00:00:21
Hello and welcome to Tech TV. Today we’ve got a special program on AI and some of the recent news out of Silicon Valley. My name is uh Bill MW um and I’m joining you here from the studios in the UK and I’m joined by my colleague Pete Warren. Pete, tell us a little bit about what’s happening with the recent acquisition that Nvidia have made and what it really means. Well, it’s it it’s it’s fascinating, isn’t it? There we were. Everybody was going getting ready for Christmas. It
00:00:57
was Christmas Eve and very very quietly Nvidia announced that it was going to uh take an interest is probably one of the best ways of saying this in a company called Grock. Now it’s very very important that we remember the Q in this. It’s GR O Q as opposed to the search engine gr. Now, Nvidia has has bought this for 20 billion or sorry, it hasn’t bought it. It hasn’t bought it. It’s very very important to say it has not bought it. Um, it’s it’s made it’s investing $20
00:01:36
billion in a relationship with Groc. Um, now one of the things that it’s interested in apparently is inference. So, Bill, uh, tell us about inference. what what what is Nvidia’s interest in uh Grock? >> Obviously, there’s been a lot of interest in AI recently and people are investing an enormous amounts in data centers and GPUs and in the GPU arena. Uh really Nvidia have been the market leader by uh some significant margin although they don’t have the market entirely to themselves. Um it has
00:02:17
allowed them to become the most valuable company in the world with uh trillions of dollars in market capitalization. But they need to be looking ahead because there’s a a concern that there’s a bubble and we need to look at well what is going to sustain this massive valuation they have. And there’s a critical difference here in AI between what you do to train the AI which is like having a race car and putting it in the um uh wind tunnel and focusing on its aerodynamics and tuning it and get
00:02:49
the suspension right with actually running it um remotely for clients where they need to get the value from it which is inference because first of all you train these AI models and then you need to run them the inference what the instructions that you’re making from them. And one of the things that’s absolutely critical to the performance of inference is really high-speed memory, hopefully on chip if you can get it as close in order to maximize the um uh speed and minimize any latency. Um
00:03:20
and the investment that Nvidia are making here is in precisely that. It’s the high-speed memory. and people envisage that we’re going moving away from separate GPUs and CPUs um two integrated AI chips that have everything on them and hopefully this will be both the CPU and the GPU and the uh MPU if you’ve got a processing uh logic to to do for uh neuralistic programming. Um but also you need to have the high-speed memory on uh on the chip in order to drive it as possible. So in the the
00:03:57
racing car analogy, this is the fuel injection that you’re actually going to drive you around the track and get you as fast as possible. So that’s the acquisition that Nvidia are making which is going to be an investment that they feel will position them well for the future, the move away from purely being a a GPU oriented outfit to actually having more of the essential ingredients for an AI chip. Well, I think no there is something very very important that we have to stress, Bill, and you need your knuckles wrapped
00:04:32
quite considerably. Uh, GPU. What is GPU? What is CPU? You’re using >> My apologies for using wanttonly a series of threeletter acronyms. The CPU is the computer processing unit. This is like the traditional brain that operates most of the software within a computer. The GPU is the graphical processing unit and that’s the thing that actually works out some of the graphical um uh thinking that drives computer games that does a lot of the magic that here we’re doing in a studio to provide you with video. But also it’s
00:05:12
been found that it is the best way to actually crunch a lot of the data for um a lot of the AI training that you need to do. But obviously if we’re taking the focus away from training and onto inference suddenly really high-speed memory comes to the four. Okay. And inference um a lot of people are going to be puzzled by a lot of these technology terms. inference. Uh um a lot of people would would assume that that basically means that the system is actually trying to come up with an answer quicker on the basis of training.
