Cornell Keynotes

AI Today: Current Trends in Generative AI Tech

Episode Summary

Each day brings a new headline on artificial intelligence. Which stories should capture our attention and which are just clickbait? Karan Girotra, a professor at Cornell Tech and the Cornell SC Johnson College of Business, joins host Chris Wofford to answer that question and help us navigate the opportunities and challenges of AI.

Episode Notes

Some business leaders believe artificial intelligence is set to replace human workers in the not-so-distant future. Time will tell. In the interim, advances in AI are helping professionals streamline their daily workflows in exciting ways.

In this episode of the Cornell Keynotes podcast, Karan Girotra — the Charles H. Dyson Family Professor of Management and professor of operations, technology and innovation at the Cornell SC Johnson College of Business and Cornell Tech — explains the current capabilities of AI and shares the most newsworthy updates about the technology. His conversation with host Chris Wofford covers:

The Cornell Keynotes podcast is brought to you by eCornell, which offers more than 200 online certificate programs to help professionals advance their careers and organizations. Karan Girotra is an author of three online programs:

Follow Girotra on LinkedIn and X, and register to attend upcoming Cornell Keynotes in his AI Today series:

Learn more about OpenAI:

Episode Transcription

Chris Wofford: On today's episode, we're catching up on breaking news this week in artificial intelligence. We called on Cornell Tech's Karan Girotra to help us sort through the hype around OpenAI's release of GPT-4o, O as in Omni, and to get a handle on what's happening at Google and Microsoft. Big news and big moves from the big three this week.

 

Chris Wofford: So we'll review what's hot and what's not. And also look at major innovations in AI that are fueling business transformation around the world. Karan makes some super timely recommendations also on how to integrate AI tech into your six month to one year strategic plan. And this is an ongoing AI series.

 

Chris Wofford: So check out the episode notes on how to register for our free upcoming summer AI keynotes. Not to be missed listeners. Here's my conversation with Karan Girotra.

 

Chris Wofford: Karan Girotra, thanks for joining us from New York City. Great to see you again.

 

Karan Girotra: It's a pleasure. Pleasure being here.

 

Chris Wofford: So let's get right into it. I want to ask you what's new in AI this week. You know, we've had a lot of conversations, a lot of back and forth. It seems we can't go a week in recent days without some huge blockbuster announcement being dropped. Like just this week we had a huge news and media blitz debuts from open AI, Google, Microsoft.

 

Chris Wofford: Now, we all know how this works, right? A lot of this stuff comes with a lot of undue hype. People are getting really whipped up this week. GPT-4o, which stands for Omni, by the way, just released days ago. So what's real? Help us sort through what's real and what's hype.

 

Karan Girotra: No, exciting times indeed.

 

Karan Girotra: What is real is, the hype is certainly real. And I think there is some substance behind it. This week, the way I think about the recent announcements in the last week or so, I think we got a lot of things that make AI more useful in a practical sense. In our daily workflows, we did not see major technical advances or technological improvements like we saw in fall '22 and all of this exploded or fall '23 when we got GPT 4.

 

Karan Girotra: So no major technical advances, but a lot of things which actually in a practical setting make AI quite useful, including this only models and Microsoft's and Google's announcements.

 

Chris Wofford: So what actually did happen from where I sit in the conversations we had, it seems like there's a confluence of multiple technologies kind of coming together in what appears to be a new and, you know, ultimately revolutionary kind of way.

 

Chris Wofford: That's what all the hype is about, at least, right?

 

Karan Girotra: Yeah, I think confluence is the right word. I think, uh, let's step back a little bit and understand what we are calling AI. What we are calling AI today is essentially two new cognitive capabilities that we've unlocked in computers. Two new things that our human brain used to do.

 

Karan Girotra: Computers could not do, but in the last 10 or so years, we've been able to make computers do them at a level comparable to or better than an average human, sometimes still short of an expert human, but pretty close to a better than average human. What are these two cognitive abilities? These are basically abilities of classification and generation.

 

Karan Girotra: Classification and generation in particular are language and visuals. What do I mean by that? Classification of language. You give me some piece of text, I can tell you what that text means, or I can put it in a bucket. This is a positive comment. This is a negative comment. You give me a picture. I can tell you it's a picture of a man, dog, or more sophistically, what exactly is happening in that picture.

 

Karan Girotra: That's classification. We unlocked that about 10 years back. Then there's generation. You give me a description of a picture and I create a picture for you. So these two capabilities are human brain can do and computers could do for numbers for the longest time. But over the last 10 12 years, we've been able to unlock classification and generation for language and visuals.

 

Karan Girotra: Now this is all cool, we can make a computer which can in a way see, and based on what it is seeing, or independently, given a description, it can generate. That's the way to think about it. See, perceive, and generate. Now, all of these capabilities, classification, generation, poor language, poor visuals, all four combinations, in a way, existed in separate discrete tools.

 

Karan Girotra: So you have a tool Midjourney for generation of pictures. We have a lot of computer vision algorithms, which help us interpret pictures. We have a lot of algorithms, which help us interpret text. And then we have GPT or ChatGPT, which helps us generate conversational text. So we have all of these tools, but in discrete uh, in separate, tools.

 

Karan Girotra: What I think what we saw this week is integration and companies are stitching together these tools, at least in one user interface. Sometimes it's a superficial stitching in front, like the interface looks the same, but in the back, it's being routed. But more interestingly, from a technical point of view, there has been some progress and also creating the backend to be common.

 

Karan Girotra: So I think what we're now, now this sounds like, okay, we had four tools. You've made a multi tool instead of like a four tool. Now this can be quite from a user point of view. It's not just like I've created a Leatherman multi tool. it actually enables end to end tasks. Let's think about a particular task.

 

Karan Girotra: Let's think about the human brain first. The human brain is an integrated tool. It's not like we have a brain A and then we have to swap out our brain B for, um, perception versus generation. So now that, if we integrate tools, consider the case of, for example, an AI based tutor for a child trying to learn math.

