AI Predictions for 2023 – The AI Times EP 2

Welcome everybody. This is AI Times by IO and techstrong. My name is Lee Baker and I'm the general secretary for the AI infrastructure Alliance or Aya as we call Oakley call it and alongside my valuable panelists.

We're going to be bringing you a session today on AI predictions between 2023. After years of slow and steady progress AI made some massive leaks in 2022. It turned into a breakout year for industrialized AI with generative AI models like chat GPT Dali majority in stable diffusion bursting on to the scene and seeing astonishingly rapid adoption and progress.

Now that we've entered 2023. Well the New Year have in store for us all. Is the macroeconomic climate going to stunt growth or will we have another year of blazingly fast progress in generative AI continual learning Robotics and more?

Hopefully we've got the panel to answer some of those questions for you. Mike Hi guys, I'm Mike brizard Chief content officer for techstrong group. And we've got our panelists here with us right now.

So we're going to jump into our first question, which is going to go to Daniel who's also the managing director for the AI infrastructure Alliance, Daniel. Welcome the show. Thanks so much for having me.

All right, so To Lee's Point 2022 is kind of the breakout year in the sense that these models emerged in. They got a lot of play shall we say it's not clear to me what people are using them for this just yet. So as you look at 2023, what's your sense of?

You know, are we going to see a lot more uses of AI and what do you think that's gonna look like in 2023? Well, I think we're already seeing a lot of uses of AI and I think we're gonna we're gonna see more and more but I think it was a breakout year because not because they just everyone came up with something new last year, but because we've been building upon it for a very long period of time and a lot of pieces came together from a giant data to having enough chips to the right algorithm to the right funding, right? And so, you know, I can tell you right away how people are using it from a generally I standpoint.

It was interesting at the eye. I had some I had some awesome. The gravity designers who do work for all the event and you know, they did work before the holidays so that there wasn't one of the best work that they've ever done.

They're usually awesome. But it was you know, right before the holidays. I'm guessing they kind of mailed it in so when I got back from the holiday took a look at I said, you know what these aren't great.

So I went into being able to Fusion in mid journey and I started iterating on some of the ideas that I had and then I fired those back over to designers that they did iteration in in their Tools Plus used all their their design sense for lettering and where things should fit properly on a page and kind of moved it together and came back with some brilliant design why I always think of these things at 10 time a lot of folks tend to think of these things is like, you know every day the media is wasting ink with like all the jobs will be gone. This is total down Absolute Total number. It's really it's been hard right?

It's humans working together and that's what we're supposed to be and then at the same time, you know, Maria who runs operations for the Iowa working. But and she speaks like four or five languages. She's really human beings.

She knows logically how to make something flow. So what she used to do in the past. She'd write the newsletter and then I do all the editing because I've been a writer for you know, 20 years in this case.

She did the right up. She knew what you want to say. Obviously, it's not as dynamic as it could be for a native, you know language speaker.

So she put in the chat DBT instead. Correct this to make it much more Dynamic I had to change like for work, but the question becomes like are people using this correctly, like all kinds of people are interacting with these tools. Like it's a human being on the other side that thing correct.

It is a it is a model that you program with words and we're gonna see more and more use cases like that as people start to realize how you interact with these things and what the huge funding that's starting to come in generate AI, you know open and getting 10 billion other folks starting to pour money in to it. We're going to see more of these use cases. I saw an artist do a turnarounds, right?

So they were you know, you have a number of artists have been upset about it, but they're not really looking at how tool is going to work. It's not the type of prompt and it's going to do everything. That's also not how it's going to work.

I saw a professional artist train up a fine-tuned model where they do the fun part, which is designed the character and then they trained it to do turnaround which is where you have to draw it in five or six different positions. Totally boring work totally tedious work and they train the mile to do their turnarounds. That's how you're gonna see you're gonna see these kind of workflows where the AI artists Ai and folks are sort of working together do things more rapid iteration these kinds of things and it's gonna be very exciting.

We're going to see more money and more more models kind of pour into it and then as more people get to it, it's gonna it's gonna refine all these things. So they get smarter and smart. It's gonna be very exciting here.

You raise interesting points to kind of use cases and something examples. Um, William as CEO of lightning AI how far can we take generative AI? Um Beyond kind of some of the examples that are we ready for text to the video text to 3D text to music.

Will I be crazy my own Symphonies and and it's so what's the quality? To be that's going to be prototypes some or am I going to be kind of publishing this for the world large? Yeah, I mean, so it's a research question right?

It's kind of like how generalist the approach to like the fusion that powers all of this, right? So far seems to be pretty pretty general, which is cool. So we're kind of starting to see evidence of that in audio.

