Automating AI Training

Deep learning derived its success from replacing hand-crafted features with learned ones. We believe training recipes—the data, environments, curricula, and reward signal that AI models train on—are the next frontier. Currently, these recipes are overwhelmingly hand-designed, relying on a mix of ingenuity and guesswork to build useful models.

Hiddenweights is automating the AI training process.

We are focused on the many settings where the bottleneck to widespread AI deployment is not lack of algorithms, or of compute, but of reliable training signal. This challenge is everywhere across the AI stack, and as the availability of compute outpaces the availability of human supervision, we expect its prevalence to increase rapidly.

Tackling this challenge demands a new fundamental research agenda—one that eschews human-designed heuristics and instead optimizes and synthesizes everything: data, environments, curricula, reward signal, and eventually end-to-end training.

Coupling this research agenda with strategic partnerships keeps us grounded in the reality of AI development and deployments. And when training is optimized instead of guessed at, XXX.