Search

Key Trends in Generative AI – The AI Moment, Episode 1

Key Trends in Generative AI - The AI Moment, Episode 1

On this inaugural episode of The AI Moment, I give my views on the key trends in Generative AI and identify some Adults in the Generative AI Rumpus Room.

The discussion covers:

  • Key Generative AI trends: LLM/Foundation model mania, Cottage industry to support LLMs, Harnessing AI compute workloads and the Enterprise dilemma – DIY or not DIY?
  • The Adults in the Generative AI Rumpus Room: Anthropic, Kolena and IBM

Watch the video below, and be sure to subscribe to our YouTube channel, so you never miss an episode.

Listen to the audio here:

Or grab the audio on your favorite podcast platform below:

 

Disclosure: The Futurum Group is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this webcast.

Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of The Futurum Group as a whole.

Transcript:

Mark Beccue: Hello everybody. I’m Mark Beccue. I’m the research director for AI with The Futurum Group, and I’d like to welcome you to The AI Moment. It’s our weekly podcast that explores the latest developments in enterprise AI. There’s this pace of change and innovation in AI, is really dizzying and unprecedented. With this podcast, The AI Moment, we’re looking to distill the mountain of information, separate the real from the hype, and provide you with some sure-handed analysis of where the AI market’s going to go, where we think it’s going to go.

So in each episode, we’ll deep dive in some of the latest trends and technologies that are shaping the AI landscape, from discussions around the latest advancements in technology to parsing the mutating vendor landscape, including the big announcements. But we’ll also cover things like AI regulations, ethics, risk management, and more. We’re going to cover a lot. And I would say this too, just so that when you’re listening, you can think about who’s listening, it’s really designed for folks to guide business decision makers with enterprises that are implementing AI, and that could be a lot of people with Cs in front of their names, but it doesn’t have to be, but it’s a CIO, CTO, CDOs, CEOs, legal, everybody. But that’s the AI vendor communities, the companies that are AI vendors, businesses and IT consultants, investment banks, law firms. We like everybody. So, you’re all welcome, but that’s what we’re aiming for.

So each show may be about 30 minutes. This one’s going to be a little bit shorter, but typically we are going to make it up of three to four segments. So the first segment is typically a guest spotlight. We’ll have special guests come on, usually from a vendor company, and talk about what’s going on with them and what they’re seeing, what’s going on with them. So that’s one segment. The other three are; key trends in generative AI, and this is just really a review of trends. I don’t want to say… We’re almost in a phase where these could be fads, not going to do fads, we’re going to do trends. It’s a important distinction. That’s second segment.

The third segment is one of my favorites. It’s called adults in the generative AI rumpus room. And this generative AI moment has produced a lot of chaos, disruption and a lot of impatience. And so some organizations have been calm and thoughtful leaders in the midst of this. And these are the adults in the generative AI rumpus room, and they need to be called out. We like to talk about them, because they’re shining examples for the rest of us to follow. So there’s that section.

The other section will be a company, we like, doing AI. That’s one thing we’ll talk about occasionally is just, “Hey, these guys are doing something cool. Check it out.” And that’s about it. We’re going to jump right in today. I don’t have a guest a spotlight today, so we’re going to start right off the bat with our key trends in generative AI. So give me one second. So our first segment is the key trends in AI today. Let’s talk about that. We could spend a lot of time thinking about this, but I’m going to break it down today into four pieces and then there’s sub pieces to that. The first one we’re going to say is; you can’t throw a rock in AI and not hit an LLM these days. So, really my first trend is going to be the LLM and foundation model mania.

If we really break it down, everything that’s really happening today is keyed around what you can do with a large language model or a foundation model. And what we’re seeing, and it started with OpenAI in October of last year, we’re almost coming up on a year when that was announced, but basically we have announcements daily of new players. I almost see three or four a week, where we’re hearing about a different LLM coming out. So tons of new iterations. And then the second piece to that announcement is these alliances. So we’re starting to see alliances between certain LLMs and bigger, let’s call them, vendor infrastructure players, cloud players, people like Microsoft and Google and AWS. And we’ll mention those here in a minute. So take that element, say, “This is one piece that’s going on.”

At the same time, to get new players and alliances, what you also have is mutations. And what I mean by that is our first LLM was really OpenAI and ChatGPT, and it’s this massively huge model really built on training of just a giant amount of data. And what’s happened is we’ve seen those LLMs mutate into smaller, more narrowly focused models. And the reason for that has to do a lot with power, and the computation it takes to run them is massive, massive, massive amounts of power taken just to train these models.