I think like it's dancers. Yes, but it's a matter of time. Right?

Will it be this year for everything? Probably not. Will it be for many things within the next five years probably like I think that's pretty reasonable quality.

It's always going to follow pattern which is like the first time it comes out. It's gonna be okay and then that whole year is going to be about getting it to be really fantastic. Remember like everyone it's kind of a worse stable diffusion today us of like last summer but this is a paper that was published two three years ago.

It's already been a while right? It's been in research for a long time open ask what's kind of the first one to apply it and then stability and so on the big thing that changed this year specifically was just that it was like built in the open source, that's the majority of it. Right?

And then now you have people using a lot of it as well, but the idea is go back a long time. So so probably the next thing so I already out in papers they're in conferences right now. You just won't see them in the public for another few years and they'll follow that pattern.

But yeah, I'm pretty excited. I think it'll it'll definitely apply to more than just this you definitely will see like text to video probably this year. I would assume Texas video will happen as well.

Like I don't think we're that far from that text to audio started to kind of happen last year. I think this year will happen more. Yeah.

Right. Hey Mitch, you are a principal analyst for text strong research and you're also CTO for the tech strong group which publishes devops and I'm asking that question because we're going to be talking about ml Ops and from your perspective is ml Ops kind of just a set of best practices that are similar to the best practices. We've seen from devops just applied to machine learning models.

Do you think this will be the year where we see more structure and how these models are created because it sounds like we're going to have I don't know tens of thousands of these things. So how will they be built? But it said the astric Ops phenomena, you know out of the devops world.

This is actually positive things that talks about applying some of the principles about how software being created now and whether you say AI Ops or ml Ops, typically what we mean by that is more than just playing machine learning algorithms into software or an AI model into software is it's it's the platform an environment for managing. Those things is you have to manage AI model like you might think of it like a data model and a database but much more sophisticated and their generations of that and learning that are applied to it and different things that you use those things for but there's always a data platform that can be the foundation or the driver for that. So what I see happening in soccer, we're just again shooting a report around observability and Cloud native applications.

And how AI is showing up called AI driven observability and it actually if people are in our survey in our data collection and talking with users of tools of observability tools and operations groups are seeing some changes that are impacted by AI specifically around kind of the toil part of it is managing alerts and quantity of them, but more importantly is as our soccer gets complex. It could really understanding the insights that they need to understand connect all the dots because we're talking about internal apps and SAS applications and things that are all all time. So that's just one Avenue.

It's interesting because you have so many advances happening in parallel, you know, whether we're talking about generative AI types of things we're talking about machine learning and applications, you know for I guess three years ago, we would have talked about autonomous vehicles and vision processing and image processing and all the things that are happening though. Those are all moving forward in parallel and I think that's part of the reasons why suddenly AI has gone from a vendor tagline to actually showing up in products and in the field as well as the research that's happening. No, dang it.

They want to add to that. Do you see a need for some sort of Define set of best practices for ML Ops and and how do we kind of bring the models and the software development processes together? Yeah, look, I think it there's always got to be best practice than ml Ops I think in my office is changing like dramatically right now.

I think it's going partially through a contraction and we're seeing I talked a bit about this the age of Industrial I'd say we see and I'll some books are you know didn't get enough traction and you know, some folks are are just not gonna make it through for this year and I think Part of the reason is that the perspective is wrong? Like what we thought was that everybody was going to have a hundred ml engineers and everybody was going to be doing data science. Right?

What they actually forgot is that the machine learning is actually really really hard and so what you're you're just never going to see the average company training up a gigantic Foundation model or something like that. It's just it's too intensive in terms of compute in terms of data in terms of people, right? It's one thing to do some xgboost and turn prediction.

All these kinds of things that incredible value. But I think a lot of these tools were kind of tuned towards trying to do everything right the fitness and data scientists doing some kind of like, you know from data managing and some basic predictions and then and also it should work for you know training on, you know, 2,000 gpus and and distributing it to billions of people for VA inference. But you know any 100 those things are going to split right you're gonna see things that's kind of specialize in one or the other and and not really, you know, there's gonna be I think a lot of these companies are actually going to end up serving the foundation model companies and what I'm calling age driven businesses, and those are those are businesses that have ai at the core of their business model, right?

And so if you think about Boeing they have ai at five different levels of the business, but their business, they'll just making airplanes, right and they could technically do it without it. But if you're generating photo realistic people or fashion catalogs online and you are able to take, you know, all kinds of generate a whole bunch of people and show, you know, a hundred thousand pieces of clothing and every possible per Nation you could never do that by hiring enough people and taking the pictures you'll still hire people who do the big photo, you know, the big hero photo shoot, but that's an AI driven business. I think a lot of those companies can start to serve these AI driven businesses that have ai as a first and they're gonna that those practices are going to start to tune those things.

