Research · Do this yourself

Run an auto-research loop with Codex to speed up a GPU kernel.

✅ He placed 12th of 183 entrants with a 232x speedup over the baseline.

Based on tosh on Hacker News (opens in new tab)

Who it’s for

Anyone who has to read a lot and turn it into a clear answer: analysts, students, founders, scientists.

Adapt it to your industry

  • Real estate: compare neighborhoods using public data
  • Restaurant: summarize what reviews say about you and nearby competitors
  • Consultant: build a short brief on a client’s market before the first call

Step by step

  1. Open ChatGPT, Claude or Grok and start a new chat (turn on web search or “deep research” if your plan has it).
  2. Write down the exact question you need answered and what a good answer looks like (a table, a one-page memo, a list).
  3. Paste the prompt below and add your question.
  4. Ask it to list every source it used with links, then open 2–3 of them to check they say what it claims.
  5. Ask for the result in the format you need, for example a table you can paste into Google Sheets.
  6. Save the prompt so you can re-run the same research next week.

Ready-to-copy prompt

Loop: propose one kernel change, implement it, run the benchmark and correctness test, keep it if faster. Log every attempt and try a different approach after 3 failures.

Start with AI

Opens the AI with this guide and the prompt filled in.

Open in ChatGPT (opens in new tab)Open in Claude (opens in new tab)Open in Grok (opens in new tab)
Cost
$0–20/month (free tiers or one AI plan)Our estimate
Time
2–6 weeksOur estimate

How to make money from it

Pricing idea: Offer short research briefs at a fixed price per report (for example $100–500), or a monthly retainer for a weekly update.

Finding your first 10 customers

  • Make one free sample report on a topic your target customers care about and share it.
  • Send it directly to 20 people in that field with a one-line note.
  • Ask anyone who replies what they would pay for a monthly version.

Ideas, not promises. Results vary.

Details: source, numbers and tags

What they did

Sankalp set up Codex to try kernel ideas, benchmark them and keep the winners, and added idea diversity when it got stuck.

How it went

He placed 12th of 183 entrants with a 232x speedup over the baseline.

Worth knowing

He says learning enough of the math to ask better questions mattered.

Difficulty
AdvancedOur rating
Time to build
Not reportedOur estimate
Monthly cost
Not reported
Revenue
Not reportedWe only show revenue the source states
Engagement
457 HN points93 comments · fetched Oct 15, 2026

Posted · Added Oct 16, 2026

Original builder recipe

  • Codex
  • A benchmark harness
  • GPU
  1. Define the question and success criteria.
  2. Let agents search, run or summarize in parallel.
  3. Verify key results yourself.
  4. Write up findings with sources.

Tags

Agent Codex
Tool CUDA
Industry Software AI

Read the original

More ideas like this

The Daily 5.

Five real AI agent use cases each morning, each with a recipe and a source. Plus a Sunday roundup. Free.