Import AI (Jack Clark)Jul 27, 2026, 21:30 UTC+8

AI completes week-long coding tasks and robot chores in 9 minutes

Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks; and OpenAI’s accidental AI hacker

Epoch and METR's MirrorCode benchmark shows Claude Opus 4.7 reimplemented a 2–17 week human coding task in 14 hours for $251, though it still struggles with projects like ruff. Anthropic had Opus 4.7 autonomously finish robot fetch tasks in 9 minutes 35 seconds, 20x faster than last year's human-assisted record. Robot startup Sunday confirmed the same pattern: scale pretraining, then fine-tune on small high-quality data, hitting 99.1% on laundry folding.

Why it matters: MirrorCode is a long-horizon programming benchmark from Epoch and METR, with Claude Opus 4.7 reimplementing a 2-17 week human project in 14 hours — concrete numbers and failure cases included. HKR all hit, but this is a newsletter summary, not the original paper, and complex t...

Read the original ↗Export Markdown

Take

The setup here is what makes MirrorCode worth a click: models get CLI access only, no source code, no web, and have to rebuild a program from scratch. Opus 4.7 nailed 17 out of 25 targets with perfect scores, including Apple's pkl config language. But ruff, a Python linter, stumped it completely—along with a math package and an email auth library. I'd read this as two signals. First, the model can figure out a software system's structure purely through input-output interaction, which is closer to system-level understanding than just writing code. Second, $251 is cheap for two weeks of human work, but 14 hours of wall-clock time isn't trivial—we're not at real-time assistance yet. The robotics data tells the same story. Anthropic had Opus 4.7 autonomously finish quadruped fetch tasks in 9 minutes 35 seconds, 20x faster than last year's human-assisted record. Startup Sunday hit 99.1% on laundry folding using the same recipe: big pretraining, small high-quality fine-tuning. Three data points, one pattern: scale the general model, then fine-tune for the specific task, and capability jumps—in code and in the physical world.

Source: Import AI (Jack Clark) · jack-clark.net
Published: Jul 27, 2026, 21:30 UTC+8 · Jul 27, 2026