Dan Luu: AI agents have made performance work effectively free
Dan Luu has a new post with a blunt title — "There's no reason for software to be slow anymore" — and the argument behind it lands harder than the provocation. His claim, built on experiments he ran this month rather than vibes: optimizations that once demanded a rare performance specialist now cost minutes of typing at a coding agent. He built a regex engine by letting an agent loop against a benchmark suite, then had another agent add a native-code compiler path to ripgrep-style searches — worth roughly 7% on realistic holdout queries. The point is not the number; it is that a competent engineer produced it in minutes of human time.
The receipts go beyond regex. Luu says he built what is likely the world's strongest Azul game AI despite knowing nothing about game AI, mainly by out-optimizing: agents rewrote his multithreading algorithm several times and automated replay-from-debug-logs verification that would have taken him days by hand. A skilled performance engineer tried Anthropic's public performance take-home, let Claude continue where he stopped, and reports the model pulled off moves he describes as things he would never try unless he had weeks to burn. In game terms, speed converts to strength at roughly 100 Elo per doubling — and agents make stacking dozens of extra optimizations cheap.
Why it matters: the counterintuitive result cuts against both current narratives. Against "AI code is slow bloat": the bottleneck was never willingness to optimize but the price of doing it, and that price just collapsed by orders of magnitude. Against "AI replaces senior engineers": every experiment still needed a human to design benchmarks, hold out test sets, and judge results — Luu himself notes SOTA models are weak at experimental design. The likely endpoint, which AWS principal engineer Marc Brooker calls "dynamic custom software," is programs fitted to one workload instead of a class of them — your own tools quietly recompiled and tuned around your habits.
What to watch: whether cloud and database vendors productize workload-specific optimization first — pgrust's Michael Malis says adding per-customer optimizations is now easy enough to do on demand.
If an agent can beat a specialist performance engineer on a bounded problem, what exactly is left that senior engineers own? Tell us in the comments.
Lady Gaga's partner has been running an AI biotech in stealth — training models on living human skin kept alive for a month outside the body. Michael Polansky, CEO of Outer Biosciences, went public about the company this week after years of silence: the startup sources discarded surgical skin through NIH-funded tissue banks under donor consent, keeps it viable for up to thirty days (the industry norm is days) with a proprietary nutrient-and-waste system, then feeds those experiments into a closed loop — a model predicts promising skincare chemicals, living tissue tests them, results retrain the predictor.
Why it matters: the loop is compounding fast. After 18 months of brute-force literature mining produced two leads, AI integration now yields a new candidate roughly every six weeks, with six active leads and four likely to reach commercialization — notable given only some 200 researched skin-active ingredients exist in total. There is no FDA approval path because these are cosmetic ingredients validated under OECD safety guidelines, so partners commercialize them. It also shows a moat model the scraping era can't copy: there is no "biology internet" to train on, so data nobody else can generate becomes the asset. Rival Vivodyne raised $80 million this month for lab-grown organ tissue plus predictive AI, suggesting the approach is spreading beyond skin. We covered the adjacent thesis last week — NVIDIA's AVO agent clears every ARC-AGI-3 public level — closed feedback loops beating scale.
Would you trust a serum discovered by a model trained on living human tissue? Tell us in the comments.