 

Karan Girotra: And this has been one of the, I think, more cool demos we saw recently. In the past, you could make it work by, ideally you want a tutor to see what the child is doing. And then provide appropriate feedback at the right step. This is right, this is wrong, carry on, good job, bad job. You want to do that.

 

Karan Girotra: Now, with the tools I described, each of these discrete capabilities of seeing what the child is doing, processing it, creating some, generating some feedback, all of them were possible technically even before this week's announcements. But they were in different products. If you integrate them in a product, if in an unintegrated setting, beforehand, before, before this week's announcements, you'd have to take a picture independently of what the person was what the child was doing on their piece of paper, put that picture into a processing algorithm, which will perceive something then based on feed that a lot of these connections would have to be men.

 

Karan Girotra: But in the modern, hopefully what they will introduce to everyone in a few days. It's GPT-4o, only this only channel is kind of where it comes from, that you can really really perceive do it all end to end. You really just point your camera, your phone camera onto the piece of paper where someone is working.

 

Karan Girotra: And from that end to end, the processing happens and the feedback happens immediately. So it's really just integration, which from a practical point of view, is making things really cool. And that's what you will see in more, most of these GPT-4o demos, that they're really just clicking uh, integrating things together to make it a lot more useful.

 

Chris Wofford: I was really stricken by the open AI demos at first. Some of them were a little clunky during the rollout, right? We had a little, you know, talking over each other kind of thing going on uh, between, um, the, and the individuals in the room who were demoing, but they were super impressive if you think about what was actually happening here, the immediate responsiveness, the latency issue completely almost, you know, we can see it.

 

Chris Wofford: It's kind of visibly gone. I was really stunned by all of it. The, and the varying applications that they applied it to. What did you think when you first saw these?

 

Karan Girotra: So, first I thought I'm an applications person. So I'm always like, okay, man, this makes all the people, all the companies I've been working with to help them create applications.

 

Karan Girotra: This makes their job a lot easier because a lot of the applications, a lot of the products we see around AI right now are products that do the integration, the wrappers or different kinds of integration between knowledge systems and systems like this. So first as a advisor to companies building these products, I was like, oh, this makes people's lives easier.

 

Karan Girotra: Then I use AI in my, I use Chat GPT in my personal life, I have to confess, sometimes I'm tired in the evenings and I have a two year old daughter. She wants to hear a new story every night. And then I'm tired. I tend to rely on Chat GPT to tell the story a few times. And it's pretty cool because she can make stories about what we did today.

 

Karan Girotra: Uh, and all the activities she had, whatever is kind of what happened at her school. So that's pretty cool. but there was a pretty strong latency and the voice indeed used to be robotic. So I think these two things from a from an individual user point of view not from a business user point of view from an individual user point of view the things that were really impressive was the latency because now it's truly conversation. In the past, if you have to wait 30 seconds for a response and it comes to touch slowly, it's not, it's, we can call it chat, but it is really chat of the IRC era.

 

Karan Girotra: It's not chatt as, that is internet relay chat for those who might not know from the modems era. But it is not really, it wasn't really kind of face to face conversation type chat. So I think the latency is a very big one. The expressiveness of the voice is a big one. Also, I think from an economics point of view, I think the costs have come down quite a bit.

 

Karan Girotra: So all of this stuff is heavy on compute and they're at least charging less for it. So this will make make a big step from a business point of view. Latency and costs were the two biggest problems I would hear from technical leaders who wanted to use it in education, health care in other places from the technical leaders from the business leaders.

 

Karan Girotra: I think there are other challenges, but on the technology side, these two challenges I think are going to be a big step forward.

 

Chris Wofford: There was a little bit of drama in the rollout in that the Scarlett Johansson estate got back to AI and said, that sounds a little bit like me. What's the deal here?

 

Chris Wofford: What did, how did OpenAI respond to that? I found that kind of fascinating.

 

Karan Girotra: Yeah. So the details are still coming out and I'm not an investigative reporter with any more clear details. So everything with that caveat in bed. From what I have heard, I have read in public information and heard from our students and others who are somewhat closer to the scene is essentially, I think, openAI, Sam Altman wanted, wanted Scarlett Johansson's voice for this kind of new version of chatGPT.

 

Karan Girotra: He reached out to her a while back. She said, no, he reached out to her the weekend before the launch. She said no again, but it turns out that while he was asking her and she said no, they had hired a voice actor who could imitate her voice, who could be pretty damn close. And that is the voice of Chat GPT.

 

Karan Girotra: Now you might ask why Scarlett Johansson? There are a million reasons for that. I'm sure Scarlett Johansson fans have reasons, but there's also a pop culture connection because Scarlett Johansson was the voice of the uh, in the movie Her, which I think is the closest that has come to the um, uh, I think the modern AI tools we have.

 

Karan Girotra: Sometimes you perceive them as terminators and other things. But I think, the OpenAI, I would like us to believe they are kind of your friendly friends like in the movie, Her, at least in the premise of the movie, I think it doesn't end so well, but so that's the connection and yeah, so I think legally, I'm not a legal scholar, but it seems like if you hire an actor to imitate a voice that doesn't seem to be a direct violation of copyright. A lot of this stuff is still gray areas and the courts have to rule on it, but it doesn't seem like a direct violation. 

 

Karan Girotra: What I found a little bit personally, a little bit jarring, was sure they didn't get the voice, they trained someone else to do it seems kind of, okay, maybe a little bit aggressive business practice, but not illegal. I think what was a little bit unprofessional, was for Sam Altman to go in and treat uh, tweet descriptively her, which is where he was kind of trying to almost mock that he could get it without it, which when you're building powerful technology is not what I would advise people to do, which is not what mature CEOs generally do to be very direct.

 

Karan Girotra: Now, I'm sure he's a mature in other ways, but but that was probably not the wisest move. Otherwise, I think there are larger copyright issues in a lot of these models and their yeah, I think they're they're not clearly gray. They're not clearly black and white issues. There are some gray areas.

 

Karan Girotra: But gray areas get diluted and ignored when there's a lot of money attention. The amount of business support or business power they will get from showing a good demo sometimes clouds the gray and makes it more black. So I think a little bit of that was happening here.

