Analyst Geoff Webb on DeepSeek and the Business of AI

[00:00:00] You know, back in the day, I had a business, six full-time employees, and though I loved the business and I loved the employees, I hated all the paperwork that went with it. I grew up in a Fortune 500 company, and I wanted everything to be that buttoned down, and I could never do it. I was always struggling to keep up with time tracking, benefits, performance management. If you're in HR, you know exactly what I mean.

[00:00:26] So I can tell you, if you're spending more time managing your HR paperwork than your people, now is the time for Bamboo HR. It's the HR tool that streamlines your process so you can focus on your team. Bamboo HR simplifies everything, so you can stop spending hours and hours on payroll, time tracking, benefits, performance, creating reports, and all that.

[00:00:51] With Bamboo HR, those hours are shaved down to minutes, so you'll get your time back to do what you really love. So, are you ready to see what Bamboo HR can do for you? Get a free demo at BambooHR.com slash free demo. Get back to doing what you actually love in HR. Go to BambooHR.com slash free demo. That's BambooHR.com slash free demo.

[00:01:21] Klar können wir Multitasking, aber wenn's drauf ankommt, sind wir froh, wenn wir uns auf eine Sache konzentrieren können. Das neue Samsung Galaxy S25 Ultra macht's möglich, als dein persönlicher AI-Begleiter. Aktiviere Google Gemini und frag die AI zum Beispiel nach passenden Restaurant-Optionen und teile sie mit deinen Kontakten. Das klingt dann so. Hey, such mir ein indisches Restaurant in der Nähe raus und sende es an Luca. Easy. Was das Galaxy S25 Ultra noch kann? Erfahre mehr auf Samsung.de

[00:01:50] Welcome to PeopleTech, the podcast of WorkforceAI.news. I'm Mark Pfeffer. What is DeepSync going to do to the business of AI? That's what I'll be talking about today with my guest Jeff Webb.

[00:02:15] He's a contributing analyst with 360 Insights and has a long history as an executive and strategist in the world of HCM tech. Most of the discussion on AI has focused on capability. DeepSync looks like it's going to shift a big part of that conversation to business and budgets. That's what we'll get into on this edition of PeopleTech. Hey, Jeff. DeepSync has been all over the news this week.

[00:02:46] Trashed the stock market on Monday. Let me ask you, can you summarize what's going on with DeepSync and where do you think it's going to end up? What do you think it's going to lead to? Yeah. I mean, obviously what's happened is, you know, suddenly out of the blue, this organization, this Chinese business, which is tiny, by the way, like a couple hundred people,

[00:03:12] have launched onto the market, an AI engine that is trainable and operatable at a fraction of the cost of what we've been expecting so far and seen so far from the really big, you know, the big AI monsters out there around, you know, OpenAI and Meta and Google and all the other folks. And even Apple's lurking there in the background. And the thing is, it's not just a bit cheaper to train and run.

[00:03:39] It is orders of magnitude cheaper. Like it's, you know, it's the talking of potentially what, 45th, 50th of the cost to train up this thing, which is huge. And what it's doing is it's sort of resetting all of the expectations around the cost models for building and operating these large language models and inference engines and, you know, chain of thought engines and all the other things that sort of people are developing.

[00:04:07] And that's huge because what it means is that when you've been thinking about what's it going to cost, what's required, what's the infrastructure necessary to move into this space and deliver on the promise, you've got suddenly an engine that comes in that's likely almost as good, if not actually as good, that is tiny by comparison from the point of view of cost, tiny by comparison, the point of view of the physical horsepower required to run it, right?

[00:04:34] So you're not building an immense data center running hundreds and hundreds and hundreds of these super expensive high-end GPUs. It's being trained on a tiny number. And I think what's interesting is we're all shocked by its arrival. I mean, they've produced it. They've launched it as open source, which is really interesting. Everyone's shocked by its arrival. And yet at the same time, I don't see that we should be.

[00:05:00] Like if you think about where we are in the evolution of applying things like large language models to business problems, we've barely started. It's just that we're into this a couple of years in. And as a result, we should expect something that is developing as fast as it is where we're so early. There's going to be huge rapid shifts that come out of the blue that are going to force us to rethink the way we plan for and execute using these technologies. Especially in business settings.

