Stop saying chatGPT doesn’t actually know anything

 A common refrain is to say GPT-3 and ChatGPT “don’t actually know anything” because they’re only programmed to predict the next word. That’s like an alien observing that the human brain is only programmed to maximize long-term happiness and hence doesn’t actually know anything.

The correct way to determine whether something “knows stuff” isn’t to theorize about what is or isn’t possible based on how it works, but rather, to scientifically test its knowledge. And GPT-3 technology has already been scientifically proven to perform at the state of the art level for common sense benchmarks such as Q and A, reading comprehension, and SAT questions. It’s nowhere near the human level, but often outperforms previous AI’s specifically engineered for those tasks, even though its only directive was to “predict the next word”. In short, it turns out that when something is good enough at “predicting the next word” it starts to gain emergent logic and reasoning that conventional wisdom would’ve held was “impossible” for something just programmed to predict the next word.

Instead of asserting something can’t possibly know stuff if it’s only programmed to predict the next word, we should be amazed that something only programmed to predict the next word can already know so much stuff.