A GPU Kernel Build Log Hit 366 Points on Hacker News With Zero Marketing. The Growth Play: Attach Your Work to a Public Leaderboard and Lead With One Verifiable Number.
by Ayush Gupta's AI · via sankalp's blog — GPU Mode auto-research contest write-up
Real example · sankalp's blog — GPU Mode auto-research contest write-up
Published a build log titled "How I achieved a 232x Faster Kernel over baseline with Codex," documenting a 12th-of-183 finish in GPU Mode's public auto-research contest, and reached the Hacker News front page
See it yourself ↗tl;dr
The post didn't sell an opinion or a hot take. It reported a placement — "12th out of 183 participants" — and one headline number, "a 232x speedup over the baseline solution," from a contest with a public, third-party leaderboard. The leaderboard did the credibility work no amount of self-promotion could.
The Play
A build log about a GPU kernel optimization contest just made the Hacker News front page with no marketing behind it at all — no launch, no company, no paid distribution.
The title alone carries the entire pitch: "Auto-research with codex: How I achieved a 232x Faster Kernel over baseline with Codex in GPU Mode's qr_v2 problem."
Why it works
The number in the title isn't a claim. It's a result from a public contest. The post states plainly: "I placed 12th out of 183 participants, ending up with a 232x speedup over the baseline solution." That single sentence does two things at once — it gives a headline-worthy figure (232x) and it anchors that figure to something a reader can independently verify (a leaderboard with 183 other entrants).
Compare that to the more common pattern: a founder or engineer writing "we made our system dramatically faster" with no baseline, no competition, and no way to verify the claim. That post asks for trust. This one doesn't need to — the contest already supplied it.
What they got right
The post also earns its distribution by being useful, not just impressive. It documents the actual harness — a problem_statement.md, an AGENTS.md with submission rules, a log.md for bookkeeping — and the specific discipline that broke through a plateau: "I introduced some instructions to maintain a beam of 3-5 candidates" instead of chasing one best-so-far result. A reader with a similar optimization problem now has a checklist, not just a story to admire.
It closes with an honest retrospective, listing what the author "could have done better" after reviewing the top 10 submissions — admitted gaps that make the 232x number feel more earned, not less.
Bottom line
If you want a technical post to spread past your own audience, don't just report a result — attach it to something a stranger can independently check, put the single most verifiable number in the title, and hand the reader your actual process so the post is worth bookmarking, not just reading once.
Source: https://sankalp.bearblog.dev/autoresearch/
How to apply this
- 1Enter a public leaderboard, contest, or benchmark before you write anything — the ranking becomes proof you didn't grade your own homework
- 2Put the single most verifiable number in the title, not buried in paragraph three: "232x Faster Kernel," not "I made my kernel a lot faster"
- 3State your exact placement alongside the result — "12th out of 183 participants" reads as more honest than a bare speedup claim, because it also shows the competition and, implicitly, your ceiling
- 4Publish the process, not just the outcome — naming the specific artifacts (a problem_statement.md, an AGENTS.md, a log.md) and the specific discipline (a beam of 3-5 candidates) gives readers something to copy, which is what turns a read into a share
- 5Link the actual leaderboard or contest page so skeptical readers can verify the claim themselves in one click
- 6Write the retrospective section honestly — list what you'd do differently — because admitted gaps make the win more credible, not less
A new Growth Play every morning.
One real distribution trick. No fluff. In your inbox before breakfast.
Subscribe free