[00:05:29] And that, I think, is what's the most interesting part of this. Sure, there's something that's revolutionary. And I imagine we'll get into the conversation about how it came about and why. But what does it mean for organizations that are thinking we need to be putting AI engines at the core of the value we deliver and building around that? When you don't know if in another six months, the numbers, the game, the sizing, the weight, the economics of all of this are going to get reset again by something else.

[00:05:59] And I think that's the question that we should be really asking. It's like, how do you, not so much what does this one mean, as how do you get ready for the next one and the next one and the one after that? Which are going to come, right? It's going to happen. Well, and what's interesting about this is it's the first story about the AI business I can recall that's really not about the technology. It's about the cost. Yeah. And the difference.

[00:06:26] I think I saw something that their model cost something like $6 million to develop compared to hundreds of millions of dollars for like a chat GPT and, you know, other of the brand names out there. Yeah. And the thing is, the reason it's not about the technology is because it can't be about the technology because they were being forced to develop this using technology that was essentially wasn't restricted by the US.

[00:06:53] So, you know, the US has been restricting sort of export of these incredibly powerful engines, you know, the physical, the hardware, the devices, the chips and so on because we wanted to maintain control of that. So what this group of guys did was they went, and I presume it's ladies too, they went, okay, well, then we'll use what we've got. And instead, we'll just rewrite a lot of the assumptions about how you build these things in order to use that to deliver it as efficiently as we can because we don't have any choice.

[00:07:21] I mean, if this is like the definitional example of, you know, invention, necessity is the mother of invention. They went and said, okay, then we'll just have to start again. And they went, you know, it's anything from changing the way numbers were used and stored to rethinking the way that the system attacks problems to how it relearns things.

[00:07:39] They just, they went, let's look at every assumption about how these things are built, rethink it to run more efficiently and faster on cheaper, simpler hardware, which we have a lot less of, and then went ahead and built that. And I think it's interesting that they were able to, again, because they had to maneuver around the very challenges that the USA industry had put in the place to slow them down. And in the process of doing so, they have completely rewritten the rules for everybody else.

[00:08:09] And again, you're like, this is going to happen. This is so early days. We should expect this. And what do you do if you're open AI who's put scads of money into your product development and now suddenly you've got someone who's, these guys, I mean, DeepSeek, they're going to compete on price, I would imagine. Well, I mean, the thing that's fascinating, again, is that first of all, they've got to look at, we're making this open source, right? So you can take a look at, you know, here's all the research papers. Here's what the code looks like.

[00:08:38] So, you know, first of all, if you're one of those other big companies, you are frantically reading through that stuff right now going, you know, asking yourself a lot of what the hell moments. The second thing is, yeah, you look for how did they, I think, again, how did they get to this point? What are the assumptions that are built into the way that these models were built?

[00:08:58] You know, one of the things that was interesting was simply the way it was storing floating point numbers was, you know, there's been an assumption that you have this massive amount of RAM required to hold all of this, the tokens, right? In memory at one time. And they kind of went, no, let's not do that because we don't have it. So let's do it a different way. Why don't we split them out into multiple different parallel systems and have each of them understand a little bit and then we'll figure out how to ask them.

[00:09:21] There was a whole bunch of things that I'm not saying that, you know, from our end, the big guys were getting lazy, but they were building on the back of assumptions that have been made and sort of driving forward. And they were these people were forced to go. We can't do any of that. So we have to start again. I think that's what, you know, if you're one of these, the big powerhouses with a lot of people, you're kind of asking yourself the question. Should we be rethinking the basics of how we do this?

[00:09:50] Because what happened was things that shouldn't have been as effective turned out to be pretty much as effective and, you know, at a fraction of the cost. I saw one estimate that said that, you know, these guys built the entire thing for about the same as the total compensation package for a small number of people running at these big organizations. So, you know, they basically built an entire disruptive set of technology for the cost of a couple of people running for a year working on some of the big stuff.

[00:10:16] And that is a level of disruption, I think, that we weren't ready for and should be ready for. Welcome to Froster Feierabend-Tip. Faster than the delivery service from... Oh, no love to cook. ...to... Hey, already done. ...by with Froster Butter Chicken. 100% free from added. Only with natural ingredients after the Froster Reinheitsgebot. Kurz... Mmm, voll lecker. ...null künstlich. Gerichte from Froster.

