Today’s thread is the internet turning into a more controlled, less trustworthy place: one story argues AI summaries, weak search, and vanishing archives are eroding the web’s memory, while others trace the spread of age checks, online identity rules, and live facial recognition, plus a false homicide alert that shows how brittle surveillance systems can be. Policy and enforcement also surface in France’s telemarketing ban and the market value of early access to Truth Social posts. Elsewhere, there’s a security failure exposing hidden LLM reasoning, a look at Nvidia’s demand risk, Mojo 1.0, and England’s progress toward eliminating hepatitis C.
-
As AI eats the web, the internet’s collective memory is disappearing
-
thewalrus.ca
- 866 comments
A long Walrus piece argues that AI summaries, weak search, and disappearing archives are making the web harder to search and less reliable as a public memory system. The comments mostly agreed the problem is real, but pinned it less on AI alone than on years of Google de-indexing, SEO spam, walled gardens like Discord, and business models that stopped rewarding people for publishing useful pages.
-
-
The UK's war on anonymity has come to America
-
effort.news
- 709 comments
A blog post argues that UK child-safety groups and laws are helping normalize age checks and online identity requirements in the US, especially through state bills that push age signals from devices to apps and sites. The comments mostly agreed that child-safety rhetoric is being used to expand surveillance, but split over whether some US laws are actually narrower parental-control rules rather than full ID mandates.
-
-
France to ban unsolicited telemarketing calls
-
lemonde.fr
- 473 comments
France has started an outright ban on unsolicited telemarketing calls, replacing its older opt-out system. The reaction was broadly supportive, but the useful signal was that this should cut legal sales calls more than scam calls unless carriers and regulators also clamp down on spoofing, foreign call centers, and enforcement.
-
-
London Underground begins scanning passengers' faces
-
btp.police.uk
- 238 comments
British Transport Police said it is expanding a live facial recognition trial into London Underground stations, with marked camera zones and alternate routes for people who want to avoid them. The comments were mostly less about the tech itself than about whether Britain has already normalized broad surveillance without getting much extra safety in return.
-
-
Stealing Reasoning Traces from Proprietary LLM APIs
-
stolen-thoughts.com
- 200 comments
A paper shows that hidden reasoning text from OpenAI, Anthropic, and Google APIs could be recovered by replaying encrypted reasoning blobs into weaker sibling models and jailbreaking those models to print the plaintext. Commenters treated it as both a real security failure and another example of labs restricting useful model internals to slow distillation and competition.
-
-
Woman pulled over twice after Flock-linked software connected her to homicide
-
guessingheadlights.com
- 223 comments
A Wisconsin woman was pulled over at gunpoint twice after Flock license-plate alerts tied her car to a homicide because police left a bad flag active in their system. The comments treated it less as a one-off software bug than a preview of what happens when mass surveillance gets wired into sloppy police databases and aggressive stop tactics.
-
-
England set to be one of the first countries to eliminate hepatitis C
-
bbc.com
- 346 comments
A BBC report says England is close to meeting the World Health Organization’s hepatitis C elimination targets after diagnosing and treating more than 100,000 people since 2015. Commenters mostly used it as a springboard for two things: how modern hepatitis C treatment became a rare real cure, and how public health wins are fragile when screening, vaccination, and trust break down.
-
-
More than 10 firms pay up to $100k a month for access to Truth Social posts
-
- 283 comments
The BBC reported that more than 10 financial firms are paying up to $100,000 a month for early access to posts on Truth Social, letting traders see some presidential posts before the public. Readers treated it as an overt market-moving access product, with the sharpest debate centered not on whether it looks corrupt, but on whether it is clearly illegal or merely another norm the system no longer enforces.
-
-
Nvidia's Risky Business
-
stratechery.com
- 134 comments
A Stratechery post argued that Nvidia’s biggest risk is not losing AI outright, but financing a huge expansion built on demand that may stop accelerating fast enough to justify it. Comments mostly agreed the danger is valuation and capex math more than product collapse, then dug into whether CUDA, local inference, and custom chips can actually loosen Nvidia’s grip.
-
-
Mojo 1.0
-
modular.com
- 118 comments
Modular announced Mojo 1.0, its Python-like systems language aimed at high-performance and accelerator-heavy workloads, especially GPUs. The reaction was less about the language features than about two unresolved questions: what Mojo is actually for beyond the pitch, and why a “1.0” compiler is still not open source.
-
-
OpenAI’s head of ethics leaves less than a year after joining
-
ft.com
- 329 comments
A Financial Times report says OpenAI’s head of ethics, Chloé Bakalar, has left less than a year after joining, after earlier serving as Meta’s chief ethicist. With the article offering almost no explanation, commenters treated the departure less as a personnel note and more as evidence that ethics roles at major AI labs are either powerless, performative, or being reshaped into operational alignment work.
-
-
Go is an ideal language for AI-assisted software engineering
-
developers.googleblog.com
- 273 comments
A Google Developers post argues that Go is especially well suited to AI coding because the language is simple, standardized, and backed by strong built-in tooling. The comments mostly agreed that Go is easy for models to generate and easy for teams to operationalize, but pushed hard on the lack of evidence and on whether stronger type systems like Rust are actually safer once AI is writing more of the code.
-
-
What's the best programming language for coding agents?
-
danluu.com
- 180 comments
A blog post tried to answer which programming languages work best for coding agents by measuring token use and task performance across languages, then concluded most common claims are weaker than people think. The comments mostly landed on a different takeaway: language differences exist, but tooling, feedback loops, and codebase structure matter more than token counts or syntax trivia.
