Faraday: the AI teammate that actually reads papers

Inherent, spun out by people who worked on DeepMind, released Faraday — an AI agent built to do the tedious part of science: reproducing papers. The lab says it beats Anthropic and OpenAI on that specific task. The point is practical. If an agent can read a methods section, set up the environment, and get the same results, it saves researchers weeks of copy-paste work and failed installs.

The broader picture is what happens next. Researchers already spend a chunk of their time chasing down code that doesn't run. Agents that can reliably replicate papers would shift some of that load from humans to machines. The risk is the same one that shows up with every new AI capability: overconfidence in a system that looks like it works until it quietly doesn't. The claim needs independent replication — and preferably a dataset that tracks how often agents actually get the numbers right, not just how they sound.

Why this matters for us: the people who do the hard, unglamorous work in labs and shops are the first to feel the gap between what AI promises and what it actually delivers.