I don't think the vast majority of people are actually never gonna interact with where they thought they were going to act with it. So we thought we're gonna be interacting at the inject data clean data, you know do experiments train tomorrow deploy a model observe a model that's still gonna happen at some level but most people are gonna move up the stack kind of like when you get Apache PHP my Sequel and then you move up to Wordpress. Yeah, you could make a word a website well before WordPress, but once you move up to Wordpress, it makes it easier for people to do that.

Then you have another layer of abstraction with drag and drop, you know, kind of editors for a terrible designer like myself. I'm a good designer but not not a great not a great color. So I think most people are gonna interact with models at the fine-tuning stage adding a little bit of code extra code adding a little glue code adding, you know, embedding go ahead and talking to the API and building kind of chains of things like that.

That's where they're gonna start to interact with it that kind of higher level. I think we're gonna see it definitive changes in the industry that these things kind of start to play out and again, not tomorrow, but over the next, you know five to ten years. Is this all starts to come together and we're going to see more specialization and most people are gonna abstract up instead of working at the lowest level.

That's like working in assembler. Most people. Are you still need assembler?

And then you still need c or building operating systems, but most people are going to code in python or you know, you rust something like that in a higher level of abstraction. It's gonna be the same thing with with ML Ops in my opinion as well. It's interesting what you're saying that that's so if we're seeing like the fuel of this is is AI driven businesses and Mitch is a testing to the maturity ongoing maturity of ml loss and MLS best practices William.

What do you think? We can expect that it? Techniques tools of AI that might either reemerge or re-establish themselves within the next year.

Yeah, um great question. So first I just want to say a really agree with them with Daniel saying right, I think like that's kind of been the The Stance of our company where we started kind of trying to do a lot and then we realized that you needed to have their individual pieces for individual workflows. Right but I think like that that kind of leads into this because it means that you the techniques that are going to come out again.

I think it's gonna be like you're gonna treat a lot of these models of some bettings which is kind of what we're doing today and you're still gonna use the other kind of traditional models xgbo's classifiers, whatever right now generative AI is a little bit different in that you're not like classifying something you're synthesizing something. So a lot of the Smith that's what really apply but there's probably like a collection of methods that haven't been introduced yet that will be introduced and next years that probably and we kind of seeing it already which is like more fine-tuning stuff, right? Generate something but like how do I make that thing be a little bit more tailor to what I'm doing?

So I think a lot of the research is going to go into like how do you find tune faster and better right into that point? Like do you need to train a foundation model from scratch every time probably not right? You should probably need a better fine-tuning technique.

And so I think a lot of people are going to start getting into those kind of things. So I think our researcher today if I were starting my phc all over I would probably start looking into fine tuning techniques and probably would be done with my first part of it in the next year and then there'll be like some crazy other stuff that will come out afterwards right because the field move so fast, but yeah if you want what no, sorry, I'll Mitch curious about your perspective one way to look at this that people talk about ml Ops but also an evolution to monologues which is kind of taking it from the data scientists research, you know, people hiring data scientists within their businesses. To more of an Enterprise platform or Enterprise capability to now place the governance around models the management evolution of those that kind of managed by the broader kind of it or CIO kind of organization is that you think that's gonna continue to be an evolution of kind of a broader management of ML and AI is this model upside idea or is that fad?

That will pass Um, yeah, so I think this whole model up thing and to some extent I'm alive and all of this like, I've always been a bearish on this because I just think it's like a it's like a temporary thing right? It's like we took ideas from software engineering. We applied them.

We said we needed them but like Daniel said like are you really coding and CSS Ross CSS or jQuery right now? Not really no one does this right like you're not coding and assembly so we moved up to the some to the to the stack. I think that the problem in the space is that the people who get into machine learning tend to be pretty good Engineers or like usually researchers that one understand every detail about everything and they just won't let go of the lower level obstruction.

So you just want to know everything but it's like do you have to learn everything about a car before you go buy it? Not really like I have to be a physicist or know anything about fluid dynamics, right? So at some point the world has to make the sleep to say hey, you know what the generation before solve that well that's build up right?

Because I think the real value now is applying this to Industry and going forward. So I think all of the stuff like we're All ops comes into I guess really it's I would say more financial services where like you really want to you really want to understand what happened to something and auditing right? So if your thing is subject to like regulatory agency and they need to understand why you did something then all of that is going to come into play right but less so about like am I use I don't know about using this model or this.