 

Chris Wofford: Good. Yeah, not to get too lost in a side plot there or any Hollywood drama, but you did discuss some really important issues.

 

Chris Wofford: The ethical stuff is the stuff we're going to be watching very closely, certainly in the coming months. Okay, so Open AI, very impressive rollout nonetheless. So what else is going on this week? We had GPT-4o, which is Omni, but we also have Google and Microsoft. What have they been up to this last week?

 

Chris Wofford: What have we learned?

 

Karan Girotra: Let's start with Google first. Google pretty much similar to what OpenAI did. They had demos of models which are integrating different what the technical folks called modalities. But for regular folks, it just means a voice language. Everything is getting integrated into one common model.

 

Karan Girotra: And I would like to point out, it's not just the different modes. It's also different cognitive functions getting integrated classification generation. And I think that creates the end to end brain. Google was doing roughly similar things. Gemini, what I noted a little bit was it looked like their demos were a little, a little, they were more like, oh, this will be out in two months.

 

Karan Girotra: This will be out in three months. So I think this seemed, it seemed a little bit like me too, that we have this stuff also but wait a little longer. Their models do have some advantages. These models have longer context windows, which means think of it as short term memory for these chatbots.

 

Karan Girotra: They can, you can give them more information to work on. That can, I think, be powerful in a business use case. You can give them a bigger document to base their answers on. So I think that, that is that is an advantage of Google, but it seemed mostly in the same direction and somewhat catch up, which is intriguing for two reasons.

 

Karan Girotra: Number one, it tells us that so far, Open AI remains the leader and nobody's been able to really kind of take the crown from them. And it's been two years now and Google has, Google is the original place where a lot of these ideas are invented. Google has a lot of resources. So that's, that speaks well for Open AI or less well for Google.

 

Karan Girotra: So that's, I think one, one interesting thing. The second thing, which is kind of intriguing is, all this money going in, all these labs throwing things, but they're converging to the same stuff. Why? Is it because it's the only stuff that works? Because remember, the promise is not just these cognitive functions that we've been able to automate, classification, generation, but the promise is we'll be able to automate planning.

 

Karan Girotra: We'll be able to automate inference. Inference means you look at some things and then you draw some patterns, not just you classify things in like one by one, but you build some meta patterns on why do x. Why do animals look like this or something? Something deeper. So those things um, the fact that they're all converging to the same things could mean uh, could also be a very simple explanation. They copy each other, they have a lot of employees going back and forth, or it could mean that the technology's frontier is not as broad that we're not going to be, we're not fighting with automation on many places, but everybody's concentrating their firepower on a small part of the frontier which means it's the weakest or the first thing that can be done.

 

Karan Girotra: So Google catch up, mostly Open AI remains on the top. And it's kind of intriguing that there is convergence in what a lot of different people are doing, which by the way, there's a third kind of group of people who are doing the same thing, Open AI, um and Google and intriguingly, it was also the startups, a lot of the startups that are in this space.

 

Karan Girotra: Actually also we're, we're doing this stitching manually or as an outside product. And so I think the fact that the startups are also in this thing and these two big players are going in the same place means it's a relatively focused place where everybody's working. By the way, bad for startups, they were startups like Google integrate Gemini into their into their search engine.

 

Karan Girotra: Really bad for another company, Perplexity, I believe, which was doing essentially an AI search engine. So on a technical point of view, Google, very similar. There are some other things that Google did better, but on the core technology, they're converging.

 

Chris Wofford: So tell me how Microsoft stacks up uh, in this picture.

 

Chris Wofford: There's a lot of chatter about small language models uh, what we'll be calling SLMs. What do they do? What do they offer? I've seen this tech a little bit. I want to know what the state of the art is right now.

 

Karan Girotra: So I'll take that part by part. First, I think what can Microsoft offer and this applies to Google also.

 

Karan Girotra: Technically, I think OpenAI is ahead, and Google is catching up, or pretty close now, one could say, maybe a few months behind. But Google does have one big advantage, which Microsoft also does. Google has distribution. So OpenAI, you've got to go to OpenAI products. You've got to download ChatGPT if they have the app soon, which they claim they will.

 

Karan Girotra: You have to go to those products. We'll find out about OpenAI because you and I and other people who are tech savvy, or interested in this stuff, follow the news. For Google's products to reach people and for Microsoft products to reach people, you don't need to follow the news. Google can integrate all of these technologies within the workflows that are supported by Google products already.

 

Karan Girotra: And that's a big advantage because in the end, these are technologies which are where their economic value will come from, is worker productivity, just making people more efficient. And how? What are they using right now to be to do their work? Microsoft tools or Google tools. So I think the big advantage that Microsoft and Google have, is that they can integrate further.

 

Karan Girotra: It's not just the integration of the cognitive capabilities. It's not building one common brain, but it's also putting that brain in your documents, be it Word documents or Google docs. It's putting that brain with the eyes of that brain sitting on your desktop PC, which is what Microsoft is calling the copilot plus PCs.

 

Karan Girotra: It's not even like you have to point your camera at the student's homework. You're doing the homework on the screen and the screen is being watched now. Privacy concerns aside, there are some uh, it might be a little, a little too much but in principle, the big advantage that Microsoft and Google both have is they can integrate.

 

Karan Girotra: Now, Microsoft is building on Open AI products. So the main capability right now they're doing is on the consumer side, integrate into the productivity tools we use, and they are becoming the big provider for businesses on, they always used to be the back end provider, and they would be middleware layer, so to say, off of technology services companies which were customizing things for bigger companies.

 

Karan Girotra: But Microsoft is playing deeply into that segment also in the enterprise segment. They're going to integrate all these AI capabilities within workflows. So that's, I think, what Google and Microsoft feature. Google is both the back end technology and integrating into consumer workflows. Microsoft is integrating into consumer workflows and into enterprise workflows.

 

Karan Girotra: So I think that's where they come in. Yeah. And you, I think, asked about small language markets.

 

Chris Wofford: Yes, please tell me a little bit about SLMs. 

