🏅 Top traders

@firstuserhere Yes.

@PatrickDelaney I'm quite familiar with how language models work having spent a significant fraction of the last few years working on and with them.

There's nothing that particularly stops one from equipping one with simple super powers in response to it emitting a token stream of a given format.

A low-key example:

https://twitter.com/f_j_j_/status/1568453579605954560

A more directly relevant example:

WebGPT exists and is effectively just a LLM hooked up to something that replies to it when it issues one of several commands Search:, Find In Page: Quote: ...

https://openai.com/blog/webgpt/

by opening a page and spewing the content of that page back into the LLM as tokens, and then with enough training data getting it to package things up in a format that effectively leads to citations off the page.

The real question is if they'll have something where they package up this behavior and/or an agent fine-tuned to do this sort of citation in a ChatGPT-like context and expose it to users.

@EdwardKmett All right well I just sold out in response to your comment, because clearly you are defining a language model as, "a language model strapped to an app." That's fine...a lot of people, in fact probably the majority of people paying attention to ChatGPT are conflating the two concepts, so that might actually be a fairer way to resolve the market based upon majority understanding. However, an LLM is different than an app - the blog you listed notes, "GPT-3" is what they refer to as the LLM. WebGPT is something else.

I put together a YouTube video trying to explain the intricacies. https://www.youtube.com/watch?v=whbNCSZb3c8

I appreciate that you are spending a lot of time on LLM's, that's great! It's a fascinating area.

So if you look closer at what that blog posts says:

fine-tuned GPT-3 to more accurately answer open-ended questions using a text-based web browser

So, "fine tuning" is what OpenAI calls, simply put, using text to modify a set of parameters in their overall LLM, GPT-X, to give the language a particular character. So you can have an LLM, and then you can fine-tune it to give different flavors of responses. Training on the other hand, is an incredibly expensive process, in the tens of millions of dollars estimated. Fine tuning is arguably a form of inference.

So what they have done is, strapped an LLM to a web browser, and fine tuned that LLM so it works. This is not the same as releasing a new LLM.

@PatrickDelaney

Fine tuning is arguably a form of inference.

I've never heard fine-tuning described that way in my life. Fine-tuning is still modifying the parameters of the network. The primary attribute distinguishing it from the "main" training phase, (e.g. next-word-prediction pretraining) is the scale on which it's done.

@PatrickDelaney A public release of WebGPT would not suffice. A fine-tuned or scratch trained model by OpenAI version of something in the GPT-3 or later generation trained to do the same sort of task, or to interact with the user ChatGPT style while providing cited references or suggested links that the augmented LLM reads or scrapes live in response to user queries would, however, so long as it met the extra caveat that the fine-tuning or training would have to be done by OpenAI to explicitly narrow the focus to them doing the work and/or packaging it up for a partner.

The reason for wanting this to be an OpenAI thing is that otherwise an end-user could just engineer a prompt, and fine-tune for a while and provide a limited PoC and call this done, and that isn't what I'm looking for.

The fact that ChatGPT's prompt seems to internally refer to itself as "Assistant" with "Browsing: disabled" seems to give indication that they likely have built up the capability in house.

@jonsimon You're right, I was oversimplifying, what you said is more correct.