There's just like a lot of hype terms in my opinion, right and it's I think it's gonna be a lot simpler than that and it's gonna be very tailor to each person really like at the end of the day that aib AI right? So so Daniel following up on that. Is there going to be kind of a two-tier play here where we're gonna have data scientists who build the initial model and then they expose that out maybe through some sort of interface or an API or whatever it may be and then that's how we democratize it and that's how we're the average individuals probably gonna interact with it the way you described it.

And is that how the whole thing is gonna play out or You know ultimately will we democratize this to the point where I'll just talk to some sort of AI machine to create an AI? Model for my personal use and I'll just describe it verbally in and we'll magically happen. I mean so that it's Turtles all the way down.

Look I think there's a lot of turtles all the way down in AI I argue that from the earlier thing that model Ops and all those things eventually just become kind of a subset of it. Right? Like we understand like how to run certain things in production.

We need auditing. We need to know where these things are. Okay, we need to know whether they're scaling there's different things like training and we haven't thought about before right that are important and and you know serving things on gpus or different too.

But what I I absolutely so I think I think that becomes kind of those those terms become a subset now in terms of The the data science that we talked to for instance stability. We're on the CIO, you know us and open Ai and you know anthropic and all these other foundational companies. They're all competing for the top data scientists this point.

It's like there's essentially a huge, you know, there. There's a huge amount of Need for like tremendous talent and and these talents are incredibly Unique Individuals, right? I mean, they get paid instrument amount of money because they not only can they just look at the papers like they can understand the Paper Understand them the math but that's not enough either for the truly creative ones.

You have to be able to combine your own ideas to be able to combine other ideas from different domains, right and this truly unique data scientist. So the average company's not gonna be hiring those those people they're not gonna be able for them. And those are the ones that like you said are going to create these kind of Large Scale Models.

They get more more generalized. I don't like term AI I think we get there at some point in time in the future, but I like generalize they are in other words models that become more and more generalized for across the subset of domains. A cat intelligence right the cats Never Gonna blue screen.

It's going to be able to jump and hide in boxes and eat and do like a subset of domain. So you might have you know robots that are able to like, you know, clean that, you know, clean the house and do dishes and whatever and you know, the neural net is able to kind of like continually learn not forget the things that it already knows learn new things. Right?

So you're gonna see these and then be trained quickly to sort of zero shot Learners, you're gonna see the kind of thing and because of that you're right. The average person is going to interact with it in a unique way and one of the things I'd like, You know about being a stability is I I came from a long line of open source for instance and at red hat and what's interesting is when when great Engineers regular like Engineers coders an average creative folks and all these folks kind of get their hands on something. There's another level creativity happens.

So there's at one level if we only end up with AI is hiding behind kind of an API that we can only interact with in the way that like they tell us to do it. That's kind of that's kind of the kind of sexual Society because some of the things when I look at some of the real creativity is I thought people you know, for instance training it up on you know, all kinds of like, you know, wave f***, you know, wave files that indicate music, right and they fine tune to model like that. They said well music is just you know, just these weights because I train it up on that and all the sudden instead of having to create a whole music but they're like creating one that has voices and guitar background and everything and no longer trying to mix it together.

Somebody just thought of that right somebody who is an engineer and a creative person. I look at some of the fine-tune models that are coming out of the Of the space where people are taking like 10 different models and saying 15% of this one and 10% of this one and and all of a sudden now, I'm getting photorealistic humans or or like, you know fantasy characters and you know cyberpunk or whatever perfectly with hands that are awesome and and great poses 85% of the time. So again, it doesn't even need to be super generalist.

Right? It's like in that creativity comes out of it when people are doing it individual level and saying wait a minute. Um, you know, I'm Hyundai.

I just want to generate trucks. I'm just gonna find you to generate these trucks and it's gonna do that almost perfectly every time so we're gonna see more more creativity. I've seen it used for synthetic brain scans and things like that something that the original designers never think about.

So the more these kind of things get out and into the world and traditional coders and traditional engineers and just smart creative, you know, internet thinkers get their hands on it. They're gonna take it in New Direction and they're also going to be better at compacting the model refining it making it, you know, run on smaller and smaller memory running it at the edge. That's not something the data scientists are ever gonna think about the original folks are ever going to think about but the people who are running it on like, you know, trying to get it to run on the desktop GPU or on your on your phone or whatever they're gonna have that knowledge to do it.

So I think there's a lot of exciting things and that's where most people are going to interact with AI and that's where it really starts to get actually just embedded into lots of applications and exciting and then, you know, we forget what it is that you know, that that AI it's no more like well, I'm consuming. Yeah, it's like no I just Of the phone and I ask it to make a recipe and it does that and I no longer think about it there because it's not problem. All right.

So even though I can't hear tune you're saying that I might win a Grammy someday. So that's good. You might they'll find you in your voice with like better pitch.