 

Karan Girotra: So SLMs, first, it's a little bit kind of clickbaity. Everybody's talking about large language models. Let's talk about small, large, small language models. They're not that small, but I think the main point there is the frontier models, which have the best capabilities.

 

Karan Girotra: They're quite heavy in terms of compute requirements. You need big computers to run them. And generally they run in the cloud, which means because our own computers are not that powerful, they generally run in the cloud. And the real time component of running this, when they're responding to the question, all the models, small, large language models in creating the model, then in really big computers.

 

Karan Girotra: Now, the difference with small language models with large language models is the day to day use part of those models that stuff is less heavy on compute. It needs less powerful computers. What that means is, it can run on your local machine. It doesn't have to run on the cloud. Now, the internet is fast, so it won't give us huge advantages in latency.

 

Karan Girotra: But if you're looking at the, looking at a co pilot sitting on your computer and watching the video game that you're playing right now and giving you feedback on how to play that video game. That was one of Microsoft's demos. That has too much latency to go to the cloud, process that image that is being seen and coming back.

 

Karan Girotra: So in those kind of use cases. You've got to have the model operating on what they call the edge, which means your local computer they're a smaller model offers a significant advantage because it can uh, it can do some very low latency tasks or tasks where you really want very quick response that cannot be done off the Internet.

 

Karan Girotra: They can be very, very useful. And then some people might feel more comfortable without sending the data to a cloud provider. Though I think we all send a lot of data to the cloud providers already. So I think that one is a little bit uh, once you know what all that you're sending on the, on the cloud, uh, you might not see the advantage too much because you're already sending a lot on the cloud.

 

Chris Wofford: I imagine that there's going to be tons of consumer demand for that, right? The degree to which we live and work in the cloud. That's going to be driving a lot of processes going on behind the scenes at both Google and Microsoft. No question. Right?

 

Karan Girotra: Yeah. Yeah. I think in the end uh, for Google has its hands everywhere.

 

Karan Girotra: Think of Amazon. Now, Amazon is mostly selling us the hardware and the backend technologies to make this work. They're having a bumper quarter also on AWS, even though they don't have arguably they don't have, they're not kind of have the frontier models or they're not making the push into anything because in the end, all of this needs computers to run on and all of those computers are either in the cloud, like, like Amazon has them.

 

Karan Girotra: And even worse on the cloud or on the, on your desktop, many of them need chips from companies like Nvidia. So at the backend, everything goes back to Nvidia and the people who are enabling all of these things. So I think it's interesting, like in any gold rush, everybody needs to use the ax made by the steel maker.

 

Karan Girotra: And everybody, the steel maker is the one who probably makes the most money upfront. And while the prospectors may get lucky or not.

 

Chris Wofford: A great analogy. Alright, a little aside here, we keep hearing about AI agents. What are AI agents? What do they do? What's the functionality here? I'm not familiar with them myself.

 

Karan Girotra: Yeah, so this is also a feature of it's also kind of integration. Right now there are tasks which AI models can do very well, which I've said many times classification, generation tasks, the other tasks, plain computation of numbers, which computers can do, but the models to do that are not the AI models.

 

Karan Girotra: It's your old school calculator however, that is coded up in the computer now. And there are other kinds of tools we use on computers. They're programming languages. There's all sorts of different tools that we have access to. Some of them are AI. Some of them are non AI. Now, the thing is, jobs or activities that humans do, they don't come labeled with, oh, this activity is for X tool.

 

Karan Girotra: This activity is for Y tool. And in fact, if you look at a job description, it probably has a whole host of different tasks, some of which are maybe not even fit for any kind of computer. So I think I always tell people when they're trying to automate a job, break things into three parts. What is still humans have an advantage?

 

Karan Girotra: What is probably for this AI tools and what can be done by other kind of tools, broadly called, let's call them expert tools. So what what agent models are that you give a real world task to an AI model to a large language model like chat GPT. And the first thing it does is it makes a plan. It says in order to do this job, I'll need these four tools. And then the large language model invokes tool one, invokes tool two, tool three, and tool four. Now, some of these tools might be the large language model itself. And some of these things will outsource to a calculator, for an example, to a web browser, for an example, to an email program, for an example.

 

Karan Girotra: So I think, I think what we're seeing here is, large language models are not just generating language, but they're also generating the language to coordinate different things. And then generating the instructions, which is also language, to each of these individual tools. So think of it in a way as this, they are the master controller, you give them a task, they plan the task. And then they kind of spread it out to different tools with the right commands in the language of that tool. So, it's actually quite intriguing because language is not just generating language. It could also be language to control and command lots of different tools.

 

Karan Girotra: And I think that's broadly what we are calling agents. They're okay for now. We shouldn't get too excited about them because this is a complicated task to be able to plan and then execute different things where have we seen it work well. People have been able to see it in, for example, fixing simple software bugs, simple bugs in code but we should be realistic there.

 

Karan Girotra: The best models, the best agent models for uh, which do the plan, they take a bug that has been identified, they plan or they take an IT task. I need to set up a computer. They break it down into four tasks and then kind of invoke individual tools to do that. They're okay. The percentage of cases that they can handle is of the order of between 10 and 15 percent 12 13%.

 

Karan Girotra: So it's not like this is really taken over but 12 percent is more than 0%. So it's not too bad. It is not nothing but it is, uh, it is not evident that we can do all sorts of planning and coordination using the large language models itself. Some of the simple tasks we can do, and that's why we're at a 12, 13, 15 percent type type performance in the 15 percent of the tasks can be done by them again, more than zero, but far short of 100 percent which a human would be able to do.

 

Chris Wofford: I want to build on your control and command analogy when we're talking about the functionality of some of this stuff you and I were joking about earlier, are we aspiring to something like a Terminator-like technology right sort of mission driven taken over, moving towards sentience or whatever.

 

Chris Wofford: Is that the right analogy?

 

Karan Girotra: Chris, I do a lot of these interviews, and all of them end up with Terminator. Terminator has to come up in every interview. I think it shows the power of how science fiction frames our understanding of technology. I think Terminator is a good ambition. Terminator is, uh, is an ambition.

