OpenAI is expanding their custom model training program. So today on the podcast, I want to dive into what this means, why this is important, and talk about some of the interesting use cases that were seen from this fine-tuned model. There's a bunch of companies that are actually using these fine-tuned models that you probably didn't know about. I personally didn't know about.
There's some very interesting use cases. I'll talk about the features, what's going on. Let's get into the podcast. So the first thing that I want to talk about here is just the fact that, obviously, this is a use case that a lot of people have been using.
This is essentially the custom model program. And it's set up essentially to help companies create generative AI models that are tailored for specific use cases. So you can imagine if a law firm or a health care provider wanted to use chat GPT, but they're like, hey, it doesn't know all this really specific stuff about the health care industry. They would be able to provide extra data, and they would fine-tune a model.
It's essentially chat GPT or GPT-4, but with all this extra industry specific knowledge. And this is what they're helping people to do. So custom models and all this was actually introduced to OpenAI's first developer conference. If you remember, they had their big dev day.
And this essentially allowed businesses to work with OpenAI researchers. So this wasn't like you could do it yourself. You pretty much had to go collaborate. And I actually saw some reports saying that you had to spend a million dollars or something crazy in order to do this.
So it definitely wasn't for everyone, but there was a bunch of people doing it. So since they launched this, they did a big blog and they said, quote, dozens. So I'm assuming, I don't know what, like dozens, like a handful of dozens. I don't know.
There's probably like somewhere between 12 and 100 customers or companies that have specifically done this with them and worked with them on this custom model program. But OpenAI said that they needed to expand the program to quote unquote maximize performance. I'm not 100% sure what maximize performance means in this context. But what I can assume is that this is probably a good way for them to make more money.
So if performance is financial returns, and this is a great thing to do. And of course, for companies, like I'm not saying they're, you know, they're sounding bad here. I'm just saying that's probably their main motivation. So one thing that is interesting, there's a new feature here and it's called assisted fine tuning.
So this is added to the custom model program. But essentially what assisted fine tuning is is, of course, you can go and they kind of have two options. The old option, which was like, OpenAI was going to help you do this and they have a new option where you can actually do this yourself. So it's kind of like self led, you can go bring in your own data, you can go train yourself without having to use OpenAI.
And that is going to be cheaper. And I think it's going to be a really interesting option for a lot of companies. So there's a bunch of different custom training models that are being developed right now on top of OpenAI is kind of base model of GPT-4. And I think this is a great example.
So they had a couple examples in their blog posts that they talked about when OpenAI kind of made this big announcement of these new features. I mean, the biggest feature here really is that like this program already existed, but now you can go and do it yourself without needing to spend a million dollars to access to one of the researchers. So in my opinion, that's huge news. They talked about a couple of different companies that have been using them.
So one is SK Telecom. This is a big telecom company in Korea that I like, I've been hearing about them do a lot of investments and a lot of moves in AI. So this isn't a shocker to me, but they fine tuned GPT-4 for help with a bunch of like specific telecom related Korean conversations and some other things that they were doing. Another famous one that I didn't realize had a partnership with OpenAI on this is Harvey AI.
This is a company that raised millions of dollars earlier last year. And I remember when there's all the hype around Harvey AI because it helps lawyers with legal cases and stuff. And I remember a lot of the hype was like, I'm sure there's way more now, but there was like 15,000 law firms that were on the wait list to use Harvey AI. So I think it's very interesting.
I watched a demo of this actually in action and essentially using Harvey AI, they were able to feed a bunch of specific data in relation to law, the law field. And I think they actually gave it a bunch of cases. So if you know lawyers, essentially what they do is when there's a lawsuit, they're going to look at a bunch of old cases, how they resolved and try to find what the precedent was on an issue. And this is something that Chatter P.T.
notoriously struggled with. There was a lawyer that asked it for some sort of legal precedent on an issue and it completely invented a case that never existed on some airline lawsuit. He submitted the case, the young big trouble because he completely submitted a fabricated case that never happened. It was a hallucination and this was kind of like a big example of some of the issues that could arise with these AI tools.
So that being said, lawyers had a lot of skepticism and that's where Harvey AI was born and why it's important is because to avoid problems like that, I'm sorry, I was thinking I'm like, that would have been like a genius marketing move for Harvey AI to hire the lawyer to just do that and then whatever his consequences are, pay for them so that it can just prove how, why you need them and you can't just use Chatter P. out of the box. But anyways, it's probably not what happened. This is a conspiracy theory.
Anyways, it came up with a really great, if you watch it side by side with Chatter P.T. and with what actually they were able to find to Harvey on, it comes up with really great results. So I saw a demo where they were like asking it for precedent on a specific issue. Chatter P.T.
gave kind of a three paragraph blurb about what you'd expect, right? Like it just kind of a point for whatever it had. And then when Harvey was asked, it was able to use the natural language processing, like the LLM kind of of GPT for both all the data and context, it was able to actually outline four real cases. It can add links to those cases and then it's able to, yeah, essentially outline like actual useful data and a bunch of different questions and things that were being asked in relation to kind of like case law.
So very, very useful and of course not possible just out of the box with GPT for. I think what's interesting here is opening. I said that they think most organizations are going to create their own customized models for their specific industries. This is actually the same thesis I have as I'm building out AI box, which is a no code AI app builder marketplace.
We believe that there's not going to be just, you know, the handful of big players we see today, but there will eventually be thousands of AI models in every industry there, you know, specifically going to doing different things. And so that's why we're building out AI box so that we will have a platform that people can access all of these different models on one account, right? You don't need to make, you know, 20 different accounts to go use all of the different best in case, best in class on AI models. You can go get them all on one account and you can also mix and match them together, link them together, change your problems together.
So that's what I'm building and I'm excited to see that open AI sees the same future that that essentially I'm building for. What is interesting here to me is open AI right now, they're getting close apparently to reaching around $2 billion in annualized revenue and they're planning a $100 billion data center with Microsoft. So right now, I think that these kind of custom model training that they're working on right now is seen as kind of a way to sustain some of their revenue growth while they're also trying to reduce like there's a huge strain on their model serving infrastructure. Like this is it takes a big toll.
So these fine tuning custom models, I think are a lot more efficient. And I think because of that, it could help alleviate opening eyes. Like obviously, this is historical. They're, you know, the compute capacity challenge that they're facing right now.
I think this could actually kind of help to alleviate that. So what's interesting is opening, I actually introduced some new model fine tuning features for GPT 3.5, right? They're free version that anyone can access. And this is this includes a dashboard for model quality comparison, the third party integration support, and they're kind of starting with weights and biases and some tooling enhancements.
Well, I think it's interesting though, is it details about fine tuning GPT four, which was actually made available in early access during dev day. They didn't actually say anything about that. I'm not sure when that's going to come out, but I'll definitely keep you up to date. I think this is a fascinating new development we're seeing out of opening eye.
It solves some problems for them. I think it helps generate extra revenue. But overall, I think this is amazing for the community as companies are going to be able to make some incredible fine tune models on top of what opening eye has. So I'm excited for that.
I'll keep you up to date on whatever amazing innovations we see there. What I would love for you, if you've learned anything in this episode, if this was interesting in any way to you, it would really, really make my day. If you could leave us a review on the podcast, I would be super thrilled to hear your feedback. You can, if you're on Spotify, drop us some stars for an Apple podcast.
Go ahead and leave a comment. Really, really appreciate it. Have an amazing rest of your day.