Claude: System Prompts (Claude Platform Docs)
Anthropic publishes the evolving system prompts used by Claude models across web and mobile apps.
Prompts have grown from ~300 to 3,000+ words over time, with some questioning the cost to performance.
Hacker News. Daily summary. Top 20 stories.
Anthropic publishes the evolving system prompts used by Claude models across web and mobile apps.
Prompts have grown from ~300 to 3,000+ words over time, with some questioning the cost to performance.
Firefox for iOS now blocks ads natively, with exceptions for search and new tab page.
The thread focuses on Mozilla's selective ad blocking and Apple's restrictive iOS browser engine policy.
Qwen 3.8 27B is a capable local LLM but defaults to excessive reasoning effort.
Overthinking is the dominant complaint; commenters suggest low reasoning or Glimmer as alternatives.
Cloudflare silently injects analytics JavaScript when users switch nameservers, even on JS-free sites.
Users condemn Cloudflare's opt-out analytics injection, calling it a MITM attack and warning about enshittification.
RISC-V's low cost and openness make it accessible to engineers in developing countries.
Appreciation for the non-US perspective, with debate over shipping costs and a subthread on HN's shifting demographics.
Anthropic's text watermarking corrupts word choice, violating promises of imperceptibility.
Split: some see watermarking as futile theater, others defend detection of AI output.
Stripe is reportedly acquiring AI model gateway OpenRouter for over $7 billion.
Skepticism over valuation and strategic fit, countered by defense of the middleman business model.
A Google Scholar search reveals research papers using 'kidney disappointment' instead of 'kidney failure'.
Commenters attribute the phrase to machine translation or plagiarism avoidance via synonym substitution.
AI labs trade world knowledge for reasoning skill as models shrink per parameter.
Debate over whether the knowledge-tradeoff approach serves non-coding users, plus a correction on benchmark recency.
Token brokers are commercializing the resale of AI credits, often at steep discounts.
The thread splits on whether the credits come from legitimate surplus or fraud, with warnings about relay operators logging prompts or swapping models.
Federal agencies used keyword lists to cancel billions in research funding.
Commenters split on whether the administration is maliciously incompetent or deliberately destructive.
LLMs trained only on K-5 material reveal post-training cannot exceed the pretraining data ceiling.
Skepticism about whether the filtered dataset truly excludes advanced concepts vs. poor model recall.
Article body wasn't reachable. HN discussion still summarized.
Two camps: those warning Softbank and Oracle are exposed, those debating local vs. datacenter AI economics.
GIMP 3.3.2 development preview introduces new project file format, non-destructive editing, and PSD improvements.
Non-destructive editing praised; zipped XML format choice questioned by some.
GitHub is investigating degraded API performance, impacting some services.
Growing frustration with GitHub reliability; many considering self-hosted alternatives like GitLab or Gitea.
The two-day weekend is only 100 years old, but hybrid work now blurs its boundaries.
The modern weekend is only 100 years old; the seven-day week itself is a social construct.
A case study on the launch, success, and political shutdown of the IRS's Direct File tax service.
Two camps: those who see Direct File's shutdown as a pure political kill, and those who argue government should never run such services.
St. Lucie nuclear reactor manually shut down after 3 control rods drop into core.
Most commenters see this as a routine safety event; rods dropping is an expected fail-safe behavior.
Anthropic finds AI agents lack coordination, exhibit uniformity, and are vulnerable to deception.
Skepticism that the study is a promo for Anthropic's next model; the behaviors are blamed on training data quality.
Reticulum is a cryptography-based networking stack for building sovereign mesh networks.
Mixed reactions: protocol praised as solid but implementation and maintainer burnout cited as risks.