[00:10:46] And that you can save. I don't know about you, but I don't know... ...very many people that are hopeful about the world of work. And I'd like to change that. My name is Marcus Mossberger, and I started the Hope at Work podcast where you'll find two things. Number one, really interesting guests. And number two, innovative ideas about the future of work. Check it out.

[00:11:12] How does this align with the movement towards smaller language models? I've been reading more and more about the development of models that are really geared toward a certain area, you know, certain tasks. So that's a smaller application with smaller overhead, et cetera. And now here you have DeepSync coming along, which, you know, smaller costs to be certain.

[00:11:41] Smaller costs, but good capability. Do you see those two things coming together to really change things even more? Well, so I know, for example, when I talk to our AI fellows and the teams, small language models are definitely something that is a useful tool because it allows you to... You want to be very focused on certain tasks, right? You don't have to build very generic large language models. You can build small language models.

[00:12:09] You can, you know, augment them with rags, which are those, you know, the technologies that basically further refine and augment the searching capabilities and the understanding capabilities. So, in other words, you ask a small language model, a question, the rag steps in and goes, by the way, look over here. This is where I think your answer is going to be. And you have a much more efficient sort of set of responses running much faster with less horsepower required to deliver it.

[00:12:33] And I think that the terminology for the way that I think, if I remember correctly, that DeepSync we're doing, it was mix of experts where you have, again, you have parallel systems that are much smaller, focusing in different areas. And essentially, you then have a couple of layers that sort of coordinate the activity of those, all of which seems to make a lot of sense, right? You're becoming more specialized and focused on the technology delivering specific sets of outcomes rather than building highly genericized platforms that could be used for anything. And, you know, for us, that's a big area.

[00:13:03] You've got, you know, you want a technology that can do this very thing over here and you focus it and you train it on that and you use other technologies around it to sort of shape and guide it and give it a faster response and a more focused response. I think that's what most businesses want anyway. Like, I think businesses are trying to, you know, they're solving, they're looking at AI to help augment, solve the way that they've been dealing with other problems. I want to do a better job of this. I want to identify this stuff sooner. I want to be able to dig a little bit deeper in that.

[00:13:31] Essentially, the core of what most businesses are doing is that they're staying focused on the capabilities they already want to deliver and they see AI tools as being focused on making that better, faster, you know, easier to use. A lot of conversational AI stuff going on. So, yeah, I think so. I think I expect that as these things become much more mainstream, they will become more focused on solving sets of problems. And what you see is instead of sort of one big AI engine lumbering along, you see a proliferation of very focused tools that are designed to do very focused sets of things.

[00:14:00] With some kind of sets of coordinating layers above that. And I think that coordinating layer is the interesting area for communication between them and abstracting up the capabilities of the layers underneath. Because, again, you go back to this whole point of things I'm going to develop that are going to be difficult to, you know, you can't kind of engineer too close to the engine itself. Because if you do, you're stuck with its particular requirements.

[00:14:27] What businesses should be thinking about is how do I abstract out the capabilities to plug in new AI engines, new AI capabilities easily when these sudden tectonic shifts occur in what's capable in the market? So I don't want to get too tied into one technology. I want to abstract it up and build a, you know, a virtualizing layer, some kind of definitional layer that allows me to go plug in whatever engine I need to, how I need to, where I need to, so that I'm not tied to the whatever was last week's technology.

[00:14:55] Which is kind of sad to say that that's where the state we're at. It seems like if you look at the AI business in general, this could go a long way toward pushing it substantially forward. Because all deep sea keeps going in the direction it's been going. The cost of implementing AI solutions is going to go down significantly.

[00:15:22] A lot more people are going to start to wade in and start trying it out. Oh, yeah. I mean, again, I think you're going to find, I mean, why wouldn't you, right? At this point, it should be crazy not to. Yeah, I think there's, and again, we're going to see, we'll see this sort of AI popping up in, again, focused AI is popping up in more places. So you've got, you know, you've got the sort of the big platform vendors out there who are going to be plugging AI in all over the place.

