Top AI Product

419,000 microseconds down to 1,805. That’s a batched FP32 QR factorization on a B200, and no human wrote the winning kernel.

What this actually is

Not a product — a reproducible agent setup. Developer Sankalp entered GPU Mode’s auto-research contest (batched square compact-Householder QR, matching torch.geqrf) and pointed GPT-5.5 Codex’s /goal long-loop mode at it. Fourteen days, 1,500+ submissions, 12th place out of 183. Claude Pro sat on the side explaining the math. Total tooling bill: $200 ChatGPT Pro, $20 Claude, plus Modal credits for GPU and NCU profiling.

The part worth copying

Most people run agents as single-point hill climbing: one incumbent, keep nudging. Sankalp ran beam search — 3 to 5 candidate idea families alive at once, so a dead end costs one branch, not the run. Memory lived in plain files: AGENTS.md for standing rules, problem_statement.md, and attempt logs so the agent stopped re-trying ideas it had already killed.

Why it matters

Same Codex CLI everyone has. The 232x came from loop design. Any problem with an editable file and a measurable number can take this shape.

You Might Also Like

• Google Search Information Agents Turn 1 Billion ai Mode Users Into Agent Operators
• Anthropic Multi Agent Turf war Research 3 Claude Agents 1 Codebase Self Replicating Malware
• Hiveterm Bets on the Multi Agent Workspace Claude Codex and Gemini in one Terminal
• Openai Codex Claude md Auto Import Makes Switching From Claude Code a two Click Move
• Agency Agents by Msitarzewski 232 Specialist Agents That Turn Claude Code Into a Full Company