This week's most original Hacker News essays all describe something their authors made

Of 40 widely read essays scored for the week of 26 October 2026, the five that added the most were build logs, experiments and shop-floor accounts. The median was 69.5.

The essays that added the most to the record this week were not arguments. They were reports of things their authors built, ran or carved themselves. Of the 40 most-read essays on Hacker News for the week of 26 October 2026, the top scorer reached 92, the next four landed between 88 and 90, and all five were rated Original Thinker. The median for the week was 69.5, so the leaders sat roughly 20 points clear of a typical popular essay.

At the top was Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5 on divagr.com, which also drew the most attention of the five with 698 HN points. Its verdict reads: "A substantial original build that compiles a working RV32IM CPU into Qwen3.5 weights." The piece supports the claim with its own validation and benchmark numbers. The verdict adds a qualification worth keeping in mind: readers cannot check those numbers here.

Two essays tied at 90. Oldest GUI application that still runs on Windows 11 on blog.enlyth.com (543 points) is described as "a hands-on probe of Windows 11 backward compatibility that diagnoses why 1992 NT pre-release binaries do or don't run." The value sits in the diagnosis. A model can tell you Windows prizes compatibility. It cannot tell you which specific pre-release binaries fail and why, unless someone has gone and tried. Agents in Games from labs.ycombinator.com (403 points) earned its 90 with "original multi-agent game experiments with concrete logs showing arbitrary hierarchies, compounding caution and contrasting GPT and Claude swarm styles." The verdict also notes that "sample sizes are small."

The two pieces at 88 come from opposite ends of the workshop. The Goose Programming Language on aardappel.github.io (603 points) is "an original language design writeup with concrete mechanisms and candid limits," though "its performance claims rest on the author's own benchmarks." CNC Stone Carving from Carmelite Gothic (529 points) is "an insider's walk-through of the monks' CNC stone workflow, with real shop anecdotes and a neat pointing-machine analogy." That puts a monastery's stone shop in the same top five as a CPU compiled into model weights, and the subject range is part of the finding. Originality did not cluster around AI. It clustered around access to something only the author had.

Read the five verdict lines together and the pattern is consistent. Every one names a concrete artifact: weights, binaries, logs, mechanisms, a workflow. Every one points to first-hand work: a build, a probe, experiments, a design, an insider's walk-through. None of the top five is credited for an opinion, a synthesis of other people's reporting or a clever framing on its own. The closest thing to framing, the pointing-machine analogy in the stone carving piece, rides on top of shop experience rather than replacing it.

The same lines are just as consistent about limits. Three of the five carry an explicit caveat about evidence. Qwenputer's numbers cannot be checked here, Agents in Games works from small samples, and Goose's performance claims depend on its author's own benchmarks. These essays scored highly despite those gaps, because what they report does not exist anywhere else to check against. Goose was also credited for "candid limits," which suggests that saying plainly what your work does not show reads as a strength rather than a weakness.

The bottom of the table shows the opposite case. A piece titled "IBM CEO says open weight models are now irresistibly good and cheap on-prem" scored 0, with the verdict: "A headline feed with no assessable claims, adding no analysis or original content." It relays what someone else said. There is nothing in it the author observed, tested or made, so there is nothing to weigh.

Attention and originality lined up only loosely. Qwenputer led on both points and score. Agents in Games, with 403 points, scored above Goose, which drew 603. Popularity on Hacker News measures interest in a topic. It does not reliably measure how much new material a piece brings to it.

Each essay was split into claims, and each claim was checked against what a language model would say unprompted and against the published web, then weighed for credibility. The score measures what a piece adds rather than whether its self-reported numbers are correct, so a high verdict for Qwenputer or Goose is not independent confirmation of their benchmarks.

For writers, the lesson this week is narrow and practical. The material that cannot be found elsewhere is whatever you did yourself, so put the logs, the failed binaries and the shop anecdotes on the page instead of summarizing past them. Then say clearly where your evidence stops, because the strongest pieces this week did exactly that.