So you’ve got more people thinking about, “Well, can we make smaller ones? Can we make more narrow ones?” And the idea around narrower ones allows those brains to think more narrowly around a particular subject. A couple of different examples there. One is Anthropic is building a telecom-specific LLM for SK Telecom in an alliance that they’re leading. So just think of it as they’re building a set with a more narrow data set that’s thinking more specifically about a certain industry’s needs. So that’s one idea.

So you have these smaller and more narrowly focused models. You also have a movement in this mutation to what I’ll call better, more secure, more accurate, less hallucinations. And this is really the crux of a lot of issues right now is there’s this great potential behind LLMs, but they have some challenges. You have accuracy, they’re not always accurate. Depends on the amount of data they’re training on, what data they’re trained on. It could be misinformed, it could be dis informed. They could be bias. It’s all based on what data they’re training on.

And the other big one that people like to talk about and we have an issue with is hallucination. So, these models are very confident in their answers, even if they’re wrong. So, that doesn’t really help. So that can cause some issues. So you’ve got these mutations and ways that LLM makers or developers are thinking about how to build these better. So all of this is going on massively fast, and that’s all the model mania that’s going on. So let’s leave it at that for a second, skip over and say the next bullet point… So first bullet point is LLM foundation model mania.

The second bullet point is there’s a cottage industry around supporting LLMs. And that’s because of all the mess we just talked about, because part of it is; well, you got to build LLM band-aids, what I call. So you have to have all these different tools that are coming out from different companies for building more efficient models, for finding inaccuracies, for catching hallucinations. So you have two things going on at the same time. You have LLMs that are trying to get better, but you also have a cottage industry of people that are building tools that are band-aids to fix the ones that don’t work. So that’s interesting.

Second part of the cottage industry to support LLMs is there’s this idea that companies are understanding they have to have much better data management and data governance. And this really reckons back to that idea that we’ve heard for years and years about this idea of big data and companies being able to leverage the data that they have. The problem’s not that they don’t have the data, it’s that they can’t use it. So companies, lots of things are siloed.

There’s certain kinds of data they can use and not use, but this is really the LLM cottage industry is how much data can you give these models to work with and what data can you do with those LLMs that’s your data? So there’s a scramble across the board of trying to figure out how to access these things, how to secure these things, having the right tools. Where does that stuff sit? Is it on-premise? Is it in the cloud? And all these different things have been spawned and triggered into action by supporting LLMs. So that’s point number two.

Point number three on the AI trends is harnessing AI compute workloads. And what I mean by that is, again, we’re going to go right back to the LLMs, these foundation models LLMs, they use the most compute-intensive workloads that have ever been put into play. The IT folks and the data engineers will tell you, it just takes massive amounts of compute to run lots of different pieces of the AI workload for LLMs, mainly for training. Inference is a little less intensive, but it’s also a big deal.

And what that has caused is it’s put all this pressure on companies that are trying to introduce AI, have to think about costs. They got to do the business models on; what’s the cost for me? What’s the outcomes for me? What’s my ROI? And they have to reduce the cost of the compute and increase the compute efficiencies. So they’re looking at two different things: how do I reduce costs that way? Is there a way to do it by increasing compute efficiencies? And they’re looking for help with that, because they can’t necessarily do that themselves.

So, what that’s caused is another bunch of work being done by different companies. One is, there is a quest, a holy quest, to build purpose-built AI chips. And many of you might know that GPUs are being used for most of the AI workloads these days. They’re built very well for that, but they weren’t designed specifically for AI workloads. It just so happens that they work okay and pretty good for that. So that meant that certain companies are really well positioned to handle that, like NVIDIA has been doing GPUs for a long time and actually they built that for gaming, believe it or not. It was what they started out, but there you have it, they’re the AI Kings.

But what you do now is you have almost every chip manufacturer is thinking about; can they build chips, purpose-built chips, for AI? And this is a massive undertaking and everybody’s working on it. We’ve got a lot of movement in this space. Really fascinating to watch everything from Intel, which is a huge player, doing those types of things to any other startups, all sorts. So, that’s really a big trend right now is to build these chips that could process more efficiently these AI workloads. And there’s one other piece to that is data centers weren’t really designed, purpose-built to process AI workloads. And so there’s some interesting movements in that space for data centers to rethink how they treat AI workloads versus other workloads, and is there ways to do that more efficiently? So there’s some intriguing players in that space, maybe challenging some of the bigger cloud providers in how they would help companies be more efficient with their AI workloads. All right, so that’s point number three into the trends.

Final one is really what I call the enterprise dilemma of do it yourself. And that means that, because we’re at the front end of generative AI and understanding all of these costs that we’ve talked about, we don’t really know, we don’t have a lot of metrics in place about the outcomes and what the ROI is. So you have enterprises thinking about, “How much of this do I ingest and try and do myself? Do I have my own data centers? Do I do some of this on site? What are my resources? Do I use the cloud? How much do I use the cloud? How much do I lean on vendors to outsource a lot of these things? How much should I do myself?”.

