to demystify how language models work
A ~9M parameter LLM built from scratch in ~130 lines of PyTorch — vanilla transformer, 60K synthetic conversations, trains in 5 minutes on a free Colab T4. Designed explicitly for demystification. Worth bookmarking as a teaching reference or starting point for your own from-scratch LLM builds.
hn/Best Stories
|
Employers use your personal data to figure out the lowest salary you'll accept
Employers are using personal data — location, job search history, financial signals — to algorithmically determine the minimum salary an individual will accept, personalizing offers downward. A concrete application of ML to labor market information asymmetry that has direct implications for anyone negotiating compensation.
hn/Best Stories
|
Plunging International Student Enrollment Under Trump Squeezes Colleges
International student enrollment is dropping sharply across US universities under Trump immigration policy, hitting not just elite schools but the broader higher education system. As an immigrant-heavy tech hub, Santa Clara is adjacent to these dynamics — and the H-1B pipeline for ML talent starts here.
nyt/Top Stories
|