I think they already did that, you know auto tune back in the day but that you know, but now maybe even gets easier and we just kind of visualize you in and then turn you into a 3D model and put you as a hologram next to Tupac up there, you know be awesome start working here now Mike. Yeah. Well as we'll wait with bated rest for my state You album let me yeah, let me take what Dan said there about kind of collaboration and kind of how more than become shared the more original the outcome.

We're all we're all kind of Engineers and Andes. Yes, he is. So that's great.

But what about the the CEO of a large Enterprise? You know, how does he Foster the type of duration that Dan is talking about to generate original ideas to ultimately create profit for his organization William. What's what's your take on that?

How should a how should a company in 2023 engage that is is not one of Dan's AI driven businesses. So I think you know it's a lot of what we do is we help companies kind of like apply and bring it together right? We're so kind of the at the level where we power a lot of these models rights ability.

I for example was producing our tools like I'm sorry several diffusion like I touch lighting and all that stuff, right so that we see we were we have the luxury of seeing across many companies since understanding like how they work and I think that the main Trend that I see with companies that are not like AI first and AI first meaning like like a civility where they have their own researchers, I would say like probably it's this is the year of like software Engineers. We're like, all you have to do is just know how to code and be creative. So I would open it up to other people that are not like status or ml people in the company, like people who are like experts in your domain right who are like the ones who are creating literally the artists we're building the things if you're like a furniture company who's designing that stuff, right?

So I would open it up to those people and basically give them like very simple way. Interacting with this like uis, right and like little abs and things so they can kind of play with and generate things. I think that's really how you're going to be able to Leverage The creativity.

I just you know again, I'm a researcher I focus in AI research. I don't I don't do you know I guess it there is creative work, but I'm not like painting here all day, right? So I'm probably not gonna be the person who's gonna be able to come up and tell you like the best way to like Leverage your particular business assets for this one particular use but partner me with someone like one of your creatives and we'll be able to do that.

Right? So that's what I would encourage. All the CEOs to do is it's like step out of your engineering domain and actually get other people in there because they're the ones who have the idea.

They just they haven't been able to do things with them before right like, you know, remember the time where you could now, I don't know Squarespace probably right? We're like suddenly everyone could have a website now you son explosion of books, right? So, I think that that's really the key.

It's make it super super that simple and allow people to to build with your things. Right and I think that's largely what we focus on as a company. I'm giving companies the tools so that they can build such apps so that they're non-engineers can actually leverage systems like this as well, right?

So does that mean decentralized AI or do you have a hundred experiments going on throughout the business? Um, I mean look, I think you will start so they're so requirements first. You must have data you must there's like some stuff you must have done beforehand, right?

So make sure you have that first. Like I think if you try to jump the gun a little bit too much it'll be harder. Although with fine-tuning might be easier.

Now this like you you may not need a lot of these requirements but no, I wouldn't just like go and like try a hundred things just because it works like it's like my advisors said during the my PhD like I was at Facebook research and now you have all the Computing in the world. What do you do? Well, you don't go try a million experiments.

You don't do that, right? You just you're smart about what you're gonna choose and you try a few things. So I would still argue like be try some pocs test a few two or three things out but make sure they're from diverse groups across the company show some success first and then go and experiment more because like you're gonna want to understand like where's to kind of take an off right?

It's just probably with recession coming. It's probably not the years to like sit here and like do a hundred thousand experiments just to see what happens. Right?

So I would be a little bit conscious of that as well William. It's In the beginning of the year and everybody gets their crystal ball out, but pretty sure that somebody will declare that 2023 is the year of AI but this has been a long time coming. So as you look back, you know, what do you think were some of the seminal moments that led to all of this that maybe people don't appreciate or because we've been at this for years.

It's not just something that came up overnight, but I think a lot of people are surprised by it but in your experience, you know, what do you look at and go gee that was a breakthrough technology. We just didn't recognize it at the time. Um, yeah, so I'll start with kind of an obvious one.

So 2018 Transformers per paper came out right that's basically talked about what Transformers were and that was like fall. I remember I like started my PhD and then so it was like pre-World the world before Transformers and the world after Transformers Transformers is what GPT is based off, right? So fall GPT, I am Transformers comes out suddenly you can now do transfer learning right?

You can take you can do you can learn something in one data set and generalize said to a different one that's not something it could be before with like text right? So in research, it's completely changes that the game then then open AI takes that model and then they scale it up. Right and that's scaling is what we know us now GPT.

So gpt123 that really proved to research and the world that like you can actually just throw more that I'm more computer. You don't have to actually fundamentally change the mask right Transformers paper fundamentally change the math then all we've done since then. At all, we've done a lot more but like the main contributions are really been scaling of this.

