Skip to main content
Andrei Moise

On this page

  1. The search loop — branch → measure → prune → deepen
  2. Two lessons
  3. References
All notes
2 min readAgentic Workflows

Beating Anthropic's Original Performance Take-Home Challenge

1,361 cycles, down from a 147,734-cycle baseline, with no background in kernel optimization, SIMD, or compiler work.

Earlier this year, I took Anthropic's Original Performance Take-Home with no background in kernel optimization, SIMD, or compiler work.

Result: 1,361 cycles, down from a 147,734-cycle baseline (108.5×), passing all 9 published thresholds.

How? I asked questions, used agentic search, switched models, and applied different techniques.

The search loop — branch → measure → prune → deepen

I use a tree-of-thoughts-like exploration over candidate optimization. I had another idea inspired by self-consistency that evolved into exploration, where I used parallel sessions with the same prompt, then compared investigations, branched on candidates, measured cycles, and pruned failures.

I switched models when I plateaued (Claude ↔ Codex ↔ external cgpt analysis). Different training data and post-training methods create different blind spots.

I also used agentic search to find what others did, how and why.

I caught him cheating several times, even though I instructed him not to change the tests/ folder and to use tests/submission_tests.py for verification.

Two lessons

Don't avoid problems where you lack expertise; AI can compensate. Be careful what you prune; don't outsource judgment, problem-solving, and understanding all at once.

We can outsource some of it, some of the time, not all of it at once. The skills we cut are our own.

References

Reach out

ai@andreimoise.com