Historical biases can impact the output of large language models that are only trained on past data, leading to a lack of forward-looking mechanisms. In the case of Galileo proposing the heliocentric model, a language model trained on historical texts would parrot back the prevalent geocentric beliefs due to their frequency in historical data, showcasing confirmation bias. This highlights the importance of not solely relying on past biases but also embracing new data and perspectives to avoid overlooking valid information in favor of conventional wisdom.
If the Wright Brothers could have used AI to guide their decision making, it's almost certain they would never have gotten off the ground. That's because, points out Teppo Felin of Utah State University and Oxford, all the evidence said human flight was impossible. So how and why did the Wrights persevere? Felin explains that the human ability to ignore existing data and evidence is not only our Achilles heel, but also one of our superpowers. Topics include the problems inherent in modeling our brains after computers, and the value of not only data-driven prediction, but also belief-driven experimentation.