Frontier AI Models are Worthless - OpenAI is DEAD

Eli (0:00)

Good news, good news, good news. It appears that CNBC must be watching Eli the Computer Guy commentary videos because they're finally coming around to my point of view. AI safety debate meets reality at Dreamforce as business leaders say last year's models are, wait for it, good enough, good enough. I'm not sure who has been screaming about this for a couple of years now. Who out there has been using the example of smartphones and saying at one point a number of years ago, smartphones got good enough, so most people really don't care about the latest version of the smartphone. Really, the only reason they replace their smartphones anymore is because the old ones, the screen got cracked or they're not updating the operating system or the battery is bad or whatever else, right? You're most likely updating your smartphone for any number of reasons other than what they were talking about at the last Apple event. And I've been discussing this again when we talk about LLMs and AI and these frontier labs and their frontier models. Now, one of the things I've been arguing for a long time now, is that the models frankly got good enough somewhere between a year to two years ago, right? 3.5 Turbo was shockingly good, but probably not quite good enough. But then you get to 4 and 4, to be clear, if you're actually trying to solve problems, if you have actually defined a problem that LLMs can solve, 4 was actually probably pretty good in 99% of the situations. And in the technology world, once you find a solution to your problem, once a piece of technology goes legacy, motherfucker, you never go back, you don't upgrade unless you have to. And so I've been talking about this for a while, and people like, Eli's so stupid. Doesn't he read tech journalism? Doesn't he see what the greatest minds of 22-year-old polysci majors think about the world of AI? Doesn't he understand he is not in step with these people? Yeah, it's like, yeah, because I've been a technology professional for 30 years. 30 years, over 30 years. I'm over 50 now. What happened, boys and girls? I'm over 50 Anyways, and the way that I talk about the technology, the reason I give the opinions that I do is because there are ways that technology works. If you are a technology professional, you understand how this profession works. And if you're a snot-nosed 22-year-old poly-sci major, you know how to write dumbass shit for the internet. There's a reason why one of us gets paid $12.50 an hour, and another of us gets paid a shit ton more than that. But anyways, again, that's why I've been trying to talk about this whole AI world thing, because so much of this is just fucking asininely stupid, because as I've said before, once you implement a technology, motherfucker, you do not touch it. You do not look at it. What is the number one rule as a technology professional? You touch it, you own it.

Oh, you were in the same room when the server crashed? Obviously, you did something.

No.

So that's the thing.

If you deploy whatever model it is, and it is actually solving your problems, what do you do? Something the fuck else. Because that's how we work as technology professionals. But anyways, 22-year-old journalists wouldn't grasp this concept. That's why I even talked about that. I just did an interview thing with Isaac today. And yeah, we're talking about the frontier models again. And again, one of the things that I've been talking about, because that's one of the things you hear in tech journalism, and in the AI-pilled world, right? Eli, what about the latest open AI model, Astra, right? Or Anthropic, right? Anthropic is about to IPO. So they're about to drop a new flagship model. It's like, how does this change things? As I, look, ask me in a year. Ask me in a fucking year. What do you mean? As I, because it's the intelligence layer, it's the intelligence layer, right? We talk about an AI stack, the stack of technologies, a lot of stuff that goes into that stack. And the intelligence component is one single layer. So basically, what has to happen now, is you have to have CIOs, CTOs, managers, team leads, whatever else, look at this new frontier model, determine what the frontier model can, more importantly, can't do, and how reliable it is to do one of those things.

Then they have to look at all the problems that are in front of them, and figure out what Astra or Anthropix's new model could actually solve for. They then have to determine how it could be implemented, right? How does everything actually connect together? Then you have to have your dev team go off, and actually write all the code, and put everything together. Then you have to go through a testing routine, and then you have to go through your deployments, and all of that kind of stuff. And yeah, look, Astra, I will be able to give you an opinion about Astra, probably about August of 2027, because that's how the technology world works.