 

Karan Girotra: Let me not say good, because remember, there are good Terminators and bad terminators, it is if you've seen if you're a certain age and you've seen all the movies, I have to make a side note, a lot of my current students haven't seen the Terminator movies. So, uh, and, and I don't know what's going to happen to this generation. This is my boomer comments, but back to Terminators. I think that indeed is a mega agent which can do lots of planning other things. I think we're nowhere close to Terminators cognitive capabilities. Or terminators physical capabilities. In fact, on physical capabilities, we're probably a lot behind.

 

Karan Girotra: Don't forget terminator. If I recall the movies correctly, could kind of could reconstruct its skin and stuff like that. We don't know that close to any of that. And we don't, we don't even have dexterous robots who can cut an onion properly. We're getting there with soft robotics, but we're so I think ignore the physical capabilities, but on the cognitive capabilities itself, which is our focus today.

 

Karan Girotra: I think even there we are, I don't think we're close to Terminators. And since the world likes, likes doing pop culture references and movie references and anthropomorphizing AI, I always uh, first thing, maybe, maybe don't think, don't try to watch, don't try to relate to a movie, but if you must relate to a movie, I always like telling people Terminator is not your right reference for where we stand today.

 

Karan Girotra: A better reference, a closer reference is another movie, a children's movie, the Minions movies. I think that is more close to where we are. I don't think we have terminators. We have minions. Now, why do I call them minions? Because they're uh, they're somewhat uncontrollable. They learn from what you're doing.

 

Karan Girotra: They kind of try to follow your patterns, but you don't really know what they've done. And then they get themselves into kind of crazy situations, which is what makes the movies funny. So I think again if you have to use a movie analogy, I think that's a lot closer. And I think the catchphrase that I think summarizes that is I don't think we have terminators.

 

Karan Girotra: We have minions. Before people start saying, oh, minions, that means AI is not going to be that powerful. I do want to make a case for minions. We have minions, but don't believe minions are useless. Minions are actually extremely powerful. Particularly when you do something in silicon, it means minions. We can make thousands of them.

 

Karan Girotra: We can scale them a lot. So I always like saying AI is not necessarily, we compare it for its brilliance of its intelligence. That's science fiction. What we really should think of is it's scaled cost in, in, in scaling up, not super intelligent tasks, but semi intelligent tasks. Somewhat cheekily, I'd say it should be, AI is a much better marketing term.

 

Karan Girotra: I'm not a marketing professor, but if I had to authentically say what was going on here. I'd probably say cheap dumbness rather than artificial intelligence, because I think that's what we're closer to then or scalable dumbness or scalable-ish ability to do things. So I think that's a better reference, Minions.

 

Chris Wofford: So many pull quotes to grab from that transcript of what we're talking about right now. So thank you for the overview. I think you've really synthesized a lot of what's going on and what it means and, and got us really up to date. So I'm thinking about our audience a little bit here. How should the business leaders among them, the policymakers people who are thinking about business transformation, actually, you know, leveraging some of this technology, pulling it into their, into their operational workflows.

 

Chris Wofford: Any advice on that before we move on to what we're calling part two here?

 

Karan Girotra: Yeah, no, I think indeed that is the focus of a lot of my research and the work we do at Cornell Tech, where we've brought technologists, people who invented some of these large language models, sit right next to people like me who, think about the business implications of these things.

 

Karan Girotra: So my research team, we are very solely focused on these things and, and very, very much focused on these things. And I think the first kind of way of thinking about all of these things that I want to give to our listeners is that these are general purpose technologies.

 

Karan Girotra: What are general purpose technologies means they're not like um, we're talking of automating. If you take the view that I just did in the last half an hour or so that, uh, so there are two views of AI. One view can be, oh, this is a cool tool that helps me do my homework. That's great for students. It's a cool tool which helps me build an AI tool.

 

Karan Girotra: Very good for tutors or very good for education. But I think if you step back, these are not just these tools. The underlying common thing that they're based on is unlocking these new cognitive capabilities, which means automating the ability to classification and generation as I've said a few times before.

 

Karan Girotra: Now that looks a lot different than a particular specific tool that looks like a general purpose tools. And if you look into economic history, what are the analogs of such tools which automate classes of work? I think the examples we find are things like the steam engine. Which automated a lot of physical work.

 

Karan Girotra: So this is a technology which is not operating in the specifics of a particular industry. It's not about making better vaccines. The underlying technology is about doing classes of work, which are across which are general purpose, which are across many different industries and our research team studies this extensively.

 

Karan Girotra: But if we look into economic history, general purpose technologies typically have a few different impacts. The first increased productivity at an individual level, like the steam engine made the viewers a lot faster, but the power load, they make the ability to transport goods much better.

 

Karan Girotra: Then they enable kind of operational efficiencies. And then finally, they enable new business models. And I think we can unpack each one of them. Our team spends quite some time doing this, but in principle, you've got to think of it as a general purpose technology, which will enable vast productivity gains.

 

Karan Girotra: New business models. Shifts in economic power across many different industries. So I think that's the first thing I want from all this tech talk we did. That should be the implication from a business point of view. That this is about classes of work that we are automating. And not about making picture making or graphic design better.

 

Karan Girotra: It's something deeper than that. It's a general purpose technology which will influence many industries.

 

Chris Wofford: So tell me, working at Cornell Tech, you work with many companies and organizations. From where you sit and your interactions with them, getting into their business processes, what is good in your estimation?

 

Chris Wofford: What's working for people right now?

 

Karan Girotra: Yeah, it's a great question. I'm going to say in the AI ecosystem, not just for companies, I think what is working very well is what is good from my point of view. Some of this is opinions, but what I think is working well is all these consumer demos that we see. It's not like a lot of people are using them in terms of the total number of people, but it did create enough hype that they created quite a lot of good energy that companies are thinking hard about it.

 

Karan Girotra: From my point of view, that's a positive. I'm a believer in the ability of technology to do things better, and I think so, the energy that these demos even the hype that sometimes certainly is ahead of itself. We can criticize it, but I think it is good in the sense that it's created a lot of energy.

