110 Unsettling Truths About Artificial Intelligence to Drift Off to

Tonight, we’ll trace how artificial intelligence grew from mathematical ideas about neurons into systems that now shape work, communication, and public life. We’ll begin in 1943 with Warren McCulloch and Walter Pitts’s model of artificial neurons, then meet Alan Turing and his 1950 test for machine intelligence. At the 1956 Dartmouth workshop, researchers gave the field its name and made bold predictions that proved far too optimistic. We’ll follow early programs such as the theorem-proving Logic Theorist and ELIZA, the 1966 chatbot that led people to form emotional attachments despite knowing it was only following simple patterns. From there, we’ll see why symbolic AI struggled with everyday perception, how two AI winters cooled funding, and how Frank Rosenblatt’s perceptron was sidelined before neural networks returned through backpropagation, deep learning, more data, and faster graphics processors. AlexNet’s 2012 ImageNet win marks a turning point; the 2017 transformer architecture then helps set the stage for GPT-2, GPT-3, and ChatGPT. As these tools spread, we’ll consider what they can do and what remains unclear: their opaque decision-making, their tendency to produce convincing falsehoods, and the questions they raise about understanding and consciousness. We’ll look at impacts on jobs, misinformation, deepfakes, privacy, warfare, fairness, energy use, and the concentration of power, alongside emerging rules and research into AI safety and alignment. Finally, we’ll consider the risks of systems helping build future AI, the limits of historical comparisons, and the choices facing researchers, companies, policymakers, and the public. We’ll end with the idea that AI’s future is still being shaped—and that we don’t have to settle every question tonight. We can also follow AI into fields far from chatbots: the AlphaFold database from EMBL-EBI and Google DeepMind gives researchers open access to more than 200 million predicted protein structures. In 2022, OpenAI reported that human raters preferred outputs from a 1.3 billion parameter InstructGPT model over GPT-3’s 175 billion parameter model after the smaller model was tuned with human feedback.