Top Stories

I don’t think AI will make your processes go faster

607 points · frederickvanbrabant.com

A pointed counterpoint to the “AI productivity revolution” narrative is dominating the front page today. The author argues that bolting LLMs onto broken organizational processes doesn’t make them faster — it just makes the broken steps cheaper, which is usually not where the bottleneck lives. Reviews, approvals, meetings, and context-switching swamp the time you save on the actual writing or coding.

The HN crowd is eating it up because it matches what a lot of engineers are seeing in practice: dashboards full of AI usage metrics, while shipping velocity is flat. Expect this one to make the rounds in Monday-morning Slack channels everywhere.

I turned a $80 RK3562 Android tablet into a Debian Linux workstation

379 points · github.com

A delightful hardware-hacking writeup: the author bought a cheap RK3562-based Android tablet, wiped it, and got Debian running on the bare ARM SoC with working display, touch, Wi-Fi, and battery management. It’s a 4-core Cortex-A53 with a couple of gigs of RAM — not fast, but eminently usable as a portable Linux device for the price of two takeout dinners.

The repo includes the kernel patches, device tree, and rootfs build scripts. Comments are mostly people comparing it to the PineTab and asking whether the same trick works on other RK35xx-series tablets flooding AliExpress.

Show HN: Semble – Code search for agents that uses 98% fewer tokens than grep

346 points · github.com

An open-source tool aimed squarely at the agentic-coding workflow problem: grep dumps thousands of irrelevant lines into your LLM context window, burning tokens and confusing the model. Semble pre-indexes the codebase semantically and returns only the snippets that are actually relevant to the query, with claimed 98% token reduction on benchmark queries.

This sits in a hot category right now — every team building coding agents is hitting the same context-window wall. The thread is full of comparisons to Sourcegraph’s Cody, Aider’s repo-map, and Anthropic’s own MCP servers, with people debating whether semantic search is worth the index-maintenance overhead.

GenCAD

340 points · gencad.github.io

Generative AI for parametric CAD: GenCAD takes a natural-language description (or a reference image) and outputs editable CAD models with proper construction history — sketches, extrusions, fillets — not just dumb meshes. The project page has demo videos of “a bracket with four M3 mounting holes” turning into a real OnShape-style feature tree.

This is the holy grail for hardware engineers who’d rather not click through SolidWorks menus for the hundredth time. Commenters are mostly skeptical it’ll handle real engineering tolerances, but everyone agrees that even rough first-pass geometry from a prompt is a meaningful productivity step.

Tesla Solar Roof is on life support as it pivots to panels

270 points · electrek.co

Electrek reports that Tesla is quietly winding down the Solar Roof product nearly a decade after Musk’s splashy 2016 unveiling. Installer count has cratered, lead times have ballooned, and the company is pushing existing customers toward conventional rooftop panels instead. The piece argues the product never solved its core problem — installation labor — and was always more demo than business line.

HN is using the thread to relitigate the original launch, with several commenters posting receipts of the now-infamous staged demo at Universal Studios. There’s broader discussion of whether building-integrated solar can ever pencil out at residential scale.

Prolog Basics Explained with Pokémon

257 points · unplannedobsolescence.com

A charming intro to Prolog that uses the Pokémon type chart as the running example — facts like effective(water, fire), rules for double-typed Pokémon, and queries that walk you through resolution and unification. It’s the rare programming tutorial that’s actually fun to read.

There are two Prolog posts on the front page today (the other being “Prolog Coding Horror”), which is unusually high Prolog density and has commenters wondering if logic programming is having a small renaissance on the back of the LLM-reasoning conversation.

CUDA Books

201 points · github.com

A curated, well-organized list of books for learning CUDA — from intro texts like “Programming Massively Parallel Processors” through deep dives on PTX assembly, Tensor Cores, and CUTLASS. With everyone trying to write custom kernels for their LLM inference stack right now, this is hitting at the perfect moment.

Comments include the predictable “but ROCm” and “why not Triton” debates, plus a few PhD students recommending older texts that taught them more than any official Nvidia documentation ever did.

Jank now has its own custom IR

154 points · jank-lang.org

Jank is the Clojure-on-LLVM project that’s been steadily shipping for a few years now. The latest milestone: replacing direct LLVM IR generation with a custom higher-level IR that’s better suited for Clojure-specific optimizations like persistent data structure access patterns and seq fusion. Benchmarks in the post show 2-4x speedups on common idioms.

This is the kind of long-haul language-implementation work that the HN crowd loves to dig into. Comments are full of compiler-writers comparing notes on IR design and lamenting how hard Clojure’s runtime semantics make ahead-of-time compilation.

Where Are the Vibecoded Photoshops?

96 points · indiepixel.de

A short, sharp essay asking why — given two years of AI-assisted coding hype — we still don’t have a single new desktop application of Photoshop-level depth and polish that was substantially built by AI. The author’s hypothesis: AI is great at producing the first 60% of an app but falls off a cliff when the work shifts to performance tuning, edge cases, and user-experience polish.

It’s a pointed companion piece to the “AI processes” story above, and the comment thread is the predictable mix of “you’re holding it wrong,” “the AI photoshops exist, they’re just niche,” and “the author has a point and we should sit with it.”