00:05:51
Is that right? It’s using the training which has created the the model and the uh logic that underpins the AI model. But actually it needs to implement your question as quickly as possible. And that will um involve writing um a lot of information immediately to memory and then bring it back again in order to process it fast. And that’s why some really high-speed memory becomes really really important in this particular aspect. Now there is also an aspect to this acquisition that is worrying some
00:06:28
people um and that is the fact that as you alluded earlier they haven’t actually acquired the company. They’ve done something that we’ve seen a few times recently which is an aquire where actually instead of paying top dollar for the entire company and all its assets which would trigger clauses within a lot of the employees contracts where they might have stock options and might be hoping to um make a a a real killing from any um investment in shares that they have or in options. They’ve got around that and also
00:07:08
importantly around a lot of regulatory skew scrutiny by investing in the company and their investment has actually particularly benefited them by allowing them to potentially poach is a bit of a a tainted word. They’re taking a lot of the talent and they’re licensing the technology and they’re compensating some of the early stage investors. But potentially the people losing out here are the employees. many of the employees who joined uh this and any other startup um for potentially lower wages than they
00:07:48
could get elsewhere in the expectation that they would get a a big uh win and a big payday when the company uh is either acquired or floats. But under this aqua model, it’s not going to float. There’s no change of ownership that would trigger the clause within their options. Um, and there’s no ownership change that would allow the regulator to step in and potentially interfere. >> So, essentially, we’ve got an independent company that one other company has put a lot of money into, but
00:08:25
it hasn’t bought it out. I mean, we we started to see the genesis of this, didn’t we, with Deep Mind because Google bought Deep Mind and Deep Mind’s head is now the head of Google. Uh they were after the talent within Deep Mind and we This is an extension of that, isn’t it? >> It is. And it’s becoming more common, particularly recently with a series of these aquaers. Now we’ve got our man in Silicon Valley who we’ll probably bring in here for for comment. We’d like to
00:08:58
welcome um James Delisanti. Now James, thank you for joining us. Um you’re our man on the spot um out there in in the US with your ear to the ground. Uh we’d be very interested to hear from you exactly what the perception is of this aqua hire trend and what the potential implications are for Silicon Valley the overall culture there. >> Well, thank you for having me uh and being able to speak on techtv.live regarding this really interesting deal. I think from our perspective, it’s a
00:09:34
little bit different than the framing of this conversation in that Nvidia’s uh licensing of Grock’s assets really looks like a s um a Christmas gift to the Gro company, its employees, and its shareholders. In this deal, if you read between the lines and how each of the um uh shareholders are being um compensated, everyone gets a lot of cash very quickly. The interesting part about this deal in the context of AI uh is that in Silicon Valley here, we know that the path from LLMs to AGI is quite
00:10:18
uh uncertain at this point. there’s a lot of controversy that LLMs alone which are probabilistic technology um won’t get us to AGI. So the other forms of AI which are starting to become um come into people’s awareness or from the prior AI um bubbles but um the machine learning alone isn’t enough because you’re training your models. You also need knowledge representation, you need inferencing and you need reasoning. And what’s so interesting about this deal is that it is the first stake in the
00:10:53
ground. It’s verifiable data to say that LLMs along alone are not the starting point to get you to the promised land of what AI AI’s potential is. You need inferencing. And why is that? Because the patterns that the machines learn are not proc are not being processed fast enough. The GPO alone is not enough to accelerate LLMs when it comes to answering questions and delivering specific answers fast. You have to be able to recognize patterns within the data that’s been learned. And that’s the
00:11:33
that’s the rationale for the LPU by Grock. So it’s a it’s Nvidia acknowledging that more technologies need to be evolved in order to move AI collectively forward which is to move from learning to providing answers and inferencing is a key core tech in AI to get that to happen. Now, what’s interesting about the deal is that for antitrust reasons, like you’re saying, how do you move the state-of-the-art forward, but solve for the complexity of Grock, which is a 10-year-old company
00:12:12
with a lot of money invested that built a different kind of business, which is hosting and acceleration on inferencing. um how could they get that technology or get access to the LPU technology and reward the employees and the shareholders who have stayed with the company for 10 years. And what’s really interesting about this deal from my perspective at least here in Silicon Valley is it’s actually elevates the bar in Silicon Valley for what an acquisition could look like. And that’s
00:12:47
because of how the CEO of Grock and the board chose to compensate themselves, their investors, and the employees. So, for example, if you understand the um nature of the deal, most employees that work at a startup, they they have a cliff upfront when they join a company and you have to wait a year before you can invest your shares. So for the most recent employees at Grock who joined within the last year, that cliff is being waved so that all of their shares vest, which means they will receive cash, an accelerated cash amount for the
00:13:27
most recent employees. So they win. For the employees that have been there for quite some time, 85% of their uh shares will receive the distribution. it’ll be paid upfront for their vested shares and a 10% is paid midyear and the remainder at the end of 2026. So in effect they are having their vested shares paid at the $20 billion valuation within a year which is an extraordinarily short time for a startup and its employees to receive the earnout associated with um the acquisition. So those employees are winning who have
00:14:11
been there who have both vested and vested unvested shares and about 90% of the Gro employees are going to be joining Nvidia and they’ll be paid cash upfront for all their vested shares and the remaining shares will be paid out at the $20 billion valuation over time uh according to a vesting schedule for Nvidia share uh Nvidia stock. So with >> So what you’re saying, Jim, what you’re saying is that actually this isn’t such a bad deal. What you’re saying is that in a sense this possibly redefineses the
00:14:49
models for how uh companies will operate in these startup environments. >> Yeah. To me, it absolutely represents a new completely distinct way to operate when you’ve been working very hard for 10 years on a technology as the market comes to you. It is absolutely a new bar in how employees have are being treated during this holiday season. And it’s no surprise that they announced it during the holiday season cuz I believe everyone is winning. In typical Aquires, they will acquire the employees in an
00:15:28
attempt to steal the assets by stealing the knowhow and the IP of the employees. Here, um, what’s so interesting is Nvidia saying, “We’re going to license your technology on a non-exclusive basis.” So the current shareholders can still uh license the technology for their data center and for other players who want and it’s saying that NIV Nvidia is confident enough in their ability to monetize this technology. They don’t have to put any constraints on the technology they’re licensing.
