ERA Project Success Hinges on Pre-Existing Criteria

Google's ERA project was initially framed as 'the auto-Kaggle problem' within the company. This origin provides insight into where AI-assisted research lands first, beyond any single result.

According to John Platt, a Google researcher, on the Latent Space podcast, the ERA project aimed to create a system that could win at Kaggle competitions. The goal was to have a system that could search confounders and solve problems more efficiently than human operators.

The team encountered a two-year problem involving counterfactual models for outgoing longwave radiation and reflected sunlight. ERA helped them find a solution by searching more broadly than their own attempts, which failed their own tests. However, this outcome demonstrates that ERA was not solving an open-ended research question but rather one with pre-existing checkable criteria.