Top Stories

LLMs Are Eroding My Software Engineering Career and I Don’t Know What to Do

995 points · human-in-the-loop.bearblog.dev

A working engineer’s raw account of watching AI assistants hollow out the parts of the job that used to build mastery — and the anxiety of not knowing where that leaves a mid-career developer. With nearly a thousand comments, it clearly hit a nerve. The thread is less about whether the tools are good and more about the lived experience of skill, identity, and job security shifting under people’s feet, which is exactly why HN can’t stop arguing about it.

Building from Zero After Addiction, Prison, and a Felony

661 points · gavinray97.github.io

A deeply personal story of rebuilding a software career — and a life — from the bottom. It resonated because so much of the HN crowd values self-teaching, second chances, and proof that the path into tech isn’t always a tidy CS degree. The comments turned into a genuinely warm space for people sharing their own non-linear roads in.

How’s Linear So Fast? A Technical Breakdown

424 points · performance.dev

Linear’s snappy feel has become a benchmark other apps get measured against, and this teardown digs into the local-first sync engine, optimistic updates, and caching tricks behind it. Engineers love this kind of post because “perceived performance” is half product design and half deep architecture, and Linear has quietly become the case study everyone points to when arguing that web apps don’t have to feel sluggish.

The 29th International Obfuscated C Code Contest (IOCCC) 2025 Winners

396 points · ioccc.org

The legendary contest celebrating gloriously unreadable-yet-functional C is back with a new crop of winners. It’s a beloved HN tradition — part art project, part flex of how much you can hide in a few hundred bytes of pointer arithmetic. Half the fun is people unpacking how each entry actually works.

DeepSeek V4 Pro Beats GPT-5.5 Pro on Precision

288 points · runtimewire.com

Another round in the open-vs-closed model race, with DeepSeek’s latest claiming the edge on precision-focused benchmarks against OpenAI’s flagship. The HN discussion is the usual mix of benchmark skepticism and real interest in how an open-weights challenger keeps closing the gap — and what that means for anyone betting their stack on a single proprietary provider.

I Design with Claude More Than Figma Now

289 points · blog.janestreet.com

A designer makes the case that prompting an AI coding agent to render real, interactive UI has started to beat pushing pixels in Figma for a lot of day-to-day work. It’s a provocative claim that cuts at the boundary between design and engineering, and the comments wrestle with whether “design in code” finally has the tooling to be the default — or whether this only works for certain kinds of products.

Show HN: Lathe – Use LLMs to Learn a New Domain, Not Skip Past It

324 points · github.com

A refreshing counter-take to “let the AI do it for you”: Lathe is built to use LLMs as a tutor that deepens your understanding of a new domain rather than a shortcut that hollows it out. Given the day’s anxiety about AI eroding skills, the framing landed well — people are hungry for tools that make them more capable, not more dependent.

Dopamine Fracking

360 points · igerman.cc

A sharp essay on how modern apps drill ever deeper for fleeting hits of engagement, strip-mining attention the way fracking extracts the last of a resource. The metaphor stuck, and the thread became a broader reckoning about feeds, notifications, and the incentives quietly shaping how we all spend our days.

The Smallest Brain You Can Build: A Perceptron in Python

207 points · ranpara.net

A clean, from-scratch walkthrough of the perceptron — the humble building block underneath modern neural nets. In an era of trillion-parameter models, HN clearly appreciated a back-to-basics explainer that makes the core idea click in a few dozen lines of Python. Great for anyone who wants to actually understand what’s under the hood.