00:16:04
But this isn’t the case with all Aquaars. what we saw recently with there was Alphabet’s uh attempt to uh buy Windsurf which actually fell apart and immediately after that a a predatory organization came in and stole some of the key staff and a lot of the employees that weren’t seen as valuable were left with very little indeed. Um and therefore while while at least with the traditional IPO model where you’ve got um options and vesting and everything like that, you know where you stand with
00:16:42
a lot of the aqua hires or if this becomes a real trend, then there’s very little certainty and there may have been a level of generosity to the Grock employees which we applaud but there’s no guarantees here. >> Yeah, I think it I think it really reflects and I’ll use this metaphor. It feels like a every uh CEO and board group and the founder group has a choice of how they want to treat the separate employee stock option pool upon a change of ownership. And I look at it that it’s
00:17:16
as clear as day. Some people are Ebenezer Scrooge before his transition, the the the miserly Ebenezer Scrooge to use a Dickens reference. And other CEOs set this up with generosity like Ebenezer Scrooge after he became the generous person. And the the reason why the acquires happen in the former case is because they don’t have a I believe an attitude of the generosity which is how do we take care of our employees during the change. The investors have been in this a long time and they clearly from the choices made
00:17:56
and the way I interpret this deal they are clearly making it possible for the employees who want to move to Nvidia they can. they get their payouts according to if it was being acquired and at the same time the shareholders who want to stay in Grock’s current business which is the data centers they can continue running the business as the CFO takes over the CEO job but I have been part of acquisitions and been part of acqu hires where the intent is in fact to not pay the employees and you can see that pretty clearly But since
00:18:34
the data on the uh compensation is so public, they are taking it public in this way, making the information and the compensation public because it’s a good deal for everyone. And those aquires you’re describing, they they will be private deals and the only way you can find out about the employees not getting their fair share for the long time they worked is um you have to talk to them personally. It’s a private deal and it’ll be held private. Okay. Well, let’s we’re going to come
00:19:04
back to that towards the end of the this relationship and and how those uh relationships develop in that rather denzian way that you’re talking about. But I mean, let’s just look at for the time being LLMs, right? Large language models. I mean, just to wrap everybody around the knuckles because they keep on using these threeletter acronyms. Um uh what we’re talking about here is this RI is $20 billion being used to actually get people their answers quicker. What we’re saying is that people don’t want
00:19:42
this latency. And what we’re saying too is that in a sense we’re also developing this anthropomorphic bond that people have been talking about because we want to give you the answer quicker. We want to actually make it look it like it’s much more of a human relationship with the AI. That’s what this $20 billion is being paid for, isn’t it? >> Yeah. And what’s so interesting is $20 billion has a lot of commas associated with it. And you can imagine the value that Nvidia places on being able to get
00:20:17
those answers faster by paying $20 billion, which even in the in the um most conservative valuation must mean that they think it’s worth a hundred billion or more because they’re paying 20 billion, which means they’re going to make they’re going to pay that money, they got to make it back. And so >> and one of the other things about this, Jim, is you know, you mentioned data centers. Well, there’s been lots and lots of concern about energy usage, about water usage for cooling. Um, if
00:20:50
you can make these systems more efficient, then of course you’re going to lower those costs, aren’t you? You could also lower the needs for this these big data center models as well. >> Yeah, that’s a great point, Peter. And what it’s showing, and we talk about this often, is that if you need 4 billion or 8 billion parameters, something inherently is inefficient about the idea of modeling the output of thinking because we’re modeling all the language processing coming out of all
00:21:21
the thinking, which is the all the content that exists out there on the internet, but that’s the output of thinking. And so we’re modeling that with with what can arguably be an inefficient method which is driving the need for power, driving the need for large data centers. And what you can see is the while they move the state-of-the-art art forward where artificial and intelligence continue to evolve, they’re actually saying we need to make uh data centers more efficient. How do we do that? We make the AI more
00:21:57
