this is foundamentally wrong. I've been instructing chat gpt about wargame rules, and it's understanding those rules well enough to run rounds of combat. that is far beyond what this article pretends chatgpt to be (a massive pretrained markov chain) as all that learning is happening after training
heck, you can make chatgpt pretend to be something that isn't, exhibiting creativity that is well beyond previous llm.
> All it knows, all it knows are the statistical relationships of how words appear together
again, you can ask nonsense, and it will not spew nonsense back. it has a cursory understanding of what nonsense is, and will not fall for easy traps (i.e. what year apollo 7 landed on the moon) that would trick straighforward probabilistic models
all in all, this article toned down my respect for the blog down two pegs, if he's so wrong about this topic, how much was he wrong on other topics where I had no understanding and took his word for good?