And that’s really been a consistent piece of the AI story ever since I started covering it in 2016, is that companies have to think through the cost-benefit analysis of how much they do versus how much they outsource. And it’s that classic story of how you think about; are you speed to market? Are you second movers? Kind of issues. What’s your real strategy overall for your company? So is it important to be on the front end or is it better to be a fast follower and those kinds of things. So, those are our big trends right now in AI. Good. All right, let’s move on. Give me a second. I’ll pull up the rumpus room.

All right, so for our next section, we’re going to talk about the adults in the generative AI rumpus room. My favorites right now, and we do this every week, so there’s always somebody that’s in there. Sometimes they repeat. I might have a hall of fame on adults in the generative AI rumpus room. I’m not going to do that right now, but we’re going to talk about some folks that are very good today. And again, this idea is highlighting industry players approaching generative AI in the right way. And so I’m going to give you a quick intro. This is what I always do to talk about what this means.

Generative AI is considered the fastest moving technology innovation in history. It’s captured the imagination of consumers and enterprises across the globe. And what it’s also done is caused copious amounts of FOMO, a lot of missteps, false starts, and these are really the classic signals of technology disruption. There’s lots of innovation, but there’s lots of mistakes. And it is a rumpus room full of a lot of kids going wild, maybe a little less so these days than it was a few months ago, but still a lot of kids going wild. And so the rumpus room needs adults and guidance through the generative AI minefield comes from these thoughtful organizations who don’t panic and they understand the fundamentals of AI and can manage the risk.

All right, so this week our picks are, Anthropic, Kolena, and IBM. So let’s go through this real quick. So Anthropic, I’m going to call it, they came out with a Responsible Scaling Policy. This was on September 19th. They published this, what they call their Responsible Scaling Policy. It’s intended to manage the potential severe risk of using their AI models. And so I’m going to read you what they said in this. They said, “As AI models become more capable, Anthropic believes that they will create major economic and social value, but will also present increasingly severe risks. With this document, we are making a public commitment to a concrete framework for managing these risks. We focus these commitments specifically on catastrophic risk, defined as large scale devastation, for example, thousands of deaths or hundreds of billions of dollars in damage, that is directly caused by an AI model and wouldn’t have occurred without it. The AI represents a spectrum of risks, and these commitments are designed to deal with the more extreme end of this spectrum.”

Yeah. I have some thoughts on this, but I’ll tell you a little bit more about what it really… We broke it out a little bit as to what they were talking about. And so they built this framework, it’s based on a concept of what they call AI Safety Levels. And they’re defining that as a series of thresholds that represent increasing potential risks. And there are two types of risks; deployment risk, which comes from the active use of the model, and then there’s containment risks, which come from just possessing the model. And an example of a containment risk is building an AI model that would enable the production of weapons of mass destruction if it were stolen. So just because you have that, it’s a containment risk.

So, it was interesting when you look at what they were breaking all this down and they said; this is an iterative process, and they admit that it’s building the airplane while flying it. And I thought that was a good way to put that and to be honest about where they are with this. And so they’re attempting to govern systems and problems that really haven’t yet even been built or hatched. So it is what it is. And I think that’s an adult move for a couple of reasons. At the very least, they’re taking some responsibility for the AI models that they produce, even though they’re not particularly… What’s interesting about that is these models aren’t particularly, they’re not specifically designed for any particular use case, and they’re not necessarily intended for any good or bad.

So it’s hard for Anthropic to go, “I don’t know how you might use this.” And what’s interesting about what they’re going to do is they are going to monitor how their models are tuned and used, and will first make it harder for these attackers to steal what they call the model weights.And second, they’re going to build in these, what they call, misuse prevention measures that would allow Anthropic to shut things down, so they’ll monitor it, and it makes sense. And I think there’s some other… We’ve heard this from some of the other makers, some idea of like, “Well, maybe we should be helping put preventions in place.” What’s interesting to me about this, it’s a good initial move, and I also think that it’s an adult move. It’s a shrewd one too, because really these become defacto standards and that can lead to legislative preventions.

Our second adult this week is a little company called Kolena. They’re a startup and they put out a product that allows you to test and benchmark model performance, AI model performance. Here’s the news, September 26th, they announced they raised $15 million, that brings up their total raise to 21 million. And they say that they’re focused on building trust in AI by making models work better. Well, remember we talked about before, you’re an adult in the AI rumpus room when you get these things right, and there’s a lot of ways you get these things wrong. I’m going to read you a little bit about their approach and then we’ll go into why they’re an adult.

