--- title: LLM foundations & interpretability description: Understanding LLM architecture and internal mechanisms behind their behavior. source: ManticsCore kind: intent_session url: https://manticscore.com/clip/s-e98f32baa683610c --- # LLM foundations & interpretability Understanding LLM architecture and internal mechanisms behind their behavior. # LLM foundations & interpretability 2 sources, one thread Understanding LLM architecture and internal mechanisms behind their behavior. # How does ChatGPT ‘think’? Psychology and neuroscience crack open AI large language models Researchers are striving to reverse-engineer artificial intelligence and scan the 'brains' of LLMs to see what they are doing, how and why. Large language models like ChatGPT produce remarkably fluent text, but even their creators don't fully understand how they generate outputs. Scientists from psychology, neuroscience, and machine learning are borrowing techniques from cognitive science and neuroscience to probe these AI systems — treating them almost like an alien intelligence whose 'brain' needs scanning to reveal its inner workings, reasoning patterns, and potential biases or failures. ## Explore next - [Explore next](https://api.manticscore.com/clip/s-e98f32baa683610c/go?u=https%3A%2F%2Fwww.nature.com%2Farticles%2Fd41586-024-01314-y) — Interpretability research in machine learning Cognitive science approaches to studying AI Ethical and safety implications of opaque AI systems Comparisons between LLM reasoning and human cognition # Attention Is All You Need Introducing the Transformer: a network architecture based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration, with the best performing models also connecting the encoder and decoder through an attention mechanism. This paper proposes the Transformer, a new simple network architecture based solely on attention mechanisms. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. **Key Results** - Achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over existing best results including ensembles by over 2 BLEU - Establishes a new single-model state-of-the-art BLEU score of 41.8 on the WMT 2014 English-to-French translation task - Trained for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature - Model dispenses with recurrence and convolutions entirely, relying solely on attention mechanisms Full interactive version: https://manticscore.com/clip/s-e98f32baa683610c