We’re Still in the Early Innings of AI – AI Field Day 6 Delegate Roundtable

I'm Steven Foskett here at AI Field Day in San Jose. It's January 20, 25. We are here learning about different AI technologies in the enterprise.

We're also having private discussions, and this is one of them, as it stands this week, we've seen the announcement of deep seek, which rose above what I call the mother-in-law test, which is when somebody completely outside our industry says, Hey, did you hear about that tech thing? And it's incredible that a Chinese AI model was able to make that kind of news when there's so much other news happening globally. But the thing that occurs to me is that all of what we've been doing, LLMs, generative ai, we are in the, it's not even fair to say the infancy of ai.

It's not, we're not in the first innings. We're in little league, we're in wiffle ball, we're still learning how to do anything. And all of the achievements that have happened so far, all the money that's been spent, all the electricity that's been burned up to make this thing happen is still just figuring out how, what, where, um, I said the other day that it's like, um, it's like we're still working with Doss and, um, and eventually we're gonna have touchscreens, smartphones and, and, and, and and, and we're not even anywhere close.

So I wanna throw that open to the crew. Uh, maybe Andy, you wanna kick it off? This was your, uh, suggestion as the topic.

Uh, we're still in the early stages. It's we're, we are essentially being the alpha test for AI and the general public is, and the, the while, while we certainly, while we're certainly helping to develop AI as it goes along, I think there's a long road before it becomes like an everyday product. And some of the comparisons we talked about off camera were a couple of the other technologies that have come along over the years, which take 40, 50 years to mature to the point where they're actually useful for everyday people and easy to consume.

And I think that, uh, AI hasn't really reached that point yet. I, um, rather than little league, my, my take on it was that we're in batting practice, so we haven't even started the game yet. And I think that things are going to progress over the years, and we're going and things are going to change.

Look very different five years from now and nobody really knows Exactly. Yeah. How, You know, the figure, the, the challenge is that AI is really 75 years old.

I mean, it was invented in the 1950s. The stuff that they're doing today, 'cause they can do today. 'cause they have the computational capability to do, they're doing today because it's available.

The, the computational, um, uh, of skills, not just skills, but hardware can do what they talked about in the 1950s. So it's, it's old technology. This, the deep learning stuff that, that came out of, uh, you know, resnet and ImageNet and all that stuff.

That was, it is 10 years ago. These things are, are taking off. This is, this is a launch phrase, launch phase.

This is going off and we're not gonna, it's a different world. Tomorrow is a different world than it's today. Yeah.

I think, Ray, you're right on. I think the, the ideas behind AI were just not able to implement is in the older days. I think what's, what's really key for AI today is that the acceleration of innovation.

We have never seen this before. And so as a result of this, you know, like Steven said, it feels like it escaped out of a lab simply because innovation goes so fast. People wanna show what the technology can do without the guardrails, without the security and all that good stuff.

And, and I think the reality is, is that it's gonna accelerate even faster, right? We see many more people now investing a lot of time in solving problems that have been unsolved for decades. And the hourglass was the data, right?

The most significant thing that happened, it's not attention is all you need. It's fa Bailey's work with ImageNet, which allowed Greg Hinton to create and his team to create AlexNet. Anyway, long story short, it's the data.

Stupid. I not told you stupid, but it, it, it's the, the, the inflection change was the mass amount of data that used it. 1947 was the first paper written by McCollum Pit that described the calculus of, and, and the point was, your point is correct.

It's been going on for 70, 80 years. But that inflection point, which was her ability to use Mechanical Turk and read her book, it's an amazing book. You haven't read fa Lily in the World.

We see, and it, anyway, the, the inflection point was she literally took ImageNet mechanical Turk and Mass train the data so that the guys out of the hidden who won Nobel Prize could create AlexNet, which was really the turning point that became, that got acquired by Google. Um, so it's the data, I Think it's a confluence too, of multiple things. It's, it's data, it's hardware that you talked about.

