> Climate change is a real field of science. AI risk is Nostradamic prognostication by people who know more than you.
Any prediction of the future is necessarily based on modeling and extrapolation.
Five years ago AIs couldn't pass a third-grade reading comprehension test. Today they pass in the top 10% of law, medical, and engineering exams for human professionals.
It is absolutely possible to extrapolate from such developments, and doing so is scientific, not "Nostradamic". Many predictions of the potential impact of climate change also include speculative elements, such as societal effects, migration patterns, conflicts, etc., which cannot be modeled or forecast with any real certainty. That doesn't make them unscientific.
It's possible to extrapolate from a horoscope too, it's just not that useful. Let's talk about the massive difference in the extrapolation being done here, climate change vs AI.
In climate change, we are analyzing historical climate data using weather models representing known physical processes. We try to predict the data using these models, and we are only able to do so if we include the forcing from greenhouse gases. From this we can constrain the range of impacts these gases could be having on temperature and forecast likely futures. The forecasts are heavily informed by a thoroughly validated base of prior knowledge, not just drawing lines through a log log plot.
None of this has any counterparts in AI. We don't understand AI systems to anywhere near the level that physics affords understanding of physical systems. We don't even understand them at a Moore's Law level, where you can at least know what engineering innovations are in the pipeline and how far they could plausibly go. Predicting the sophistication of future AI is just Nostradamic prognostication.
Yann LeCun recently gave a presentation arguing that LLMs are a dead end and proposing a completely different approach. His arguments were extremely heuristic and unconvincing, but this at least shows that both sides have bigwigs with unconvincing heuristic arguments.
Any prediction of the future is necessarily based on modeling and extrapolation.
Five years ago AIs couldn't pass a third-grade reading comprehension test. Today they pass in the top 10% of law, medical, and engineering exams for human professionals.
It is absolutely possible to extrapolate from such developments, and doing so is scientific, not "Nostradamic". Many predictions of the potential impact of climate change also include speculative elements, such as societal effects, migration patterns, conflicts, etc., which cannot be modeled or forecast with any real certainty. That doesn't make them unscientific.