Thomson Reuters Trained Custom AI LLM for $40 Million - OpenAI is DEAD

Eli the Computer Guy (0:00)

Okay, everybody, let's say it all together.

Open AI is dead.

Anyways, how are you doing? You may not know this. You may not know this about me, but in my opinion, I think OpenAI is highly effed. I do not believe there is a bright future for OpenAI, and I come here every day to explain my argument. So one of the things with OpenAI, really the thing with OpenAI, is Sam Altman believes that intelligence should be like a utility. So just like you buy electricity or water or sewage or whatever else, you will go to a single provider or a few different providers. You will get intelligence services from them, and basically it'll come in at a metered rate. And that is Sam Altman's and OpenAI's concept for what the business of intelligence will be going in the future. One of the things that I've argued for a long time is that that's just dumb on the face of it. That's just not how things work in the technology world, right? People want control over their systems. There's architecture design things to think about. There's their security, their safety, there's resiliency. There's all of these things to think about. And the idea that you're just going to have these centralized providers to provide all of intelligence just doesn't make a lot of sense. Beyond that, with Sam Altman, Sam Altman believes basically the more data equals more intelligence for some reason. And remember, with a lot of the stuff that these AI systems have been trained on, they're literally trained on Reddit. That's not a joke. A profit center for Reddit is selling access to all of the communication on Reddit, right? One of the things that Elon Musk really crowed about with XAI purchasing Twitter was that basically Grok would be able to learn about real human communication.

I'm sorry, who the fuck has ever looked at Twitter and went, wow, this is real human communication, right? But that's the thing, right?

They're training their models on all of this garbage. And one of my arguments for a long time has been, is if you really clean up your data, you train your models on clean, sanitized data that targets specific areas you want this particular model or product to solve for, then you're going to end up doing a lot better. Apparently, one of the best LLMs out there for Koreans is a Korean model, a South Korean model that was created, that was only trained on Korean material. So instead of having an LLM that's trained on Russian and French and Swahili and everything else, and all of that comes in, you simply train it on Korean material. So it's really good on Korean matters, right? Because it's highly focused there. And that's where it's kind of interesting right now, is we see Thomson Reuters come out, and basically, they've come up with their own model. They've trained it for a mere $40 million, and they're using it for a real world product. So they have this product over here called CoCouncil, fiduciary grade AI, built for high stakes professional work. And I think this is interesting. Like one of the things we talk about with using these AI systems is what about the hallucinations and everything else? One of the problems you run into is that most of these AI systems, they are not enterprise class in the way we used to think of enterprise class. Like you would not have gone into a 10,000 person company and expected to see Linksys switches back in the day, right? They would not have been called enterprise class equipment. So Linksys, I don't know, do you kids know Linksys anymore? Anyway, I remember Linksys and Yahoo and Myspace. Anyways, there is this consumer and Soho network equipment manufacturer called Linksys. They actually did create some pretty big switches at some point. They started creating 48 port switches that were actually pretty good for small business environments. Again, business environments where you had less than 100 people, their switches were actually pretty good and at a relatively decent price point. The thing is, they're Linksys switches. They weren't Cisco switches, they weren't Juniper switches, they weren't HP switches. They were Linksys. They were good for what they did, and that was about it.

You might have to reboot the switch once a week or something like that, and that was just the thing. So again, if you went into a corporate environment that had 10,000 employees and you went into one of their networking closets and you saw Linksys switches in their networking closet, you would have thought that was a bit weird. You're like, huh, that's weird.

I do not think I want to work for your company anyways, right? But what's weird is the artificial intelligence world. If we look at these AI systems, we look at these LLMs, basically most of these LLMs, if you think about it from an enterprise class capability, is they're much like Linksys. They're kind of like that quirky thing. Like, yes, they do their job, but they're a bit quirky. You might got to reboot them once a week. They probably have the life expectancy of a carton of eggs, so you're probably gonna have to replace them every year or so. Like, do they do the job? Yes.