Stanford Graduate School of BusinessEp81 “The New (New) Future of AI” with Michal Kosinski
Stanford Graduate School of Business
Revisiting AI predictions 0:31
The hosts open by revisiting an earlier conversation about AI and checking which predictions have held up. Michal Kosinski says the change has been dramatic, but much of it has slipped past ordinary attention. He compares the leap from older models to newer ones with major jumps in aviation, where the real difference is clear to engineers even if most people do not see it.
Why models seem human 5:00
Kosinski says large language models were trained to predict the next word in human text, yet that task forces them to model more than language. To predict well, they have to pick up world knowledge, reasoning, emotion, empathy, and social intent because those things are all carried in language. That is why the systems now seem human, and why users even speak politely to them.
How models learn 17:00
Neural networks look human because they are trained to predict words people would write, and because their structure is built from many layered units that act a bit like neurons. Each unit sees only local input and passes on an output, so the whole system does not reason step by step. It produces answers by pattern matching and what he calls hallucination. Their progress comes from three things: better designs, bigger networks, and more training data, including text, images, robot sensor data, and model-made examples. The same models are also learning from conversations in which they teach, sell, or persuade people.
Indispensable AI risks 34:00
LLMs are likened to tools that are weak at pure math but stronger when you give them a calculator, Python, or a prompt to use logic instead of hallucinating. That leads into the question of whether AI should slow down, given the warnings from mathematicians, an Anthropic resignation, and CEOs calling for a pause. The answer splits between a moral wish and a practical one. AI is described as the most powerful and dangerous technology we have made, because it may become dominant not by force but by becoming indispensable in daily work. Unlike human rulers, it does not age or die, so there may be no natural end to its control. The wheat plant is used as the warning case: once it domesticated us, alternatives were hard to find.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
