Hacker News. Daily summary. Top 20 stories.
Smarter elevator algorithms don't always reduce wait times; simple up/down buttons often outperform destination dispatch.
Two camps: those defending simple up/down buttons and those noting real-world constraints like cap-ex and code.
Inference APIs are moving toward non-portable session state, locking users into single providers.
Broad agreement on lock-in concerns, with pushback on whether all opaque features are unjustified.
DeepSeek releases V4-Flash public beta with major agent benchmark gains.
Mixed sentiment: excitement over benchmark gains tempered by harness comparability concerns.
QM is a multi-agent harness for startups, offering scoped workspaces in Slack and on the web.
Most commenters defended the policy of accepting only human-written text contributions, not code.
Tailscale acknowledges its zero-trust product didn't prevent lateral movement after a credential leak at Hugging Face.
Praise for Tailscale's candid response mixed with skepticism it's primarily a marketing push.
DeepSeek V4 Flash 0731 offers strong performance at roughly one-third the cost of OpenAI Luna Max effort.
DeepSeek V4 Flash is called 2-3x cheaper than OpenAI Luna Max, but slower.
Google used AI agents to find and fix over 1,000 Chrome security bugs in mid-2026.
Skepticism about the blog's one-sided success metrics vs. belief in AI's real security benefits.
Hugh Howey marks the end of a 20-year window where writing was hard but publishing was easy.
Debate over whether LLMs are a printing press for text or a substitute for human writing.
WASTE runs the 2.78-trillion-parameter Kimi K3 model on a 64 GB laptop by streaming expert weights from NVMe.
Sitting still and being content for three minutes reveals a deep human unease with the present moment.
Multiple commenters reject the premise, saying sitting still comes naturally and the article describes anxiety, not universal experience.
Satirical dialogue about AI workers being laid off via video call.
Split between reading the story as AI agents or humans dependent on tokens for survival.
Food companies have filed 239 lawsuits from 2010-2025 against public health laws globally.
Split between critics of corporate lawsuits and skeptics of the report's framing
JEP 401 Value Objects (Preview) merged into OpenJDK master after years of work.
Praise for the project's long-term discipline, with debate over Integer identity and the 63-bit atomicity limit.
AI speeds prototypes but not the hard work of building production-grade software.
Two camps: those agreeing the hard part remains, and those saying AI compresses the whole pipeline.
OpenAI claims a next-gen model solved 10 long-standing math and CS problems.
Two camps: one excited by AI's potential, one fearing math careers will collapse like chess.
Vienna Standard Mean Ocean Water (VSMOW) costs $120,000 per gallon.
Commenters compared VSMOW to other costly NIST reference materials like peanut butter and cigarettes.
BMW delivered a full-screen Spider-Man ad to paid-for vehicles' dashboard displays.
Broad criticism of BMW for violating ownership expectations after promising no in-car ads.
A blogger modded a cheap Thunderbolt 3 adapter for 25 GbE on a Mac Studio.
The thread centered on macOS SMB without RDMA as the true speed limit.
Gander is an open-source Android file viewer with zero permissions.
Privacy enthusiasts debated whether zero-permission apps are truly isolated from the network.
Research suggests AI 'reasoning' models may rely on pattern matching, not logical deduction.
Models work, but their 'reasoning' may be pattern matching, commenters are split on significance.