Firefox on iOS finally blocks ads by default, but conveniently spares Google's own search ads.
What the article says
- Firefox for iOS now ships a built in ad blocker that strips out ads at the network level using an EasyList based filter.
- It leaves search engine ads alone, so ads on Google, Bing and DuckDuckGo results still show up.
- Sponsored posts on Firefox's own home and new tab page also stay untouched.
- The article text wasn't available to read, so these points are pieced together from the title and the discussion.
What HN is saying
- Most commenters welcome the move and think ad blocking belongs in every browser by default, not as an add on.
- The sharpest complaint is the carve out for search ads. People call Google's search ads a major source of scams and malware, and suspect Mozilla is protecting the money Google pays it rather than protecting users.
- Several people say dedicated tools like Safari's uBlock Origin Lite, Wipr, or AdGuard already block more than this does.
- A side thread blames Apple for forcing every iOS browser to run on WebKit, which is why Firefox still can't bring its own engine or full extension support to the phone outside the EU and Japan.
Anthropic quietly publishes Claude's full system prompts, and the comments dig into what they reveal.
What the article says
- Anthropic keeps a public page tracking the system prompts used by Claude's chat apps, the instructions that set tone, safety rules, and behavior before every conversation starts.
- These prompts have grown enormously over time, from a few hundred words in early versions to several thousand now.
- One newer addition tells Claude how to explain that some of its replies are actually being handled by a different, more cautious model for safety reasons.
- The prompts also cover small stuff, like avoiding emojis unless the user uses them first, and steering away from words like genuinely or honestly.
- The page does not include the tool definitions or the separate prompt used for Claude Code, which people find more interesting and harder to get.
What HN is saying
- Commenters compare prompt versions over time and note some sections read like legal cover more than genuine model guidance.
- Several people say the instructions to be brief and avoid stock phrases clearly are not working, since Claude still gives long, hedge-heavy answers.
- A few point out contradictions in the prompt and argue that conflicting instructions quietly hurt output quality even outside the contradiction itself.
- One person shares that a prompt nudging Claude to prioritize wellbeing over blind task completion actually helped them step back from overwork.
- Others question why the system prompt is not just trained into the model, with replies explaining it needs to stay editable and is cheap thanks to prompt caching.
A Novo Nordisk funded study says Ozempic may cut dementia risk, but it never tested anyone's memory.
What the article says
- The article itself couldn't be loaded, so this is pulled from the study details discussed on HN.
- Researchers reran data from a major weight loss drug trial, but instead of testing memory, they tracked a blood protein score meant to predict future dementia risk.
- People taking semaglutide saw that risk score rise much more slowly over two years than people on placebo.
- The study was funded by Novo Nordisk, the drug's maker, and some authors work for or hold shares in the company.
- Novo's own earlier trials testing whether semaglutide slows cognitive decline in people who already have Alzheimer's failed to show a benefit.
What HN is saying
- Many commenters are fans of GLP1 drugs personally, but they push back hard on this particular study.
- The sharpest complaint is that it only tracked a proxy blood marker instead of real dementia diagnoses, and that a drugmaker funding research on its own blockbuster deserves extra scrutiny.
- Several people argue any brain benefit might just be ordinary weight loss and better blood sugar control, nothing unique to the drug.
- One detailed comment points out older adults often lose weight quietly in the years before a dementia diagnosis, which could make the drug look better than it actually is.
- There are also plenty of personal stories, big weight loss and real health gains, alongside complaints about low energy and joint pain.
A solo hobbyist used Codex on autopilot for two weeks and made a matrix algorithm 232 times faster than the baseline.
What the article says
- A GPU coding contest asked entrants to speed up a standard matrix factorization step used in machine learning math.
- The author let an AI coding agent run for hours or overnight, testing ideas, profiling the results, and improving the code on its own.
- He kept several competing versions running at once instead of just one, which helped the AI escape getting stuck on the same tweak.
- The approach placed him twelfth out of a hundred eighty three entrants, beaten mostly by people with deep GPU expertise.
- He also shares the exact instructions and workflow habits he gave the agent to keep it productive over many days.
What HN is saying
- Commenters swapped their own stories of AI agents finding real speedups in codecs, video transcoding, and networking libraries.
- A sharp pushback noted the top contest entries often broke on any input beyond the exact test cases, since agents optimize narrowly rather than write general, robust code.
- Some think this works so well for GPU code because performance can be checked automatically, unlike say a typical web service where correctness is fuzzier.
- One heavily downvoted comment claimed Anthropic deliberately cripples Claude for this kind of self improving research to protect its edge, which others did not confirm.
- A few people simply praised the writeup itself for reading like an actual human wrote it, not an AI summary.
An engineer in Trinidad explains why ten cent RISC-V chips matter more than clean instruction set design.
What the article says
- A Trinidad based embedded engineer answers a widely read essay that trashed RISC-V's design.
- His real argument is about access, not elegance. Shipping to his country can cost far more than the chips themselves, so cheap open parts change who gets to build things.
- He points out the original critic actually worked out from scratch that RISC-V is perfect for cheap microcontrollers, then complained anyway.
- ARM keeps small chips and powerful chips walled off by licensing. RISC-V lets the same instruction set scale from a ten cent part up to a Linux capable chip with no permission needed.
- He argues the gap between cheap and powerful hardware isn't really about physics. It is a business decision that RISC-V lets people route around.
What HN is saying
- Commenters largely agree the two authors are arguing past each other. One cares about high end performance, the other about access for people outside rich countries.
- The sharpest pushback is on the shipping math. Several readers note that if postage costs sixty to two hundred dollars, the ten cent versus one dollar chip price barely matters.
- Others explain the nuance. Western suppliers charge more for compliance and verification, while Chinese sellers on AliExpress ship cheaply because they skip all that.
- A few commenters say RISC-V will win the cheap end mainly on price and openness, not because the chip design is actually better than ARM.
- Several people note the linked article kept crashing under traffic and shared archive links so others could read it.