Um, I I think it's also kind of where we are and how we use what we want, um, technology to do for us. Reminds me of could all carry around a Motorola brick pump. Um, but that wasn't something the masses did.

We eventually got to a place where it was in a, a form factor and, uh, in a, uh, maybe a cost factor, but, you know, an iPhone or something that really take it, took it to the next level. We're trying to figure out what that app is, what that thing is, how does generative AI become that? And there's a lot of ideas floating around, you know, is it the, uh, the Zoom phone?

Is it something else? Well, what's, what's the iPhone of ai, generative AI that's gonna really take off? And so that's why we see so much innovation happening all at once.

And just to bring together those two streams of thought, I think that if you go back to that Gibson quote that's overused. The, the future is here, it's just not evenly distributed. Mm-hmm.

And that, that pulls in every device that we're gonna see coming out at every major consumer electronic show is gonna have AI in it. And this is where we are, Bob. So not as the math guy.

Look, not every graph is a hockey stick, and not every graph has one inflection point, right? Um, I think we're gonna look back in 10 years at where we are now and the excitement we have now, and we're gonna be even more excited by what we see. This though, going back to Steven's comment at the beginning about dos, well, you know, more than once over the last couple of days, I wanted to fire up Linux and, and, you know, get my bash uh, cheat sheet out there.

I mean, we, we were really watching to do another metaphor, uh, how this, how the AI sausage is made, right? And with that understanding saying, well, things are gonna change, things are gonna get better and they're gonna deliver a lot more value. But this experimentation is really important.

I guess the experimentation part is what scares me a little bit, right? So we've got these, we've got, um, generative AI and all sorts of other narrow marketing bucket. You put everything in a marketing bucket, um, that can do amazing things.

Um, but the point is, it's, it's all with, with data. It's all with data that's been gathered over the last, you know, 500 years. You know, some of it's digitized, some of it's becoming digitized, but lots of it ha uh, comes from a few people in the world and not a lot of people in the world.

And in fact, uh, generative ai, uh, the, the big players were able to make their applications by stealing all of the world's data and not crediting anybody. So we've gotten a long way to go. I just read a fantastic article on LinkedIn about the Maori language and how they're being so careful even to look at Whisper and to see how they were surprised and shocked to see their language was there and they hadn't been consulted.

And for, um, native people, especially, you know, that are trying to reclaim their language, it's, it's a different, um, project than it is with European people. And, um, a lot of those principles weren't followed. And they're lucky enough to have been at this and digging at it and have the equipment and the, the, the technical people to, to go back in and do things the way they would do it and to keep preserving their language the way they should.

Because language, I've been saying this all day, language and words, construct worlds, and the context of how we do things depends on our language. So now all of a sudden we have these great big algorithms that can slice and dice language and spit out something. It's not necessarily the truth.

And in the context of what the language is even describing ai, we don't have the language down yet. All very good points. I I wanted to actually go back to what Bob had mentioned with the, uh, you know, not every graph is a hockey stick.

Uh, the ML commons presentation yesterday gave us some really good information on that, where they showed a bunch of different performance benchmarks that they had started using and abandoned because they no longer fit the models that were available. And I think that's where we are with this business. Where there, you can't actually come up with a benchmark of figuring out how this stuff works properly.

It, it continues to diverge and go in different directions. And that's another reason why I think that we're, we're in the very early stages of ai. Uh, there's, there's no way to actually come up with a, a way to figure out how to measure the performance of it.

So my prediction is, you know, uh, you've been to Vegas lately, and on the way to the airport you'll see these signs. Have you been injured in an Uber ride? Right?

That, and every time I go back there, every few months, there's some other new legal thing, you know, uh, have you been hurt in, uh, as a worker in hotel, right? My prediction is in another year it's gonna be, has your data been taken without your consent? And w has your privacy been violated in by ai?

Without your knowledge? I thought You were gonna say, have you been injured by an AI agent? Uh, you're not.

