Scale did most of the visible work
The models got better because they got bigger. GPT-2 was large by the standards before it; GPT-3 reached 175 billion parameters when comparable language models were in the hundreds of millions to a couple of billion. That is not an increment, and the outputs changed accordingly.
What size buys is a better output distribution: a more accurate sense of what token plausibly comes next given everything before it. Better, and still imperfect.
The imperfection compounds, and that is structural
Generation samples a token from the distribution and feeds it back in. But the network was never trained on its own output. It was trained on text people wrote, and it is now being asked to continue text it wrote itself.
So a small error does not stay small. Each sampled token moves the input slightly further from anything the model saw in training, and the next prediction is made from that slightly stranger place.
Why a raw model cannot follow an instruction
Ask GPT-3 to summarise an article and it will produce something that is not a summary. That looks like a failure and it is the training objective working exactly as specified.
On the open internet you almost never encounter an article, followed by a request to summarise it, followed by a good summary. That pattern is not in the distribution, so the model is not good at it. It is not refusing; it has genuinely never been shown the shape of what you are asking for.
This is the gap between a language model and something you can talk to, and it is not closed by scale. It is closed by training the model on examples of being instructed.
What that means when you are choosing one
The differences that matter between providers are mostly in this second stage and in the data behind it, and none of them publish it. So evaluating a model on parameter count is evaluating the part that stopped being the interesting part.
It also means a small permissively-licensed model, tuned on the right instruction data, performs far better than its size suggests. That is the part worth knowing if you are building rather than buying.