-
-
Apple Silicon and macOS VMs: Faster LLM Inference with llama.cpp
-
github.com
- 43 comments
A blog post from the Lume team shows a big llama.cpp speedup inside macOS virtual machines on Apple Silicon by changing what GPU capabilities the guest VM reports to Metal. Commenters quickly clarified that this is not a general Mac inference breakthrough on bare metal, but a fix for a VM-specific path that had been forcing slower kernels.
-
-
To save C, we must save ABI (2022)
-
thephd.dev
- 137 comments
A 2022 blog post argues that C cannot evolve cleanly unless the ecosystem gets better tools for changing binary interfaces without breaking old code, using `intmax_t` and similar type choices as the motivating example. Comments mostly agreed the core problem is real, but said the post overstates what the C standard can fix because ABIs are set by platforms, toolchains, and existing libraries, not by C alone.
-
-
Compression is prediction
-
ngrok.com
- 92 comments
An ngrok blog post argues that data compression and prediction are two views of the same process, using examples from file compression and language models. Readers liked the explanation, but the useful pushback was that the equivalence only really holds inside the data distribution you trained or compressed against.
-
-
What I learned by putting GitHub Copilot behind a MitM proxy
-
lighthousenewsletter.com
- 24 comments
A developer put GitHub Copilot behind a man-in-the-middle proxy to inspect what the coding assistant actually sends and stores. The writeup shows how Copilot routes requests, pulls extra file context including possible secrets, and keeps local session history, which raised more concern than surprise in the comments.
-
-
Chicken Scheme 6.0
-
code.call-cc.org
- 50 comments
CHICKEN Scheme 6.0 is a major release of the Scheme-to-C compiler and runtime, adding full R7RS support, native UTF-8 strings, and better C interop. Comments treated it as a real step up for people who want small compiled binaries and easy embedding into C-heavy systems, while also resurfacing Scheme’s old portability and library-fragmentation headaches.
-
-
OpenSSH 10.5/10.5p1
-
openssh.org
- 30 comments
OpenSSH 10.5/10.5p1 is out with routine fixes plus a small but practical new `ssh -Z` mode that shows which public keys the client will try, in order. The release notes also stand out for saying AI-assisted security bug reports are welcome, which pushed most of the conversation toward how far that acceptance really goes.
-
-
Faster floating point math with Rust's new API
-
pythonspeed.com
- 45 comments
A blog post shows how Rust’s new floating-point API lets you opt into algebraic rewrites like reassociation, which can make numeric code faster by relaxing strict IEEE 754 behavior. The comments mostly sharpen the tradeoff: this is useful, but only when you fully understand that you are changing program semantics, not just flipping on a harmless speed boost.
-
-
Show HN: Git-knife – edit commit messages, authors, and dates like a spreadsheet
-
- 85 comments
A Show HN post introduced Git-knife, a small desktop app for bulk-editing Git commit metadata like messages, authors, and dates in a spreadsheet-style UI. Commenters mostly saw it as a handy wrapper around painful history-rewrite workflows, with the sharpest caveat being that rewritten history breaks signatures and should stay confined to local cleanup or carefully managed branches.
-
-
Show HN: iPhone app takes simultaneous images from 2 lenses, fuses into 1 photo
-
photosynthesis.camera
- 189 comments
An indie iPhone camera app called Photosynthesis says it captures from two rear lenses at the same time and merges them into one image, aiming to keep a wider field of view while adding more detail in the center from the second lens. Readers found the idea clever, but most of the real energy went into whether the image improvements are actually visible and why an offline camera app is sold as a subscription.
-
-
Grok Bot
-
x.ai
- 105 comments
xAI launched Grok Bot, a hosted AI agent product where each bot gets its own cloud computer, can stay logged into apps, run continuously, and coordinate with other bots. Readers saw the product shape as familiar and potentially useful, but the response was dominated by distrust of giving a Musk-owned service persistent access to accounts, data, and browser sessions.
-
-
Jolt: Clojure compiler implemented with Chez Scheme
-
jolt-lang.github.io
- 51 comments
Jolt is a new Clojure implementation that runs on Chez Scheme and can build small standalone native executables instead of targeting the JVM. The post drew interest for its drop-in compatibility ambitions and native FFI, but a lot of the conversation centered on whether its breakneck pace and visible LLM use should change how seriously people take it.
-
-
H3-metal – Native MiniMax-H3 inference for Apple Silicon
-
- 93 comments
A new C implementation brings MiniMax H3 image and video generation to Apple Silicon without the usual Python stack, aiming to make a large open video model run natively on Macs. The comments focus on what that actually buys you in practice: modest speedups on high-end Macs, far lower performance than Nvidia GPUs, and a lot of confusion around memory needs and quantized model formats.
-
-
Closing Canario Terminal source code
-
rapha.land
- 51 comments
A developer says they are stopping public source releases for their terminal emulator, citing the overhead and spam that came with being open source, especially AI-generated issues and pull requests. Most of the discussion agreed maintainers owe nobody support, but argued that closing the code is a different choice from simply turning off contribution channels.
-
-
Show HN: Mcptoon – Token-efficient MCP CLI client
-
- 43 comments
A Show HN post introduced mcptoon, a command-line wrapper for Model Context Protocol tools that claims big token savings by shortening tool schemas and outputs. Commenters liked the general goal of cutting MCP overhead, but quickly found the implementation appears to confuse character count with actual LLM tokenization and may drop information models need to call tools correctly.
-