Adaption Labs: Gradient-Free Continual Learning — Sara Hooker, Adaption
Auto Scientist aims to automate model-training choices across data, alignment, and architecture. The builder-relevant claim is broader recipe search, though frontier training remains compute-heavy and safety stays unresolved.
Hooker estimates **fewer than 5,000 people** can train frontier models at scale. Adaption’s Auto Scientist co-optimizes data and model adaptation across dense and mixture-of-experts architectures, with support planned from day one for **242 languages**.
The practical idea is to encode scarce training know-how into an agent that searches more configurations than one specialist can. Builders evaluating automated training should inspect whether the system controls **data quality and the full adaptation loop**, not only hyperparameters.
Hooker estimates **fewer than 5,000 people** can train frontier models at scale. Adaption’s Auto Scientist co-optimizes data and model adaptation across dense and mixture-of-experts architectures, with support planned from day one for **242 languages**. The practical idea is to encode scarce training know-how into an agent that searches more configurations than one specialist can. Builders evaluating automated training should inspect whether the system controls **data quality and the full adaptation loop**, not only hyperparameters. The talk does not remove the need for large models or substantial compute, and the reported win rates were capped by a stopping rule above 60%. Wider access also expands the safety obligations around powerful model training.
This moves automated model adaptation beyond hyperparameter search toward co-optimizing data and architecture, while confirming that scarce training expertise can be partially encoded rather than eliminated. It strengthens the case for evaluating the whole adaptation loop, but the stopping rule and continuing compute demands narrow claims that automated training broadly democratizes frontier capability.