How do you just grow these models get more data train them for longer, right? And that becomes GPT 1 2 3 and so on right along the same timelines. I didn't follow the deficient work too closely, but I think it really the paper started coming out around 2018, right?

It was like a while back then they start to bring it in into like Dali then you start to see it the thing that we were using before that was Gans, right? So there was like a lot of hype around Gans for a while because they could generate things but the they were super blurry they went great. So the second we start using the fusion we're like how interesting this works a lot more.

So I think that when the world switch to diffusion that was a big seminal moment there and then probably the third one from I'm really limiting to like last three years because I've been many other some in our moments and AI before that was when we open source when we had a version of say well deficient that was up in Source, right that really changed the game for people as well. And that was this summer and that's I think what you're saying with the explosion today specifically, right? So I think probably the next probably I would assume the next big explosions like this are going to be when we have a pretty good like we've already done so whisper, right?

So we've done voice to text we've done text to image, which is great text to videos gonna come next. We don't have anything yet for really good like text to like speech right? Like there's not like a great system for that yet.

I think that'll be the next thing that happens. There's some inklings that is probably already happening probably will happen this year. And then I just want to understand like can you just keep scaling GPT forever?

Like literally just give it more data like the math hasn't changed right so that from a research for fact that it's interesting. I mean, Maybe I'm a little bit more more like research heavy. But like what I what I'd like to see is actually can we get better results with smaller models?

Like I don't actually think going to like a trillion parameters just really the goal. Like what if you could go down to two billion parameters or 100 million and still get a really good model this unlocks the ability to deploy this models on the edge to put them on headphones to put them on cell phones, like all the stuff right when this models are massive, which is really hard to actually productize. So yeah, I think probably this year will be a good year just like last year was I guess and then look they're all of this work Transformers Builds on a lot of research for many years, right?

So there were many seminal moments. I think attention mechanism was a big one in the research World obviously convolutional networks viaes. Like there's a lot of stuff that has really enabled us to get here.

Right? So it this is a just a combination of like 30 40 Years of research for sure maybe more. All right, cool day you want to add to that and anything and what's in your crystal ball?

Yeah, look, I think I mean you I think you get you know text image and then take in an image to 3D, right? It takes the 3D and these kinds you get this whole workflow pipeline kind of happening at once. I think you hit on a number of different things that have been happening.

I do think we get to trillion parameter models and I do think they're going to be quite interesting because you start to get these emerging properties out of them the other interesting things you start getting more and more chips you get faster and faster chip, right? So it's like we're already starting to see a competition. Now, you've got that third risk building gigantic wait for size Chips To Train You Know training primer models it'll experimental.

But even if they like, you know, get bought out nobody's gonna let that check disappear because now it's like now you don't need to go over the network with infiniband. You're just kind of going over on the chip with like, you know hundreds of millions, of course on there, right? And then you have you have everyone in there in their world kind of getting into these like these new chip designs.

They even have Intel now coming out with the Gowdy too architecture, which has a different architecture than me 100s. They will be day 100s. I'm not sure where the age 100's come into it, but now you're starting to see more.

Years, you're going to see faster and Pastor chips. We're gonna commoditize all these months like it's gonna get there's always gonna be a bleeding edge of like super powerful models that can only be run on the top of the line things. But as the years go by the chip, you know, whatever the future of Moore's Law is right?

It's you know, whether that exist or not the chip are going to get better faster more diverse It's Gonna Get Easier to run those models than compress those models, right? So they start to they kind of start to trickle down and you're not going to start to It's the interesting thing is the combination of technique the chaining of these techniques, right? And so it's like you have you know, anthropic doing the Constitutional AI for instance and then saying okay cool.

We're gonna give it these principles and then we're going to have the AI you have conversations and then score itself on whether it fits those and then do like automated reinforcement learning. So that's a combination of techniques in automation or you know, when I use the automatic 111 automatic 11 11 interface to do stuff they've kind of Open Source communities done amazing stuff kind of jamming all the stuff together. So I always using a model to kind of generate, you know posters for my for my for my workout room.

I wanted to have like old school like propaganda style posters of like, you know, people working out or whatever as I was trying all these different models to make it work. There's a there's a there's a faith scan in there. So you can check the box you face scan to restore the face and one of the things that's when it's scaled up, you know, it moves the detail which way sometimes you get the mangled hands or the face doesn't look right because it's it when you know, it's fine when an arm is small but When you like scale it up, you can figure out the general piece of but like the hand maybe not that fate.

Well, they have a faith Gan that just restores the basically looks perfect. And then another upscale or it changed to that right? So it's the chaining of these things that starts to make these things look incredible and and then you know, we're gonna we're gonna have more breakthrough techniques.

