Google DeepMind has published research on JEST (joint example selection), a new AI training method to reduce computing costs and energy consumption. The researchers claim that JEST allows a 13x improvement in performance and a 10x improvement in power efficiency compared to other methods.
JEST differs from traditional AI model training techniques by focusing on entire batches of data rather than individual data points. It creates a smaller AI model to grade data quality from high-quality sources, ranking the batches by quality; it then compares this grading with a larger, lower-quality set and then trains a large model using the most suitable batches identified by the smaller model.
Analyst QuickTake: If proven to be a robust method for LLM training, DeepMind will likely look to integrate the training method for future iterations of its “AlphaFold 3” foundation model (announced in May 2024 ), geared toward molecular structure prediction.
By using this site, you agree to allow SPEEDA Edge and our partners to use cookies for analytics and personalization. Visit our privacy policy for more information about our data collection practices.