AI platforms and control set the tone today: Anthropic’s Claude Fable 5 arrives as a stronger coding model, but debate quickly shifts to safety filters, retention, pricing, and claims that it may quietly refuse help on competing AI work. Apple’s new Siri AI preview lands in the same frame, with skepticism about what has really improved, then broadens into a fight over platform access as Apple drops the EU rollout after failing to win a DMA exemption. Elsewhere, xAI’s compute-rental business draws scrutiny, developer security gets a hard look after poisoned Microsoft repositories, and policy stories touch web trust, phone ID rules, and H-1B fees.
-
Claude Fable 5
-
anthropic.com
- 1321 comments
Anthropic released Claude Fable 5, a new top-end coding and reasoning model that shares weights with its more restricted Mythos 5 variant. Early users said it is materially better than Opus 4.8 on hard software tasks, but the launch was overshadowed by aggressive safety filters, 30-day data retention, and pricing that many think pushes the best model out of subscription reach.
-
-
Siri AI
-
apple.com
- 684 comments
Apple posted a new “Siri AI” preview for iOS 27 that promises a more context-aware voice assistant tied into apps, messages, email, photos, Shortcuts, and on-device indexing. The reaction was mostly eye-rolling: readers saw recycled demo tasks, unclear shipping promises, device and regional lockouts, and little evidence Apple has solved Siri’s long-running reliability problem.
-
-
Apple decided not to roll out Siri in EU after denied request for exemption
-
reuters.com
- 578 comments
Reuters reported that Apple asked the EU for a temporary exemption from Digital Markets Act rules so it could launch its new Siri AI in Europe before opening the same system capabilities to rival assistants. The EU refused, Apple pulled the rollout, and the comments mostly saw this as Apple choosing platform control over compliance while wrapping the fight in privacy language.
-
-
xAI is looking more like a datacentre REIT than a frontier lab
-
martinalderson.com
- 530 comments
A blog post argues that xAI’s latest multibillion-dollar compute-rental deals make it look less like a cutting-edge AI lab and more like a company monetizing datacenter capacity. The comments largely agreed the rental business is real and lucrative, but split on whether that proves xAI found a smart temporary business model or exposed that it has fallen behind in the model race.
-
-
Microsoft's open source tools were hacked to steal passwords of AI developers
-
techcrunch.com
- 177 comments
A TechCrunch report says attackers compromised Microsoft-owned open source repositories and planted files that could steal developer credentials, especially from machines running AI coding tools or VS Code tasks. The comments focused less on the article and more on what the incident says about modern developer security: broad GitHub tokens, weak isolation, and auto-executing tooling are turning developer laptops into the softest target in the stack.
-
-
If Claude Fable stops helping you, you'll never know
-
jonready.com
- 152 comments
A blog post argues Anthropic’s new Claude Fable policy lets the model quietly give worse help on work related to building competing AI systems, without telling the user. Commenters treated that as both anti-competitive behavior and a broader warning that cloud AI tools can silently become unreliable in ways teams cannot audit.
-
-
Let's Encrypt bans certificate usage in any US sanctioned territory [pdf]
-
letsencrypt.org
- 244 comments
A diff to Let’s Encrypt’s subscriber agreement looked like it barred certificate use by anyone in US-sanctioned countries, which set off fears that a major chunk of the web’s free HTTPS provider was cutting off Russia, Iran, and other restricted regions. The key signal is that Let’s Encrypt says this was mainly a terms clarification, not a broad new service shutdown, but the wording was bad enough to remind people how much internet trust infrastructure sits inside US legal reach.
-
-
FCC wants to kill burner phones by forcing telecoms to get all customers' IDs
-
404media.co
- 264 comments
A proposed FCC rule would require phone carriers to collect customer identification for all telecom accounts, effectively making anonymous prepaid “burner” phones much harder to get in the US. The reaction was overwhelmingly hostile, with most readers arguing it will expand surveillance and data leaks while doing little to stop spam calls, which mostly come from caller ID spoofing and VoIP.
-
-
Cleaning up after AI rockstar developers
-
codingwithjesse.com
- 320 comments
A blog post argued that AI-assisted “rockstar” coding creates the same kind of hard-to-maintain mess once blamed on cowboy developers, just much faster. Commenters mostly agreed the real risk is not AI writing code, but companies shipping prototypes and unreviewed output as production systems.