Yep. There, there you go. And I'm, and it could definitely, I, if, if I were younger, I would go back to law school, right?

And I would chase that because Just a shout out, legal week will be, I think the second or third week in March. So if you're gonna be in New York City, go to legal week. But just a stitch penis concept.

And another concept that Annie brought up, imagine for a moment if those benchmarks were applied to ethics and empathy. Uhhuh, was that legal weak or legal weed? Sorry, Legal weed is actually Nevada now.

Okay. League Week is in New York City in March. Bless You.

I I think what you're seeing is a Precambrian explosion of models that exist today and, and, and the ML performance and ML common things while those models are going away. 'cause they're old models, nobody's using 'em anymore. You know, it's like they've got newer versions.

So they went from NLP, which was, I don't know, 64 tokens to, you know, GPT-3, which is 32,000 tokens. I mean, it's, it's just the world is coming. I'm done.

I'm done. Don't Get interrupted. It's, it's exploding.

It is exploding. And we're, we are not, we're at just the beginning of this hockey stick, which is a hockey stick that's been going on for 70 years, Bob. Well, uh, you know, there's several ironies here.

One of which is the attitude saying, Hey, I stole this data fair and square for me. Um, right. Well, you know, who, who's the establishment?

Uh, right now? The the other factor, um, in, in, in terms of languages, you know, um, and I don't, or shouldn't just think in English, right? This goes to the language, it goes to cultures and things like this and shouldn't just be translated to English, much less American English.

Hmm. There's so many more things that have to be understood about local cultures. If you think, if you are fluent in a completely different language, let's say from English, you think differently, right?

You express yourself differently. You, you, you express concepts, you synthesize ideas very differently. So it just shows how, how early we, we are with this and frankly, how much more work has to be done to be much more inclusive for this new technology on the global market.

And, and I want to jump in there as well back to the, the premise of this discussion. Um, you know, Bob reminded us of it how early we are, we are stone knives and bearskins here, and what is this thing gonna look like once you take this same technology? I, I've been saying this, you know, it just kind of came to me this month and, and it probably came to other people earlier, but you know, we're trying to train artificial general intelligence based only on sentences and paragraphs and words discovered at random on the internet.

That's not, that's not intelligence. Like it can't know things that aren't appropriately and adequately expressed in language. And, and if it can't, then, if it can't, then, then it can't be intelligent To the point of are we in the early stages?

It's really a frame of reference. And I think that's what Ray and John we're getting at is if you look at the history, we're 70 years into understanding the math behind it, building the math, the mathematical models and all of that. So from a, the foundational technology of how do you build this, it's very old, but there's a whole lot of new stuff going on in terms of computation, power, storage, networking that enables us to do these things that are very old.

So your frame of reference could be, it's really, really young and we're in early stages of war. It's actually fairly mature if you're, if you're a mathematician, a lot of this stuff is very mature and old hat to you, right? It's not new stuff.

But to, to follow on to, uh, Jay's point a little bit, um, the authenticity of the stuff is, is something that we do need to benchmark as well. And something that I've pushed for every time that I've been participated in any AI related discussion is AI has to be auditable. You have to be able to find to, to go back to the source of information.

And I think that's another thing that needs to be built in that we haven't been building into AI so far, is if if you're actually going to provide authoritative answers, you need to have some ability to show what authority they're based on. Yeah. I mean, before I, I, I mentioned the philosophy of ai, but since it came into a little bit on a discussion, if, and, and I'll keep it really short if you're not interested and you don't wanna hear me geek out, but, but if you read a book by Eric Lawson called The Myth of ai, he makes the argument that why we're not gonna get to a GI and, and a lot, and, and it, it delves on the history, which is the difference between deductive inductive and abductive reasoning and abductive reasoning is the argument against a GI.

And there's a lot of literature from Dr. Woods, from Eric Lawson's book Myth of ai. And it, and it really sort of bounds on that abductive reasoning is sort of the, the, the Columbo effect, the thing that the innately human, so I I, I think that the A GII think back to what you said about the history, you need to understand the history.

