[1hr Talk] Intro to Large Language Models

A large language model is, at its core, just two things: a file of model parameters and the code that runs them. Taking Llama 2 70B as an example, the parameters are the neural network's weights, while the inference code can be as little as a few lines of C and runs locally with no internet connection. The model works by training on a 'lossy compression' of vast amounts of internet text, learning to predict the next word in a sequence — and in doing so mathematically captures a surprising amount of knowledge about the world.

Building an LLM happens in two stages. Pretraining burns millions of dollars of GPU time to process terabytes of text, compressing internet-scale knowledge into the weights. Fine-tuning then uses a smaller, high-quality set of question-answer conversations — usually human-labeled — to turn that raw 'document generator' into a genuinely useful assistant. A third stage, reinforcement learning from human feedback (RLHF), pushes usability and alignment further still.

Modern LLMs are no longer simple text generators; they increasingly act like the kernel of an operating system that orchestrates external tools. By calling a browser to search, a Python interpreter to compute, and multimodal abilities to handle vision and audio, the model can tackle complex, composite tasks. The analogy to an OS is deliberate — the context window serves as working memory, much like process management, dramatically increasing the model's practical usefulness.

The frontier of LLM progress lies in giving models a human-like 'system-2' mode of thinking — letting them reason and iterate over a tree of possibilities on hard problems rather than answering on pure intuition. Inspired by AlphaGo, a major open question is how a language model might automatically self-improve within a narrow domain. Through repeated training iterations, models may eventually surpass human performance on specific tasks and spawn a rich ecosystem of customized variants.

As LLMs spread, security threats grow more prominent. The talk breaks down jailbreak attacks (using role-play or encodings to slip past safety limits), prompt injection (hijacking the model with hidden instructions), and data poisoning (planting trigger words via the training data). These attacks show that LLM security is the same cat-and-mouse game traditional software faces, and underscore how important evaluation and defense research are in this new computing paradigm.