According to @DeepLearningAI, researchers introduced the Energy-Based Transformer EBT, which scores a candidate next token by energy and iteratively lowers that energy via gradient steps to verify and select the token, source: DeepLearning.AI on X, Sep 27, 2025. According to @DeepLearningAI, in 44-million-parameter trials on RedPajama-Data-v2, EBT outperformed same-size vanilla transformers on three of four benchmarks, source: DeepLearning.AI on X, Sep 27, 2025. According to @DeepLearningAI, the post links to a summary in The Batch, while the tweet does not specify compute cost, latency, code availability, or release timeline, so cost or speed implications are not provided, source: DeepLearning.AI on X, Sep 27, 2025.