-
-
Federal judge blocks H1B visa $100K fee
-
alaskasnewssource.com
- 332 comments
A federal judge blocked the Trump administration’s new $100,000 H-1B visa fee, which would have sharply raised the cost of hiring skilled foreign workers. The story looked like another tech immigration fight, but most of the useful discussion centered on rural Alaska schools and hospitals that rely on visa workers because domestic hiring there is exceptionally hard.
-
-
OpenCV 5 Is Here: The Biggest Leap in Years for Computer Vision
-
opencv.org
- 119 comments
OpenCV 5 is a major new release of the long-running computer vision library, with a redesigned ONNX inference engine, performance gains, and broader AI-model support. Readers liked the speedups and OpenCV’s reach, but most of the conversation fixated on a release post that looked AI-written and on whether OpenCV is still the right production tool versus dedicated runtimes.
-
-
Apple Core AI Framework
-
developer.apple.com
- 105 comments
Apple posted Core AI, a new framework for converting and running neural network models across Apple CPUs, GPUs, and the Neural Engine, with commenters trying to sort out how it changes the existing Core ML and MLX stack. The useful signal is that Apple now seems to be splitting its AI tooling into distinct lanes for on-device neural nets, older non-neural ML, and separate bring-your-own-weights workflows, while developers are more excited about the broader push toward local and OS-level models than this API alone.
-
-
Upcoming breaking changes for NPM v12
-
github.blog
- 44 comments
GitHub says npm v12 will make a few security-focused defaults stricter, most notably turning package install scripts off by default unless you explicitly allow them. The reaction was mostly "finally" because this addresses a long-known npm supply-chain weakness, though people also pointed out gaps around build-time code and release-delay protections.
-
-
We Think the SpaceX IPO Is Overvalued
-
morningstar.com
- 270 comments
Morningstar published a skeptical valuation of the SpaceX IPO, arguing the offering price depends on very optimistic assumptions about Starship reuse and a new business line of orbital AI data centers. Comments mostly agreed the stock may still soar anyway, because buyers see it as a Musk trade and a momentum event rather than a normal fundamentals-driven investment.
-
-
EU-banned pesticides found in rice, tea and spices
-
foodwatch.org
- 280 comments
Foodwatch posted lab results on rice, tea, peppers, cumin, and spice mixes sold in Europe, finding residues of pesticides banned for use in the EU and a smaller set of samples above legal residue limits. The core allegation is a “boomerang effect”: EU firms can still export some banned chemicals abroad, then the resulting food is imported back into Europe.
-
-
Facebook is paying people overseas promoting Alberta separatism
-
cbc.ca
- 162 comments
CBC reported that Meta’s creator payouts are rewarding overseas Facebook users for posting Alberta separatism content because outrage drives engagement. The comments mostly treated this less as a Canada-specific influence plot and more as a predictable side effect of paying for ragebait at global wage differentials.
-
-
Making Graphics Like it's 1993
-
staniks.github.io
- 127 comments
A solo developer posted a deep technical/art breakdown of a retro-style first person shooter built with early-1990s constraints, including software rendering, palette-based shading, and a sprite pipeline generated from 3D models. Readers loved both the final look and the unusually concrete explanation of how the art direction, engine choices, and homemade tools fit together.
-
-
Looking Forward to Postgres 19: Query Hints
-
pgedge.com
- 42 comments
A blog post says PostgreSQL 19 is likely to gain a new "query hints" style feature called plan advice, which lets users nudge the database planner toward safer execution plans without overriding it completely. The comments mostly treated this as overdue pragmatic tooling for edge cases, with a strong warning that it can also become a crutch that hides bad schema, stats, or query design.
-
-
Where is the AI jobs crisis?
-
apollo.com
- 207 comments
An Apollo chief economist argued there is no broad AI-driven jobs collapse yet, pointing to strong U.S. payrolls and job openings. Commenters mostly agreed the post overreached from thin aggregate data, while also saying the pain is real in specific pockets like junior software hiring and the hiring process itself.
-
-
Ask HN: Are you still using a Vision Pro?
-
news.ycombinator.com
- 148 comments
An Ask HN post asked whether people still use Apple’s Vision Pro after two years. The answers split hard between a small group of daily users who swear by it for travel work and private giant-screen media, and a larger group who say weight, glare, battery fuss, text clarity, and social awkwardness turned it into an expensive shelf item.
