|
📖 Read In Depth
|
GLM 5.2 beats Claude in our benchmarks
Semgrep's engineering team benchmarked GLM 5.2 (a Chinese open-weight model) against Claude on cybersecurity tasks and found it competitive or superior. This matters beyond the headline — it's a concrete signal that the AI capability gap between US and Chinese frontier models is closing on specialized domains, with real implications for export controls and the compute race.
hn/Best Stories
|
Asian AI startups launch Mythos-like models
Asian AI startups are shipping frontier-class models as Anthropic's export restrictions drag on, effectively ceding market ground and accelerating capability diffusion outside the US. The export ban meant to contain AI development may be backfiring by creating a vacuum that Chinese and other Asian labs are rushing to fill.
hn/Best Stories
|
Anonymous GitHub account mass-dropping undisclosed 0-days
An anonymous GitHub account is mass-publishing undisclosed zero-days — this is a significant security event with implications for the entire software ecosystem. The HN thread likely has context on who might be behind it, which targets are affected, and what the responsible disclosure norms breakdown looks like at scale.
hn/Best Stories
|
HackerRank open sourced its ATS. My resume scored 90/100. Oh wait 74. No – 88
HackerRank open-sourced its applicant tracking system and the author tested their own resume — getting wildly inconsistent scores (74, 88, 90/100 across runs). A practical investigation into how non-deterministic ML scoring systems create the illusion of objectivity in hiring, with real consequences for engineers navigating the job market.
hn/Best Stories
|
Age verification is just a precursor to automated attribution of speech
Argues that age verification mandates are a trojan horse for building persistent identity-to-speech attribution infrastructure — once you've verified who someone is to check their age, that verification layer can be repurposed. A technically grounded civil liberties argument worth reading carefully given the KIDS Act momentum.
hn/Best Stories
|
Zuckerberg's war on whistleblowers
Cory Doctorow on Meta surveilling and silencing whistleblowers — specifically the 'Careless People' author who claims Meta tracked her for 12 months to enforce a gag order, with Google reportedly assisting suppression of critical articles. A sharp piece on platform power, legal intimidation, and what corporate surveillance looks like when turned against critics.
hn/Best Stories
|
AMD Strix Halo RDMA Cluster Setup Guide
A practical setup guide for building an RDMA-connected inference cluster using AMD Strix Halo APUs — consumer/prosumer hardware running vLLM at cluster scale. Relevant for anyone thinking about inference infrastructure outside the hyperscaler stack; the AMD angle matters as an alternative to Nvidia-dependent deployments.
hn/Best Stories
|
A.I. ‘Employees’ Might Disrupt Work in Unexpected Ways
Scholars are documenting 'unknown unknowns' of AI in the workplace — second-order effects that undermine productivity gains and create unexpected failure modes that don't show up in standard ROI analyses. More substantive than typical AI-at-work coverage; engages with the incentive structures that lead companies to deploy before understanding consequences.
nyt/Business
|
Suspicious Discontinuities (2020)
Dan Luu's classic 2020 post on suspicious discontinuities in performance benchmarks — cases where results jump implausibly at round numbers, suggesting measurement gaming or selection effects. Resurfaces on HN now with fresh relevance given the benchmarking wars between frontier AI labs; the framework for detecting benchmark manipulation applies directly to current AI capability claims.
hn/Best Stories
|
|
⚡ FYI
|
Historical memory prices 1960-2026
Stanford's historical dataset of memory prices from 1960 to 2026 — a clean longitudinal view of one of the most important cost curves in computing. Useful context for understanding why Apple is raising Mac prices now (Article 10 cites soaring memory/storage chip costs) and for grounding any reasoning about AI infrastructure economics.
hn/Best Stories
|
Meta improves Brain2QWERTY, a system that can decode text from brain activity to enable typing using non-invasive technologies, MEG and EEG
Meta has improved Brain2QWERTY, a non-invasive BCI system (MEG/EEG) that decodes intended text from brain activity. Still early-stage but represents meaningful progress toward high-bandwidth, non-surgical brain-computer interfaces — a different track than Neuralink's invasive approach.
reddit/r/singularity
|
LineShine, a Chinese supercomputer, has topped the global supercomputer ranking
LineShine, a Chinese supercomputer built on Huawei CPUs, has taken the top spot in the TOP500 global rankings. Significant for the semiconductor geopolitics story — it demonstrates that US export controls on Nvidia GPUs haven't stopped China from building world-class HPC infrastructure through domestic chip development.
reddit/r/singularity
|
Voters Think A.I. Is Terrible. In Campaigns, It’s Everywhere.
AI-generated content is ubiquitous in 2026 political campaigns even as voters say they distrust it — voter data analysis, micro-targeted messaging, and synthetic media at scale. The gap between public perception and actual deployment is itself a data point about how adoption outpaces public deliberation.
nyt/Technology
|