We're gonna have more sort of zero shot Learners, you're gonna get more of these sort of general, you know, these sort of generalized models when you look at what Google's done with their robots in their office, which we can now like open doors clean the office, you know do windows. What do they do? They just jammed Lambda into that they already had the the robots and they just want that smart that they jammed Lambda into it which by scaling it up it had these emergent properties and like chain and freezing logic all the stuff.

I also think and one of the things work experimenting with is Is I think you're going to get like people who go all in on like a neural symbolic logic like they're gonna go They're gonna go find the best math and like logic data set. They're gonna train the heck out of the model on that thing and then they're gonna like free, you know, they're gonna consolidate those weights and freeze them and then they're gonna train up the gigantic model and then all of a sudden when you get like maybe new emergent properties, I think you're gonna see continual learning research start to happen more and more too where you're able to say. Look we did this one thing if I teach it to, you know, recognize dog, that's and then teach it to recognize cats that forgets how to recognized dogs.

You're gonna start to see it do that. And then you're gonna get to a point where these Mountain these True Foundation models become a moat that you that you can't really be in other words. Like if I if I'm trained on a trillion, you know, the images over time not just fine tuning just fine tuning eventually destroy the model over time.

If you go too far with it continually learning what allow you to keep adding stuff to it all the sudden you get these sort of smarter and smarter model for the time, but I think really the chaining of things exciting and then kind of doubling back lastly to the You were talking about you know, where did where did Enterprises consume it? I think Enterprises consume it from the ai-driven businesses from Fast services and things like that, like people who build like what I'm calling what we're calling kind of the tier 2 companies and not not because they're lesser but just because they're most likely going to take the foundation models embed them into an awesome interface or something that does something very specific for business drug Discovery or like generating music or doing turnarounds for for and they're gonna get the audience for that. Right?

They're gonna like, you know, they're gonna have the end audience for that. Then some of them will level up to generating their own models or fine-tuning their models or but that audience is gonna be super important. So you're gonna see all these companies kind of developing the tool sets.

The Enterprises are able to utilize In their environment if they wait a minute, I'm now augmenting my process of materials design or like I'm augmenting all of my all of my non-native speakers. I'm gonna be able to like edit their you know, edit the stuff so that they have a tool that allows them to speak and like easier in it in another language, etc. Etc.

These kinds of things just gonna start, you know, something's gonna come into analyze all the legal documents really quickly go wait a minute, you're gonna have to fight in 20 different jurisdictions based on all the MS, you know, all the things he's kind without realizing it you should try to consolidate these right. So you're gonna see more and more tools built in that utilize this and more more people are going to just be serving that Enterprise at that level and that can be really really exciting. And again, it's not always it's not all this year.

No, but the speeds are already there and what we talked about earlier is, you know, this feed and this feed and this feed and this feed and somebody comes along goes. Wait a minute. If we put the airplane together just like this it flies, right but the propeller and the engine and the wings and the materials all that stuff.

There it's the theme it's just when it comes together is the question. That's when those exciting moments are and when we really start to expand as the dieting when the technology. So so we've got generative AI we've got improvements in computes be that new chips or Quantum Computing.

We haven't forbid we've got AI driven businesses who are kind of crafting these these models and these interfaces Mitch. What does this mean for the heartland of the AI infrastructure Alliance the the independent software vendors who are helping build this infrastructure. What does it mean for those companies who typically are talking to Enterprise to help build out their infrastructure.

How do you think their message will change and how will generate today AI impacts the the ml software ecosystem? Let me answer it in two parts. It's a great question Lee at the CEO the Border level the sea level in an organization.

One of the reasons one of the big reasons why they fund research either externally or inside their own organization, but especially externally and literally quoting you kind of what I've been told which is I'm paying you so I don't get surprised either because I lost an opportunity because my competition took advantage of something I didn't know about your job is to make sure I don't get surprised and where there's an opportunity to help me determine if that's something we can leverage in our business. That's how business people look at. This is is we don't want to get our butts kicked in Market because we were sleeping and but we also don't want to miss an opportunity, but we can't live on the bleeding edge on everything either right, you know their business.

They're not researchers. So I think so the second part. I want to answer this from there's a saying we kind of use as there's a Tipping Point of Technology where we kind of emerge out of talking about it and we don't talk about it anymore because just certain the fabric how we do things.

I'm not saying researchers. Don't talk about it. Of course they do but in the general population, so and it's one of the reasons why there are all kind of specialized or novel or different applications that happen same thing happened in security, whether it's whether it was blockchain rightly.

That's the magical technology that's gonna change the industry and has made some changes but not everything. It's gonna have same thing. We're talking about generative AI we're talking about natural language and translation into different languages.

