Vertical Foundation Models

Specialized foundation models for fields like manufacturing, energy and biology that work from few or no examples, matching general models at a fraction of the inference cost.

YC wants research labs building foundation models for specific fields such as manufacturing, energy and biology. They should work with no examples or just a few and give typed decisions with probabilities instead of generated text. The bar is matching or beating general-purpose models with orders of magnitude less inference compute, cheap enough to run everywhere. Google's TimesFM-3, a 330M-parameter forecasting model that needs no task-specific fine-tuning, is the example given.

The hard part: high-quality, verifiable rewards for the field, rights to the data, and a clear accuracy and compute win over existing specialist models.

Read the full request on YC →