It’s a bad take for several reasons. - It’s reductive. The model may be architecturally relatively simple, at least insofar as humans are involved, but the dat

It’s a bad take for several reasons. - It’s reductive. The model may be architecturally relatively simple, at least insofar as humans are involved, but the data that was used train it is anything but. Code is data, data is code. The model is more than its architecture. It’s all the weights and biases within. - It’s qualitatively wrong. The temperature settings change it from a predictive system (auto complete) to a generative one (digital assistant). - It’s quantitatively ignorant. How do you eat an elephant? One bite at a time. For that matter, how do you speak if not one word at a time? It happens that the model generates one token at a time as well, but the million dollar question remains: what chain of reasoning did it have to follow to come up with that token? Just because it‘s only outputting a single token doesn’t mean it’s not “thinking” about the structure of the sentence, paragraph, or the entire document - several tokens ahead, to say the least. Indeed, having been trained on a such a large internet corpus it has the capacity to model entire personalities and all the quirks and neuroses that go on behind the scenes. So, yes, it is a bad take. That being said, it’s still a useful one because it encourages people to think about the models as tools, which is exactly correct given how they’re structured. So I still describe them as such.