They give access to text-davinci-003 and code-davinci-002 models through their API which are strongly related.

But they have done massive amounts of refinement and continue to do so I believe. But I think the amount the refinement has helped improve may not be as much as you would think. For example I think a big part of the refinement is the guardrails.

I've been noticing Claude(AnthropicAI’s new model), seems to be matching the performance of ChatGPT https://twitter.com/goodside/status/1611622200573829121

I had a chance to use LAMDA a few times. In my experience, LAMDA produces a slightly better outcome (less BS basically) but not a fundamentally different level. I guess MS won't be the only player in the league.

What do you use it for exactly?

Not parent, but basic scripts and programming, debugging, error codes, outlines for emails, reports, dry documents, code translation, brainstorming, explaining code snippets, playing fake text adventure games, creating lists, and on and on.

Anyone saying GPT isn't useful because it's not Wikipedia or right 100% of the time are completely missing the mark and imo, is not able to think outside the box. It's insanely useful.

It's useful (obviously) but not as a 1 to 1 replacement of search on it's own.

That’s my question - is it that it’s a huge advance that’s revolutionary or is it an incremental innovation that can be replicated from current state of the art with some effort.

It can be replicated, it's basically clever fine-tuning of GPT-3 with human assistance. Google has model that can do what ChatGPT can for some time already.