What was an induct, what was an expert system? What is the symbolic system? What is the subic system?

What's the difference? One's an expert, one's a neural network, and then where does abductive reasoning fall on? And I think there's a lot of uncovering of where we're going to go and how far this things gonna go until people actually start investigating and think about the epistemological effect of what this is.

Yeah, I don't think we, I know I hate that word. Yeah, I use it a lot, but I hate it. Yeah.

That word. I feel we're able to wait for that. Um, because I think that's part of the problem of, of hype.

There's so much AI hype because we don't have the vocabulary to talk about what AI is. Ai all of the terms that you think about with ai, like learning, it teaches, it learns, it, it, it reasons it hallucinates that we have, we have a common English language, meaning for all of those words, they don't apply to computer science, which is what AI is. We really have to have those terms defined for us.

We start need to switch context from our common knowledge as a English speaking society for those terms to our engineer minds for what those terms mean. So we can start looking at this stop anthropomorphizing, um, ai, it's like, it's this living thing and it's gonna turn into a living thing. If we just have enough data and we just train it fast enough and we do all these things, no, it's garbage.

It's always gonna be computer science. And there's so much good things that are happening just with what we're able to do now, that could be even better if we got everybody on the same page thinking about, this is what we're doing and we're going to use these problems. We're gonna think about how to solve these problems instead of racing to a GI or thinking that it's something different.

We need to start being engineers and go back to that. Yeah, interesting. Kinda taking that approach is, I think one path is the race for general ai, right?

And that's, that's causing a lot of innovation to happen. It isn't the only path. I think we're at the beginning of the, of the industrial age of ai.

How do we apply this? How does this change how we work, how we create things, how, you know, manufacturing, whether it's, you know, mental capacity and using data or applying technology to it, or it's physically making things. And that's why we're in this experimentation phase, right?

It's okay, we have seam engines, running belts on conveyors that are doing things in a, in a, uh, lumber mill, but now we can use a, you know, a gas powered engine or we can use something else to drive the process. Now we have even better ways, oh, we've standardized things about how we work. I think that's what we're seeing is all this experimentation about how it's changing what we did by hand before and, and now in today's version of by hand to what does it look like in a world of ai.

You know, I I go back to, uh, a cartoon I saw probably 10 years ago, and it was like, uh, various game checkers, chest go, and, you know, all these things, age ai, ai, ai is super human and it's, I think HEI is an interesting discussion, but in the end it's a question of how much can this, this, this computational entity do for you? Is it, is it, is it something that you can depend on to do things like write reports or do, uh, you know, conversions of data or, or you know, c create, uh, you know, an object list from, from an image? And the answer is yes, yes.

And yes, it could do all that stuff today. Ah, it's, it, it, it, is it a GI who the hell cares? Well, and, and maybe that's, that's a good way to kind of end it right there, is that yes, we're, I think we're all in agreement that we are still in the early stages, maybe not of the mathematics as been, has been pointed out, but of the application of the creation of, of whatever AI is.

It would be great as, as Gina was saying, if we could agree on, on certain terms, it would be great if we didn't use inappropriate anthropomorphic terms for things that aren't actually what they are. Uh, maybe that's foolish. Humans do that all the time.

Um, but at the same time, we have no idea what this is gonna look like in the future. One thing I will say is that the technology that's being developed is almost certainly going to lead to just mind-boggling advances. And once we get past this fixation on ever bigger LLMs, once we start applying that same technology to other problems, you know, as we heard about it, the CES keynote from Nvidia Jensen Wang was talking about building a simulation of the real world and training an AI model in that instead of in language.

That's a really interesting thing. And because the same math works in that context, and, and, and so we'll see where we go, but I think it's way too early to be too scared of what we've, what what we've built or too supportive of what we've built. We've only just started scratching the surface.