Andy and Kevin Surace discuss GENIE's near-term direction. The main challenge isn't core performance but handling thousands of edge cases. Transformer scaling shows diminishing returns, so big capability jumps aren't coming soon. Multimodal advances in vision will shape what GENIE can do next.
So, Kim, when you look out there and you think about advance and GENIE in the future and you go, you're just getting going on this journey now, a year out, we've got a more mature GENIE rolling to hundreds and hundreds of enterprises. You know, there's this great quote, I think it was Bill Gates maybe who said it, that we overstate the impact of technology in the short term and underestimate it in the long term. The problem with that quote today is the long term is now the long term kind of shrunk quick. How do you keep a. With just the speed at which AI is moving? But more importantly, how's that going to show up in Geni? This notion that just the exponential growth in intelligence and performance of AI is staggering, how does that affect GENIE in the next year to two years? Because the GENIE you have now, I'm assuming is text. A year from now.
Yeah, you know, could be. Look, I think you have to look at it this way. We've created a GENIE offering that is so wickedly fast that, that speed and cost are not the limiting factor. Right. They're just not. It's. Compared to humans, it's less than a tenth the cost. Right. Or maybe a hundredth the cost when you get to really big numbers. So you're looking at an incredible reduction in cost and an incredible improvement in speed. And so as models improve, that may reduce the cost of using those models on servers. As GPUs improve, it'll reduce the cost again of the those servers. But in terms of capability, what we're trying to do is widen that aperture to accept more range, a larger range of test cases and a larger range of applications and oddities that we're going to see. And again, those who've been in this business a long time might smartly say what we're trying to deal with is the edge cases. And it turns out they're all edge cases. Yeah, yeah, yeah, that's how it is. So the stuff in the middle turns out to not be a very big hump. It's, it's a tiny little bit and then spread, spread, spread, spread, spread. So, so dealing with the edge cases is a forever problem, and we know that. And again, this is where the startups that they get in this space don't realize that there's a million edge cases, not 12. And they give up the ghost after a year and say this place is impossible. Right? And so we know that we've got to deal with those edge cases in genie now. We've dealt with it across our platform. And we'll continue to do that. I think that will be the big thing. Clearly we're going to do more in our generation of test cases, which we already do in AI script generation. We'll do more about converting that to English, more readable English. There's lots going on in terms of improvement of models, core models, since the transformer hit in 2017, while we have been taking the transformer and applying it to bigger and larger data sets. You know, let's take a trillion tokens in, right? And it started out, huh? I wonder if we can take a million tokens. Wow, we can. How about 10 million tokens? Eventually we're at a trillion tokens. You know, I think what we're going to find, and we have been finding is that the, the actual outcome or output of the transformer is starting to level off, and it's been leveling off over the last year. GPT 5 is not going to be that much better than 4.5. Not. Or 4. Oh, not much better than 4. Not much better than 3.5. Right. And so for our uses, we might get more efficient, but in the end, the outcome has to be the right outcome. Here are the test cases, here's the scripts that implement them. Right? That's it. That is the 100% win. And honestly, the transformer, you know, transformer improvements over the next couple years, changing that much, that's going to impact this. Um, so. Hey, interesting. Just recently this week, GPT4O has implemented a new vision model, a new image model, Image generation model.
Yeah, Correct. Yeah.
And that's better than it used to be based on Dall E3. And anyway, it's better. But the big improvement was this. Up until this week, if you asked almost any of the image generators to generate images that included text, the text was a jumble, right? So generate an image of my book called the Joy Success Factor. And Andy's reading it. It'll. It'll generate that and say, didn't know how to spell in images, right? So the, the new model comes out and it can spell and there's all this information and accurately. So generate a poster that. And, and now the spelling and the data and the, and the, the numbers are correct. Do you know how they did that? That was not an improvement in the transformer. They literally took 1,000 people, right? And did supervised learning to actually ask it things and then tell it which one was correct. Like so spell the following words and tell it which was wrong. So literally, human supervised learning, literally millions of images generated until it now knows how to spell in the image generator. So this wasn't an advancement of the transformer. It was going right back to the way I did AI in the 90s when we developed the first virtual assistant. Human trained humans fixing it. Right. And training it, that is. They went back to exactly what we have done for literally 20, 30, 40 years. So again, not an improvement in the core model, just going back to old thousand person training and we'll fix it. Yeah, that's how they did it, actually.