[00:15:49] You've got people who are the sort of edge compute vendors who are going to suddenly have access to a lot more horsepower, a lot more easily. You know, Apple's very big on how we're going to plug AI into solve, you know, to the edge devices that we sell. The smaller, more nimble the engines are, the more efficient they can be run, the less horsepower it requires to train them and the less horsepower it requires to run them, the more places you can put them. And suddenly you can expect to have a lot more intelligent things in a lot more places.

[00:16:16] You know, whether that's sitting in some piece of business software doing something or it's actually physically out there, you know, a much smarter lawnmower. I don't know. I'm trying to get an example. You know, smarter devices that run warehouses, for example, or smarter devices that go out and, you know, perform other tasks that would have been just way too expensive or way too compute heavy, way too slow, right?

[00:16:39] With the previous sets of technologies or way too communications heavy, suddenly a lot of those problems start to go away and the door is open to be putting much smarter technology in many, many more devices in much more, much more sort of a proliferation of intelligence. And, you know, the Internet of Things suddenly becomes kind of the Internet of, you know, the AI of things kind of thing out there. Everything gets a lot smarter, a lot faster.

[00:17:07] So, yeah, these kind of changes are just going to continue to accelerate the process at which AI itself accelerates change. So, you know, if you're a CHRO or a CIO about implementing AI in your day-to-day operations, does this change your planning, do you think?

[00:17:30] Does it change your vision or is this really you get to do the same things faster and cheaper? I think it does a bit of both, right? I think it accelerates your ability to achieve the goals you already had. I think it opens the doors to thinking more broadly about what is and is it possible?

[00:17:48] You know, and it's things like training and personal development and utilization to, you know, to be more, you know, more efficiency in the workplace and better communications and more targeted information. I mean, who wouldn't have? I tell you, I would, if somebody's got an AI that will figure out which is the junk emails that I'm just copied on and I don't need to ever read and just puts them to one side or summarizes today's emails for me

[00:18:14] and then, you know, sends it over and then does a quick call with me and lets me know what I needed to know. I'd be delighted. I'm sure there's something out there already before somebody sends me a message saying, hey, we already built one. But just like the ability to filter out the noise that we have to deal with and focus on the stuff we have to do, I think would be huge. Those are all things that I suspect will get easier and faster. But, but, but here's the other thing.

[00:18:38] I think businesses need to be very, they need to be aware of the fact that the more stuff you do with AI, the more that will potentially, you know, the workforce themselves becomes a little unsettled around what does this mean for me? Are you replacing my job? What's your strategy, right? And I think that's something you need to be clear of. The second thing is use of data. Got to be so careful, right?

[00:19:02] I think we've already seen this with businesses attempting to control the use of some of the existing sort of, you know, large language models of don't put sensitive information in in order to get an answer out because you never know who else is using that. And there's a lot of opacity around what happens when you put information into these, these systems and where does it reside? And is the model using it to learn? And does it pop up somewhere else? And who's going to know about it?

[00:19:29] And I think when you think about personnel information, that is especially important. You can't be put, you know, you don't want to dump your whole employee base into a model and tell me who's most likely to leave next week or who might get sick in the future or who's going to, who's good for a promotion, only to discover, even if you should be asking those questions, that that information is suddenly no longer fully under your control.

[00:19:51] And the reason beyond just pure ethical reasons is you don't want to put an AI engine at the core level of a business function only to discover that regulatory changes require you to tear it out tomorrow or pay fines. And then you have to re-engineer a whole bunch of stuff that you were relying on. So, yes, you should use AI. Obviously, yes, we should be using functionally focused AI.

[00:20:17] But the data that it consumes is there's every bit of the same requirements to manage and control and secure that data as there is whatever else you have into it. It doesn't obviate the requirements to be careful with people's information. Jeff, thanks very much. Thanks for dropping in on short notice to talk about all that. I'm sure we'll talk again soon. We absolutely, you know, lots of short notice conversations are going to be coming up, I think. So, yeah, great to see you, Mark.

[00:20:56] My guest today has been Jeff Webb, contributing analyst with 360 Insights. And this has been PeopleTech, the podcast of WorkforceAI.news. We're a part of the Work Defined podcast network. Find them at www.wrkdefined.com. And to keep up with AI technology and HR, subscribe to WorkforceAI today. We're the most trusted source of news in the HR tech industry.

[00:21:25] Find us at www.workforceai.news. I'm Mark Pfeffer.