 

Karan Girotra: Energy also means funding. That means a lot of groups are exploring the problems with them trying a lot of different things. That's why we're seeing all these updates, that's working well. The second thing which I think people underestimate what is good about this technology is, this technology is more accessible than previous technologies to build something using to build a we were talking of individual productivity.

 

Karan Girotra: One of the exercises we do with organizations is pick a job role and create a copilot for that and insurance claims adjudicator a learning designer at a corner. Pick a role and try to create a co pilot for that role, not a generic co pilot like Microsoft is doing, but do this specific co pilots.

 

Karan Girotra: And we come up with the specifications or knowing whatever we understand about these technologies, what tasks could be helped by having an AI co pilot and now many people after that are like, Damn, how will I actually make it? Do I need to hire a tech company to do it? The good news is the core creators of the technology, Microsoft, OpenAI, one can criticize them for many things, but this is one good thing they're doing.

 

Karan Girotra: They're making themselves, the platforms, very easy to build on. So this technology is way more accessible than people imagine. I'm willing to bet most of our listeners, with a little bit of kind of spending time, a few hours, should be able to generate, create their own GPTs. And be able to make pretty powerful tools for their particular use case.

 

Karan Girotra: So that's the second thing that's going well, a lot of energy and funding. And at least from a user point of view, this technology has a lower barrier of entry than most people perceive. It's a lot easier to use. You don't need to have IT consultants and hundreds of millions of dollars. And the third thing which I see which makes me happy is companies are experimenting.

 

Karan Girotra: It's good. It's good. Companies are experimenting a lot. Now there's a bad side to it. I think companies are only still scratching the surface of what can be done. And sometimes we have these clickbait experiments where we're also exaggerating a little bit the outcome. They're not what I was an academic would say double blind studies.

 

Karan Girotra: We have some of those and we're getting there. We've done in our research group, some of those on creative tasks, other groups have done them for consultants and other tasks. So I think there are academic studies, but right now we're just scratching the surface of what can be done and probably it'll take a little while to find the deeper uses and really know how well it will work.

 

Chris Wofford: I have a question from viewer Peter who wants to know how to stay on top of this stuff in a way that you do. Peter says, the pace at which AI is expanding and developing can be overwhelming. What do you do or where do you go to stay on top of AI development? And what kind of updates do you find more important than the others?

 

Chris Wofford: How do you separate the wheat from the chaff, so to speak?

 

Karan Girotra: It's a great question. And it's a challenge for me also.

 

Chris Wofford: Sure. 

 

Karan Girotra: So I think what I always advise folks is twofold. Number one, of course, you've got to follow the news. And one can, lot of problems with social media. But generally uh, Twitter or X is a pretty good source of knowing what, what, what is going on.

 

Karan Girotra: Now the challenge with what is going on, if you just take that as face value, there's a lot of propaganda and there's a lot of information. There's a lot of noise in that. So what helps me is have a conceptual background on what these things are. And I've been trying to kind of make some of that conceptual background clear to our listeners.

 

Karan Girotra: Capabilities, cognitive capabilities, classification, generation. There are some other ones we haven't crossed yet. Inference and reasoning. So every time you see a news item, I'm like, okay, where does that fill in that framework? Is it kind of going into new territory or is it the same thing? And so that's uh, one way to kind of put things in perspective.

 

Karan Girotra: Also, that's on the output side of technology. It's also good to have a framework. So on output side of technology, it's like, what is the space of cognitive work and what pieces of that space uh, where is the frontier of man versus machine moving on that space? So that kind of mental model helps in perceiving the capabilities of technologies.

 

Karan Girotra: Also, I think this is a little bit technical, but we should always remember all these AI technologies have three ingredients. Compute, data, and algorithms. And one can always ask, what changed? Why did GPT-4o get better? And in most cases, it is, oh, they just built a bigger computer, so they could build a bigger model.

 

Karan Girotra: Or they could, they had some new data to learn from, or they had some new architecture. So I think I always put these, so that's my framework in terms of capabilities, how do they go forward and in terms of ingredients, what's changed. And that kind of gives me a unifying framework to think of everything that's going on.

 

Karan Girotra: And without these frameworks, you'd be lost in uh, and remember, there's a lot of incentive for people on Twitter or other places. To keep you last because then they can sell you the magic notions and their 999 course on eight great things that happened on AI this week. Sorry, I'm being a little bit perhaps unfair, but it does bother me that that there's a lot of sometimes bad information also out there.

 

Karan Girotra: The way to immune to make yourself immune to bad information is just have solid grounding in frameworks. And once you think of frameworks, things haven't changed that much. Everything I've described was true 10 years back, and it's gonna be true for the next five, seven years. And what, what are the progresses here? So I think good understanding, conceptual understanding, at least as an academic, I believe, helps you be immune to a lot of the noise and, and understand conceptually what is changing and what's not changing.

 

Chris Wofford: Great question, Peter. Nailed it, Karan. Great answer right on the money. So we've talked about the ease of adaptability good energy going on within companies, accessibility.

 

Chris Wofford: We all have access to this stuff, very powerful tools. Let's talk about what, or some organizations may be not doing well. Where are people dropping the ball? Where, what are some pitfalls that we might run into as we get into the early stages of business transformation, whatever our endeavor may be.

 

Karan Girotra: Yeah, I think we're so I like calling telling people you're stuck in assistant land. I think we're still thinking of an assistant land is great. We have assistants who are helping on the jobs. But this technology is a lot more than that. I think it doesn't have to be just your assistant, it can be a partner just to kind of give a small thing or when we start, and that's still looking at an individual level.

 

Karan Girotra: When we start looking at a business level, there's so many business processes which can be automated in different ways. So I think can be made more flexible, less flexible. When we start thinking of business model type things, whole new kind of industries or whole ways in which we operator design kind of businesses and industries could change.

