AI Alliance: Exploring OpenAI's Unexpected Collaboration with Scale AI

0:00OpenAI, of course, is the tech firm behind the popular chatbot ChatGPT, but it is essentially extending its partnership horizons to third party vendors to simplify the fine tuning of its AI models for enterprise enterprise level developers for enterprise. have shortcomings. I think what's interesting here is, you know, CheckGPT launched back at the beginning of, or the, or, you know, the end of last year, and they've known that fine tuning of these models was a very hot topic. They have a ton of money from Microsoft and they probably could have tried to spin up some sort of company or solution to this problem.

0:45But I think what's really interesting here is they've decided and they've opted to kind of focus more on their core competencies. I mean, really creating GPT-4 and then GPT-5 and that kind of core thing to the point where I think they've like sort of neglected their image generating or like it a few people know this but opening I actually has like an audio or like a music generating AI model that they created quite a while back that is essentially like no one really updated and nothing's really happening with it they have dolly 2 that generates images which in my opinion is sorely lacking compared to competitors like stability or you know stable diffusion and a lot of other ones Mid-Journey obviously is crushing in this department, building on top of a stable diffusion from Stability AI.

1:28So I think, you know, really OpenAI I think is kind of buckling down and focusing on ChatGPT as like the big focus here. And they're okay to work with other people, which I think is really a good sign for the industry. So specifically the company announced a partnership with Scale AI, which is a San Francisco-based startup specializing in data labeling. So the collaboration aims to blend the fine-tuning functionalities of Scale AI's data engine with OpenAI's GPT 3.5 text-generating model. So for those that are keeping track, GPT 3.5 is the immediate predecessor to OpenAI's flagship GPT-4 model.

2:05GPT-4 is obviously the best. It's what everyone uses now if you have premium. Previous to that, you had 3.5. But here's the difference, right? the reason that a lot of these integrations are going into the 3.5 is because 3.5 has 3.5 turbo which is a super super fast model so while it's slightly worse arguably than gpt4 it's way faster and so by fine-tuning essentially you're able to get the quality of gpt4 but at the speed of gpt 3.5 which is some really big improvements that people are really interested in so fine-tuning an AI model essentially means adapting its functionalities to accomplish specific tasks effectively.

2:45So for instance, a business can refine a model to align with their unique brand voice and tone or to respond to customer inquiries in a specific language, right? Telling it that always responds in German or French. And then clients of scale will also be able to leverage its data engine to prepare and, you know, quote unquote, enhance their data. So they can then use this customized data to fine-tune OpenAI's GPT 3.5 model, even adding features that allow the model to cite or reference the proprietary data in its generated responses. This is really, really important, right? Like let's say they fine-tune it to have a ton of their company's data to help their customer service reps, right?

3:26They want this thing to be specifically a fine-tuned model that is answering all the questions people might have about their customer service. So in that data, they can have references to their terms of service and a bunch of other things. And when the model spits out a response, if it's been fine-tuned, it's going to say, you know, like, this is our policy regarding this as per, you know, our terms of conditions. And then they can cite exactly where someone could go and actually find that. So I think that's going to be really, really useful. It's also going to help combat like hallucination because you could say, if it says anything, you could say, can you cite where that like actually came from?

4:00Like, are you making this up? Or is there a specific citation I can click on that's going to bring me to the exact location of the information I'm searching for. So I think to really ensure high quality output, scale promises that the fine-tuned GPT 3.5 models are going to be under, they're essentially going to undergo a review by human experts. So these assessments will ensure that the models not only meet, but exceed performance and safety standards. Scale AI's CEO, Alexander Wang, expressed a lot of enthusiasm about this collaboration in an email interview that he recently did. He said, quote, Scale is OpenAI's preferred fine-tuning partner to deploy the most performative models, and we look forward to expanding our relationship.

4:42OpenAI has chosen to partner with Scale because of our expertise and proven track record of delivering high-quality models with real business impact for our customers. So I think this partnership announcement comes kind of on the heels of OpenAI's revelation that it's going to be allowing customers to fine-tune a lighter version of GPT 3.5. So that's GPT 3.5 turbo, right? The super fast version and using its fine tuning API. So while OpenAI is hinted at developing an in-house tuning tool, it remains unclear whether this forthcoming tool is going to compliment Scale AI's platform or serve as a different function.

5:17I feel like it's going to compliment and they're going to keep Scale AI around. Otherwise, I'm not sure why they've, you know, come up a big partnership if it's going to become redundant very soon. I guess that will be a good indicator of what type of company OpenAI is, how cutthroat they are. But in any case, Brad Lightcap, he's a CEO of OpenAI, he kind of emphasized the value of the Scale AI partnership for their customers. He said, quote, we want to serve customers wherever they are. We're excited to work with great partners like Scale that help us serve the need for our customers as they put AI to work in their organization.

5:50So according to Lightcap, the new alliance is not only going to allow companies to fine-tune OpenAI models as they have before, but also enable them to take advantage of skills, data expertise, and also enterprise AI capabilities. So as for the future, I think Lightcap remained kind of coy about whether OpenAI plans to extend fine-tuning capabilities to other models like GPT-4, which is expected to, you know, gain those kind of functionalities this fall. I think actually specifically, even if he's being coy and like not responding to that, Sam Altman tweeted and said that coming this fall, the GPT-4 is going to get fine-tuning.

6:30So I think this partnership either way signifies a really important step in evolving the fine-tuning landscape for AI, aiming to streamline the process for enterprise and really ensuring that they can customize AI models more efficiently and also making sure that these are more effective.