Top Stories

DeepSeek v4

1827 points · deepseek.com

Far and away the biggest story of the day — and one of the highest-voted HN posts of the month — DeepSeek dropped v4, with a 1428-comment thread that is mostly people running benchmarks against it in real time. Early consensus from the practitioner-heavy sub-threads: it’s another meaningful jump on coding and long-context reasoning, the price-per-token continues to be aggressive, and the open-weights release is structured to actually be usable rather than ceremonial. The recurring meta-conversation is the same one HN has been having for six months now — that the assumed gap between frontier closed models and open-weights releases keeps compressing, and that pricing pressure on the closed labs is no longer hypothetical. Pair this with the Google–Anthropic story below and today reads as a clear inflection in the AI capital cycle.

Google plans to invest up to $40B in Anthropic

383 points · bloomberg.com

Bloomberg reports Google is preparing a fresh round of investment in Anthropic that could reach $40B — a number that, if confirmed, reshapes the AI infrastructure conversation more than any single funding event since Microsoft’s original OpenAI commitment. The 422-comment thread is, predictably, the place where every flavor of HN argument is happening at once: cloud-strategy people parsing what this means for TPU vs. Trainium vs. Nvidia, antitrust-watchers asking how this clears given the existing Anthropic stake, and a long sub-thread on whether the math actually pencils out at current Anthropic revenue multiples. Worth reading alongside today’s “I cancelled Claude” post — the two stories sit awkwardly next to each other and the thread has noticed.

I cancelled Claude: Token issues, declining quality, and poor support

797 points · nickyreinert.de

A pointed customer post-mortem from a longtime Claude user explaining, with receipts, why he’s churning — covering the recent Claude Code quality regressions, what he describes as opaque token-counting changes, and a frustrating support experience. The 482-comment thread is the one to read if you want the unvarnished developer view of where Anthropic’s product reliability sits this week, and it dovetails directly with the CC-Canary regression-detection tool also on the front page. Anthropic engineers are in the thread responding in places, which makes for a useful real-time look at how the company is handling the heat. The broader tell: posts like this getting 800 points used to be rare; in 2026 they’re a recurring genre, and every frontier vendor should be reading them.

Sabotaging projects by overthinking, scope creep, and structural diffing

369 points · kevinlynagh.com

Kevin Lynagh’s essay on the specific patterns by which engineers torpedo their own projects — over-architecting upfront, expanding scope in response to imagined requirements, and getting stuck doing “structural diffing” against an idealized version of the code that doesn’t exist yet. The 93-comment thread is the kind of A-tier engineering-management discussion HN does well: lots of senior engineers recognizing themselves uncomfortably in the descriptions, plus a useful sub-thread on whether AI coding tools amplify or attenuate these failure modes. The piece is short, opinionated, and worth reading for anyone running a codebase or mentoring people who are.

There Will Be a Scientific Theory of Deep Learning

161 points · arxiv.org

A position paper arguing — against the current vibes — that deep learning is going to get an actual scientific theory, not just an ever-growing pile of empirical scaling laws. The 50-comment thread is a serious one: ML researchers debating whether the field’s resistance to theory is principled or just path-dependent, and how the recent run of mechanistic-interpretability and feature-circuit results changes the priors. If you’re tired of “deep learning is just alchemy” takes, this is a meatier read in the opposite direction.

SDL Now Supports DOS

229 points · github.com/libsdl-org

A merged PR adding native DOS support to SDL — yes, in 2026. The 83-comment thread is a small celebration of preservation engineering: people talking about which retro-gaming and demoscene workflows this unblocks, the specific DOS extender story underneath the patch, and a tangent about how SDL’s portability story keeps quietly extending in directions nobody asked for and a few thousand people are delighted by. A reminder that meaningful infrastructure work still happens at the long-tail end of the platform list.

MacBook Neo and how the iPad should be

217 points · craigmod.com

Craig Mod on what an actually-good iPad would look like if Apple let it be one — riffing off the rumored MacBook Neo to argue that the iPad’s hardware has long since outrun its software model, and that the gap is finally costing Apple meaningful pro-creative users. The 123-comment thread is a familiar but well-argued debate: tablet-OS purists pushing back, longtime iPad-as-laptop users sharing the specific places they keep hitting walls, and a sharp sub-thread on whether iPadOS 26 actually moved the needle. As an essay it’s the most coherent version of the “iPad needs to grow up” argument that’s circulated this year.

My audio interface has SSH enabled by default

175 points · hhh.hn

A reverse-engineering post on the RØDECaster Duo: the author opens up the firmware, finds SSH listening on the device with default credentials, and walks through what a curious owner can do with that. The 47-comment thread is the usual mix of “this is a feature, not a bug” and “this is a security disaster waiting to happen,” but the more interesting current is people pointing out how often consumer audio and video gear ships with embedded Linux running services nobody on the marketing page mentions. If you own one of these, the post is a fun read; if you’re shipping firmware, it’s a cautionary one.

Email could have been X.400 times better

134 points · buttondown.com/blog

A nicely-done historical piece walking through X.400 — the OSI-stack email standard that was supposed to be the future before SMTP’s “worse is better” trajectory ate everything — and asking what the modern email stack would look like if we’d taken the structured route. The 135-comment thread is unusually thoughtful by HN protocol-history standards: people who actually shipped X.400 systems weighing in on what was real and what was marketing, plus a side-thread on whether the structured-messaging dream is now being lived out in MCP and agent-to-agent protocols. Worth reading if you’ve ever wondered why email looks the way it does.

Show HN: Browser Harness – Gives LLM freedom to complete any browser task

89 points · github.com/browser-use

The browser-use team shipped Browser Harness, a more general framework for letting an LLM drive a real browser end-to-end on long-horizon tasks — auth flows, multi-page workflows, the whole bit. The 40-comment thread is the practical kind: people comparing it to Anthropic’s Computer Use, OpenAI’s Operator, and Playwright’s MCP server, plus the inevitable “but what about CAPTCHAs and ToS” discussion. With Anthropic, OpenAI, and Google all betting heavily on browser-driving agents, the open-source side of this stack is finally getting interesting, and Browser Harness is a clean reference point for where it’s at.