 

Karan Girotra: So I think we're in stuck a little bit in assistant land. I don't know stuck, but I think, uh, we're still in, in what I would call level zero of the benefits that, that maybe I'm a little impatient that bothers me a little bit. The other thing that, that I think is going a little bit less good is sometimes to make the best of AI, I think it is important to understand the right business context and integrate with the business use cases. So I think what one phrase that I think captures it quite nicely with this modern AI, it's quite easy to build stuff. Like you were saying, it's accessible. It's easy to build, but it's really hard to know what is the worthwhile thing to build with this thing.

 

Karan Girotra: So easy to know how to build, but hard to know what to build. And I think that is the challenge companies are. So the, they're stuck in assistant land because the other higher end uses a higher KP higher benefit uses. Um, you a lot of these technologies require a little bit deeper understanding and work on trying to make this work.

 

Karan Girotra: And what kind of understanding you need. You need what I was saying before conceptual understanding of the technology and understanding of business processes, understanding of business models, genetically and in your context. So it really is what we try to do at Cornell Tech and few other people are doing.

 

Karan Girotra: You get business people talk about this, but they don't have a conceptual understanding of technology. You got my great technology friends who are talking about all of this. They don't understand how, industrial processes work, for example, or what is the economics of business models in different industries.

 

Karan Girotra: What you need is to marry the two things. And I think maybe it's an aspiration rather than a complaint. I think the next level of use cases will happen when people will be when the hard part, what is the useful thing to build gets tackled, which will happen by combining technology and business, which is what we're trying to build at Cornell, like really putting these two things together in ways that might people don't even like being put together in these things, but that's where the magic will come.

 

Karan Girotra: One third thing that bothers me a little bit. So, it was stuck in assistant land. I think the business understanding and the integration is very important. And then I think the third thing that bothers me a little bit is sometimes companies are restricting use, like don't use this stuff.

 

Karan Girotra: It's too dangerous. And you know what that leads to people using it in more dangerous ways that leads to people using the free versions, which are much worse. It leads to people kind of posting it on personal devices, free versions don't have the privacy protections, the paid versions have privacy protections.

 

Karan Girotra: So it leads to kind of bad use without knowing how to use it. I think so restricting is bad. Ideally, you should, yeah, you shouldn't restrict maybe train and enable, in a more controlled way because restriction never gets anywhere because these technologies are available to consumers anyways. So I think deeper integration of business and technology to get the higher end applications and really dealing with the restricting is not working.

 

Karan Girotra: And if anything, all of this restriction, other things creates fear, anxiety and probably more negative emotions in your employees than makes sense.

 

Chris Wofford: You know, as an authority and an educator, you're at Cornell Tech, you're a faculty author of a couple certificate programs that are offered to the university for AI for productivity, generative AI for business transformation.

 

Chris Wofford: So I should ask you, you know, who are you teaching now, either in your in person cohorts or your synchronous online cohorts? What are you advising as far as telling people what to get together for their like six year or six months, rather to one year plan. How do you kind of make recommendations? And in what sequence, how does this work?

 

Karan Girotra: I think different recommendations for individuals and business leaders. I'll start with business leaders. I think if you're leading a business and could be a small business, your business, the end point you want to have in six months to a year is definitely you should have run a few experiments by now.

 

Karan Girotra: I generally have a rule. If you have 10 million of revenue, you should be running at least one experiment. If you're 100 million of revenue, 10 experiments. So I think that now it's it's a number I made up, but I think it's roughly right about not a super precise number, but something of that nature.

 

Karan Girotra: So you want to be running several experiments and hopefully in six months to one year, you have a couple of those experiments that seem to be good, but more importantly, you have that nuanced understanding of the business technology intersection that you can come up with beyond assistant land bigger, bigger experiments to run next. And if you're a big company, some systems, that's the end point for business companies. Now, end point is easy for me to talk, but then the next question business leaders ask me, okay great end point. How do we get there? How do we really get there? And I think that is where, where I have something more, uh, more, the end points, I think most companies also are aspiring to that.

 

Karan Girotra: How to get there is where I think we can offer some some insight. What I've seen work well, I've worked with a few big companies by now to kind of see it work well. I call it three pronged approach, the bottom, middle, and top approach. You got to create some bottom up energy. What does that mean?

 

Karan Girotra: That means do a short training for everyone, stop restricting use, and give a large number of people, in the thousands, access to these tools in a sandboxed environment to try it. Run some contest. Amongst Chris or anybody can vote in and type in what was the best use case this week also on some bounties on what could be problems with or I did this and it led to a really disastrous outcome.

 

Karan Girotra: So let create the bottom up energy by enabling people to do it, bringing it in. Don't make it kind of too., you have to do something, but at least enable people to do it and you will get some intelligent people doing it. You'll remove some of that fear, anxiety and the individual productivity benefits will probably come from that.

 

Karan Girotra: Middle, or I would say not middle, but just below the C suite, some people you need to get a good number of your N-2, N-4 people kind of to really sit down and think of how this will change the business. For these folks, we typically run a 3-4 day bootcamp, get in there, try to learn how we, there's a lot we haven't talked about.

 

Karan Girotra: How AI creates use cases. We have time. We can talk about it next. But in principle, go through a bootcamp type setting to systematically think of how you would change, not just an employee using it for writing their emails better, but how that would change the business process, the business models, all the things we we research about.

 

Karan Girotra: So that some group has to be dedicated to do it. That is where you will get the big productivity gains. You'll get an AI ready organization. It'll protect you from competition, which might be doing it. And they might even identify some new business models and growth opportunities. All of that will fail if the top leadership doesn't become intelligent sponsors.

 

Karan Girotra: So, so typically we'll run all three versions of these this knowledge that we share with companies, something for everyone, then some focus thing for a group to really get there, really find these things. And then all of that fails. Whenever I do with the N minus two or N like couple of people, a couple of layers below the CEO, what should I, their complaint is, well, our bosses won't support this.

 

Karan Girotra: Bosses need to become intelligent sponsors, and I think so basic conceptual awareness of the kind we were talking about early on about technology. Where it's no longer possible to say, Oh, let the IT folks figure out the conceptual underpinnings of this or let XYZ figure out about this. I think to become intelligent sponsors, the leadership also needs to do at least some training, education, practice themselves to get, better of this better with this.

