Interview, Fireside Chat, Keynote
How Far Are We From An AI Einstein? - Adam Brown
- Large language models are expected to potentially derive general relativity from Newtonian physics within approximately 10 years, which may represent the final step before human tasks are largely superseded.
- Human intelligence is projected to be fully encompassed by artificial systems around the time general relativity is invented, despite acknowledged disanalogies between current models and human cognition.
- Large language models are predicted to process more examples than any human lifetime allows and develop sophisticated, though potentially non-human, representations of complex geometries.
- Systems are expected to exhibit a significant advantage in knowledge retention compared to humans, accompanied by distinct strengths and weaknesses that will evolve as performance improves.
- The evaluation of large language models is anticipated to become increasingly difficult, potentially requiring PhD-level expertise to create tasks that can distinguish their capabilities.
- Recent developments show models advancing from zero success on graduate-level general activity exams three years ago to essentially mastering these exams within the last couple of months.
- While models may currently meet the standard of graduate courses in general activity, such academic benchmarks remain necessary to reach the level required for meaningful performance evaluation.