A Peek Behind the Curtain: How A.I. Works

The talk opens by arguing that "artificial intelligence" is too broad a label, since it covers very different tools in the way "computer" covers laptops and an old job title. The speaker narrows the focus to large language models, image and video generators, and reasoning agents, and explains how each one is built.

The middle section follows the pipeline. Training data has to be cleaned, because duplicates, mislabeled records and inconsistent formatting push a model toward poor decision boundaries, and a live demo showed that a partial fix can even reduce performance. Text is split into tokens and mapped into numerical embeddings, with word, character and subword schemes compared on cost and nuance. Training adjusts large numbers of weights, and precision and recall, illustrated with a bowl of marbles, show how results are judged. Temperature and attention shape output, and retrieval tools are presented as a way to curb invented answers. Image generation is shown as gradually turning random noise into a picture, and reinforcement learning is illustrated through reward hacking in a simple game.

The closing discussion covered audience questions on whether to block or sanitize risky prompts, whether AI-generated content should be labeled, and how long prompts should be retained given security and legal exposure. The main takeaway is that a model reflects its data, its reward design and how clearly people state their goals, and that many of these tradeoffs have no single right answer.