 

Karan Girotra: Because there were two advantages of this. One, there will be some completely new growth opportunities, which will probably come from people who really kind of think of the business strategically, not so much tactically. And second, I think you got to, this is going to be big enough that you have to start creating proper structures of change.

 

Karan Girotra: And that will require senior leadership to be involved. So for companies, if you want to get many things, get some experimentation and real outcomes going and to get their substantial experimentation. People are doing a little bit of experimentation. I think I'm talking of a much bigger scale of experimentation than most companies are right now doing, enable that deep kind of pronged approach, bottom up energy.

 

Karan Girotra: Some just below the top. Sit down on the ground boot camp to figure this out. And then at the top at least everyone needs to be an intelligent sponsor. And the middle in the top what I'm calling could be very overlapping. Many times the top leadership will be part of this kind of bootcampish approach to do it.

 

Karan Girotra: But at a minimum, they also need to get the top level things. From an individual point of view, I think I'll give a shorter answer. You got to upskill. You've gotta know what is going on here. You've gotta upskill in all these elements, understand the technology, how the technology will create opportunity in your job and then become more experimental, more innovative in trying these things out.

 

Karan Girotra: So that's kind of my, what does all of this mean for individuals?

 

Chris Wofford: In my notes. I've got a recommendation that leaders create structures for change. What do you mean by that? I want to understand that a little bit more clearly.

 

Karan Girotra: So I think, remember, any technology we create, it is it is exciting, but there's a lot of, we don't know how it will exactly play out when we put it in your business. How will people react to it? The unintended consequences. We don't really know how much productivity advantages these things will give.

 

Karan Girotra: So any of these technologies, they have a lot of promise, but a lot of stuff fails also. And when I say structures of change, I think what has become more important is creating the structures that allow for wide experimentation and intelligent failure, not just any failure, because it's very, it becomes cliche to say we should encourage failure.

 

Karan Girotra: It doesn't mean blow up and burn money also. And too often big companies do that also. So I think we need to create structures of change. This means overall, what I like calling agile leadership, more things we used to teach as, in innovation. I like saying innovation used to be a nice to have, now it's a need to have, if the technology frontier and other things are changing.

 

Karan Girotra: So the structures of change would involve systematic ways of thinking about generating opportunities, systematic ways of evaluating opportunities, and then intelligent experimentation, which leads to sharp, which leads to cheap, quick failures, rather than long drawn out failures, which typically happen in the corporate world.

 

Karan Girotra: So you've got to get that innovation engine working very well. And that is a lot of that is classic stuff. We would we used to do. It's just become a lot more important being agile, being experimental, being creative and thinking of things, having systems too, so structures in three dimensions, identification of opportunities, evaluation and experimentation.

 

Karan Girotra: I think that's what organizations need to do. Experimentation while watching the legal, ethical and other risks around these things.

 

Chris Wofford: Yeah, I want to off ramp for a minute into the ethical territory. You know, there was a little bit of chatter early on in our discussion about this, and now we're thinking about leadership, right?

 

Chris Wofford: Is there anything that has happened related to some of the developing news, the breaking news that we've heard, that throws up a red flag for you as far as ethics goes aside from the Scarlett Johansson thing, which we're not going to talk about anymore?

 

Karan Girotra: There are many things one could, one could talk about the ethical, intended, unintended consequences about AI.

 

Karan Girotra: But I think what this week brought into the focus in terms of ethical issues is, is the dirty secret of all of these models. These models are trained on data. We don't really know which data. And it is unclear if the companies had rights to train on all of that data. Now, what do we mean train on the data?

 

Karan Girotra: It means doesn't mean they're copying the data or storing the data. They're looking at all the data on the internet and interpreting or learning something from it, getting inspired from it in a loosely spoken way. So a lot of models are doing that. And it is unclear if that was a copyright violation or not.

 

Karan Girotra: Now there are two schools of thoughts here. One school is like, okay, there are fair use doctrines, which means when you read a New York Times article, you can get inspired by it, you can know, know that information and you can write an email to your friend about it. That's generally not considered illegal. But, and in a way, the computers are doing the same thing. Except they're doing it on a massive scale. They're not reading a article, they're reading all the articles. And so, so it is not a hundred percent clear because the laws were not written imagining such a scale of reading. So it's a gray area in a way one could make reasonable arguments on both sides, potentially.

 

Karan Girotra: But what is uh, and what, how is it going to play out? Hopefully the courts will resolve it. But in absence of this resolution, what the companies are doing, it is ignoring it. They're assuming it's okay and completely going and training on these models. So what I worry a little bit about is, and they're moving really fast, why are they doing it?

 

Karan Girotra: Because I think two years back or three years back, it would have been like, Oh, too risky. Let's not go there. But right now I think companies are like, damn, this is such a lucrative opportunity. Let's take the risk. Let's just go there and see what happens. So I think basically, the AI companies have decided to ignore the tech regulation uh, the copyright regulation, sorry, and we'll see where it plays out.

 

Karan Girotra: What I hope happens is the courts come up with some fair arrangement and that leads to some markets for appropriate compensation of everyone involved. Right now, the tech companies are, we're doing it, too bad for us to the world. We're taking these technologies. We're going to do what we're going to do with it.

 

Karan Girotra: And that to me is uh, I can understand financial incentives to do that, but it is, I'm afraid it can lead to, it might not be something we can pull back on. It might not be able to kind of, it might be a last battle by the time the courts adjudicate, which is how it happened in several industries on books before when Google Books started scanning all of them.

 

Karan Girotra: It didn't happen too much in music, but could go there. Could go there. So that's what worries me in terms of the societal impact of these things. Other ethical issues also but that would need a longer conversation on that one.

 

Chris Wofford: Thank you for listening to Cornell Keynotes. Check out the episode notes for information on two important AI certificate programs from Cornell University, authored by today's guest, Karan Girotra. Check the links for both the Generative AI for Productivity and the Generative AI for Business Transformation certificate programs here at eCornell.

 

Chris Wofford: Thanks again, friends, and be sure to subscribe to stay in touch.