This one novel application of that to me where it shows up in the Enterprise is from all the people taking you to me course Sarah class. I'm using ml algorithms which are really largely. Data analytic kind of algorithms in their python code or in their react framework or those kind of applications.

That's where the everyday developer software engineer is going to it's going to show up there. So I think that's the that's the translation of the transference that we see. It's all happening parallel and it's gonna be one thing that suddenly makes AI everywhere.

It's gonna be lots of different applications of it. The one thing about this unique about AI though is and it was sharing with you. I started my journey.

May I quite early kind of back in the list but prologue days and the whole idea was mimicking human thinking and it's that still shows up in our thinking today about AI doing what a human could do artistic kinds of things creative kinds of things which are what you're all very interesting. There's also the I just want to throw a lot of data at this and do unsupervised or supervised Al It's against it to do specific kinds of applications. So there's there's different is of that spectrum and I think it's more of kind of the robot that's doing your your work for you.

But in a business application, right? What can I do with all this data? What kind of insights what kind of Novel uses of that data that are going to help me deliver better customer experience and new product gain markets chair that kind of thing.

That's the money end of what turns that part of the crank and the research dollars or what turns, you know gets those ideas started and flowing in that evolution. All right, guys, you only got a few minutes left. So I'm just want to kind of end with a quick short question a little prediction from everybody going once around William will start with you.

They said the wise man is able to expect the unexpected. So what's that unexpected thing that we should expect the unexpected thing that we will yeah, I think we're not gonna displace all these jobs. Like we said earlier in this call, right?

I think everyone's like super scared of this but I think it's just gonna augment a lot of people and they're gonna be better. So maybe what we can expect this normal non-experts whatever their craft this actually being able to operate at like a closer expert level, right? So kind of like, I don't know normalizing just skill sets that are distributed across people.

So like your best artist, I don't know your best painter and you're kind of worse painter and a team may actually be able to kind of like get a little bit closer to each other now, right same for writing same for these kind of things. right, Daniel I mean, there's always like an unobtaining and that's the most interesting thing. So whenever I think about the future, it's like a Monte Carlo and also it's like these are like the strong branches but it's based on like what we know now, right?

So like there was the cartoon that I just tweeted from 1923. It showed the the cartoon Dynamo, you know generating like, you know cartoons into them in the year 2023, right? So and you go wait a minute, it kind of nailed generate AI but the way it thought was very analog.

It's like this pan drawing on a page from the like analog brain telling it what to do, right? So there's always like the things you can't see and what they couldn't see we're computers and the rise of digital Nets and all these kinds of things. So I think the most interesting things are right now, it's all about the data.

It's all about scaling up. It's all about doing these things but there's could be some kid working on a whiteboard somewhere right who's like figured out how to like praying, you know train something the way you train the squirrel or dog or the way you train, you know, like human being if I take the kid out Garden for the ball to him, you know for a few weeks. I forgot to catch that ball and throw the ball back may not, you know go to the major leagues, but we'll know how to do it in a very short period of time.

I think if somebody will the thing you always look for the changes. Directory of things if that one thing that zero shot learner that mimic learning that like new technique that's fundamentally different the changes the game that I think is the thing that you have to you have to try to anticipate the way you do is you build for the technologies that we have right now, but you always be prepared to drop everything and kind of like shift in the New Direction when you can, you know, not overnight because you have these sort of Legacy, but you have to be prepared for when that for when that that don't hit the water and and makes this huge Flash and change that you want to look at it. They like that just changed the way we do everything.

I think we're on the verge of a couple of techniques like that that really accelerate how artificial intelligence learned and how it kind of contributes things into society and that to me the thing that gets me excited. I'm always looking for those things out there. All right Mitch real quick couple of predictions and then Lee take us out.

All right. My prediction is we'll continue to hear about sort of the human centered and we'll hear about the data Centric applications of AI and I think the more things the translated into kind of data Centric make that cross that Chasm help all of us and we'll still be fascinated by the chat gptu and with the next whatever the next thing is that kind of falls into the the augmented AI category or some variation of that. So it's interesting times to be in this industry for sure.

Thank you everybody. Thank you guys for what it's worth. My prediction is that Mike's debut album will go flatten them, but But I think everybody else's predictions.

They're gonna be much more useful. So let's let's plan to revisit in a year and see where we got to it remains for me just to thank our panelists William Mitch Dan and some Mike for helping coordinate. Today's today's show.

If you've liked and enjoyed today's content, please like And subscribe. We have additional content that we can create it with tech strong, and we look forward to seeing you again soon, please join us either text on TV or AI infrastructure Alliance that all take care. here