So, Kolena can provide insights into identifying gaps in an AI model, test data coverage. And the platform incorporates risk management features that help to track risks associated with the deployment of a given AI system. Using Kolena’s UI, users can create test cases to evaluate a model’s performance and see potential reasons that a model’s underperforming while comparing its performance to various other models. So they can manage and test all this in specific scenarios that the AI product will have to deal with, rather than applying a blanket aggregate metric like an accuracy score, which isn’t very helpful, and can obscure the details of the model’s performance.

So that said, okay, that’s what they’re trying to do. This is an adult move, because while AI has got all this… There’s really, it’s built in that there’s a degree of nebulousness to it. How does it work? How do we know it’s accurate? With rules-based development, you always have a trace, you can see where you went. And with AI, it’s much less so. You just don’t necessarily know all the time. It’s a black box.

So, they have to be tested and monitored. These AI systems have to be tested and monitored. And building and deploying accurate AI models isn’t the only… Kolena is not the only company trying to do this, and they join a growing number of others who are helping enterprises think through this. And I think it just shows more momentum. And here’s hoping that these monitoring tools are more widely adopted, so the more of these companies that get into the space, the better. We’re going to see better results. And those are adults. So, Kolena’s adult number two.

Adult number three is IBM. I’ve talked about IBM A few times in my writings on the rumpus room. They’ve been adults in the rumpus room before in my research notes. So this is my first podcast. We’re going to talk about them in this sense. Today, on September 28th, they announced that they would provide standard intellectual property protection to IBM-developed Watsonx models. And what that really does… It’s interesting why we consider this an adult in the rumpus room space. Here’s why; we’re in this time where there’s a lot of activity and interest in AI. It’s also a time when people are just learning about AI, what it can do, what it can’t do, as well as what it should and shouldn’t do. And this time, it’s a time of really minimal trust in these outcomes. And consequently, companies who build trust in AI outcomes will help the market advance.

So, I think what IBM’s doing with this IP piece is it’s helping build trust in those outcomes. It’s guaranteeing. It gives those companies that are trying AI some confidence that IBM’s behind them on when they try and do these things. There’s some similar provisions, in a different sense, from Adobe, for their AI-fueled product called Firefly. And I just think that there’s going to be more and more of these adults in the AI rumpus room. Will back the AI that they put into production. And that’s interesting, because they’re doing it downstream. They’re a provider, they’re not the end provider. So they’re a supplier of these things. But I think we’re going to see more of that. And hats off to IBM for being an adult in the rumpus room again this week.

All right, that is going to wrap it up for this week’s session of The AI Moment. I’m going to thank everyone for joining me here. Be sure to… describe. Be sure to subscribe, rate and review the podcast on your preferred platform. We’ll see you next week. I’m Mark Beccue.

Author Information

Mark comes to The Futurum Group from Omdia’s Artificial Intelligence practice, where his focus was on natural language and AI use cases.

Previously, Mark worked as a consultant and analyst providing custom and syndicated qualitative market analysis with an emphasis on mobile technology and identifying trends and opportunities for companies like Syniverse and ABI Research. He has been cited by international media outlets including CNBC, The Wall Street Journal, Bloomberg Businessweek, and CNET. Based in Tampa, Florida, Mark is a veteran market research analyst with 25 years of experience interpreting technology business and holds a Bachelor of Science from the University of Florida.

SHARE:

Latest Insights:

CTERA Adds Cyber-Resilience with New Honeypot Functionality
Mitch Lewis, Research Analyst at The Futurum Group, covers CTERA’s new honeypot cyber-resiliency functionality.
Quantinuum Announced a Dramatic Improvement in Error Rates that Should Lead to Faster Adoption of Quantum Error-Correcting Codes
The Futurum Group’s Dr. Bob Sutor discusses Quantinuum’s announcement of achieving better than 99.9% 2-qubit gate fidelity and what this means for quantum error correction.
On this episode of The Six Five Webcast, hosts Patrick Moorhead and Daniel Newman discuss Apple Vision Pro developers losing interest, U.S. awards Samsung and Micron over $6B in CHIPS Act funding, does AMD have a datacenter AI GPU problem, Adobe’s use of Midjourney, Samsung knocks Apple off of number 1 market share, and Arm says CPUs can save 15% of total datacenter power.
In Recent Years, the Concept of a Sovereign Cloud Has Gained Significant Traction Among Nations Seeking Greater Autonomy and Security in Their Digital Infrastructures
The Futurum Group’s Steven Dickens observes that Oracle's recent $8 billion investment in Japan not only expands its cloud infrastructure but also strategically aligns with the growing global trend toward sovereign cloud solutions.