-
-
Solar Energy Saves Europeans $135M a Day
-
cleantechnica.com
- 135 comments
A Cleantechnica post says Europe’s solar generation is avoiding about $135 million per day in fossil fuel purchases. The comments mostly agreed solar is strategically valuable, but pushed hard on the headline’s framing, saying this is displaced fuel spend rather than net profit once installation, grid, storage, and subsidy costs are counted.
-
-
The iPhone's Last Stand?
-
stratechery.com
- 205 comments
A Stratechery post argues Apple may not need to win the "AI platform" race because the iPhone remains the main consumer device, while Microsoft is pushing a more cloud-and-agent-centric future. Commenters mostly rejected the headline’s drama and focused instead on whether people actually want AI embedded into phones, how Apple will charge for its AI stack, and whether Microsoft can execute on hardware at all.
-
-
Show HN: Gitdot – A better GitHub. Open-source, written in Rust
-
gitdot.io
- 290 comments
Gitdot is a new open-source Git hosting service that can already handle accounts, orgs, public and private repos, and GitHub imports, but it ships with a stark keyboard-driven UI and very few of the collaboration features people expect. The launch drew attention less for the Rust backend than for a polarized reaction to the design, the missing basics, and the mismatch between the “better GitHub” pitch and what exists today.
-
-
How much do amd64 microarchitecture levels help in Go?
-
lemire.me
- 44 comments
A benchmark post tested Go binaries built for newer x86-64 CPU feature levels and found meaningful gains from moving the baseline to amd64 v2, smaller gains at v3, and little current benefit from v4 because Go does not yet emit AVX-512. The comments mostly agreed the practical target today is "v2 by default, v3 when you control the hardware," while arguing over portability, runtime dispatch, and whether compilers should automate multi-version binaries.
-
-
Is Grep All You Need? How Agent Harnesses Reshape Agentic Search
-
arxiv.org
- 52 comments
A paper compares plain text search like grep against vector and hybrid retrieval inside LLM agent harnesses, and claims grep often wins on long-context search tasks. The comments largely agree the result is real inside coding-oriented harnesses, but say the title overreaches because the benchmark favors literal lookup and the bigger lesson is to design the harness and data layout, not worship one retrieval method.
-
-
The beauty and simplicity of the good old C-style void* in C++
-
giodicanio.com
- 134 comments
A blog post argued that plain C-style `void*` plus a size is the nicest way to pass a raw memory blob into C++, and mocked alternatives like `uint8_t*` and `std::span`. Most commenters rejected that framing and said modern C++ should make byte intent, size, ownership, and type boundaries explicit, especially because raw object bytes are full of portability and correctness traps.
-
-
What it feels like to work with Mythos
-
oneusefulthing.org
- 144 comments
A Wharton professor with early access to Anthropic’s “Mythos” family posted a first-person account of using its coding model “Fable” to build projects like travel-time maps, games, and research artifacts. The comments mostly pushed back on the article’s hype, saying the demos looked polished but shallow, often wrong, expensive to run, and light on the code quality questions that matter in real software work.
-
-
Corrupting a ZFS File on Purpose
-
oshogbo.com
- 19 comments
A blog post walks through deliberately corrupting a file stored on ZFS to see how the filesystem detects and repairs bad data. The comments turn into a practical discussion of what kinds of storage corruption really happen in the wild, and why end-to-end checksums still matter even when drives have their own error correction.
-
-
Show HN: Gravity – interactive solar-system simulator, from Newton to Einstein
-
qunabu.github.io
- 34 comments
A Show HN post shared a browser-based solar-system simulator that teaches basic orbits step by step, then switches to an N-body mode and a conceptual Einstein finale. Readers liked the teaching approach and responsiveness, but quickly pointed out several physics and visualization mistakes around orbital speeds, Earth’s axis, and what the “relativity” section actually does.
-
-
Can LLMs Beat Classical Hyperparameter Optimization Algorithms?
-
- 15 comments
A new paper benchmarks large language models against standard hyperparameter optimization methods and finds that LLMs alone do not win, but a hybrid “centaur” setup that combines them with classical search can do better. Commenters mostly read it as another case where LLMs are useful as a wrapper around established optimizers, not a replacement for them.
-