efficient and deepseek is a perfect example. Better algorithms, better chips, cheaper inferencing, less learning and more thinking. So we’re moving from learning. Well, you don’t want to always keep learning. You have to start logic and reasoning from that learning. And what’s also missing from the equation ultimately is knowledge that you already have. So something >> that’s one of the fascinating things about this because so many people are thinking, “Oh, AI the the model of it is
00:22:30
all sorted out already. That’s that’s what we’re going to move towards. We’re going to move towards these great big data centers. We’re going to need this investment in all of these data centers. Nobody’s thinking about the speed of evolution that is occurring. I mean, is this a move towards an artificial intelligence chip? A uh an what what should we call it? An AIU? >> Yeah. Yeah. Well, one one could hope. Um there’s arguments both on both sides of getting to artificial intelligence. But
00:23:03
um the notion of reasoning and thinking as a human brain thinks, it stores knowledge in frames. So that um and anam example Fifi Le and who’s the basically the u the mother of um of uh image processing today in in most cases when they’re learning uh the AI can get confused between a man on a horse and a statue of a man on the horse in Trafalgar Square. But human beings can tell instantly that those are two separate things. And a one-year-old who’s a toddler can tell the difference between a live bunny rabbit and a toy
00:23:47
bunny rabbit. They know that instantly. And AI is still confused by that. So, so we are at the beginning of although we’re trained on all the knowledge and Apple had a great paper on this, the illusion of thinking. In fact, we’re at the beginning of real thinking. Now, you can rationale and use probabilities to arrive at the right answer. But when the AI can’t tell the difference between like a husband and ex-husband because probabilistically those are very similar in the way they occur within language.
00:24:20
Those those are completely different distinctions. And if you can’t tell the difference between Mustang a car and Mustang a horse, then you’re in separate. You’re confusing the domains. And and to some extent, that’s where these hallucinations come from and that’s where some of the creativity comes from. But in the end, we’re moving forward to thinking. So what we have to do is train on all the content and then start working on how human beings use knowledge and logic to reason based upon
00:24:51
the frames of content that they know. >> And and there’s another point here as well, isn’t there Jim? Which is we actually don’t want a lot of data to keep on moving around. We don’t necessarily want data to go to these data centers. If we are going to think about this, we want to keep data as close as possible because a lot of people don’t actually understand that it costs money to move data around. They don’t think about the mechanics involved in all of this. And so what we actually
00:25:24
do want is at some point to have an artificial intelligence chip. We want to h have a lot of the operations done locally, don’t we? >> Yes, I I believe that that’s uh a good a good point. Um you want and not only do you want it to not be moving around, you don’t want the meaning of the data to be moving around. So for example, since LLMs are based on probabilistic learning, deterministic learning is like the definition in science like 2 + 2 = 4 or H2O is a water molecule. You don’t want
00:26:03
to learn that H2O is a is a water molecule. You have to know that it’s a molecule and you never want the data behind the definition of H2O to change. You want to make sure that that kind of knowledge which is know which is actual knowledge. It’s not learning what an H2O model is. It’s knowing what an H2 model is. You don’t want that to shift. Right? Because all of our science, math and science and chemistry and biology, there are certain laws in the universe, physical laws, um, biological rules and
00:26:42
the science behind it that you don’t want the AI to learn incorrectly or guess at the answer. And even in our data, for example, in data security or identity, you don’t want to be 80% confident that it’s someone’s American Express card. If you’re looking for an express card number, you want to be 100%. So, you have to verify the data. And in the last um Neur uh uh conference that just happened uh this fall, a big discussion was around the concept of knowledge graphs and bringing knowledge
00:27:16
and knowledge graphs into the um the AI alchemy, if you will. And so, Google, as you know, built the world’s largest knowledge graph. So they’re sitting on quite an amazing um uh quite an amazing asset. And then now you’re seeing this notion of moving toward um recognizing smaller patterns with uh language chips uh in the form of inferencing. So again, we’re going back to the basics of the human brain. The brain stores knowledge and it makes inferences from the knowledge that it has. Let’s come up
00:27:54
with an analogy for that because it’s a little like if I walk through a town, I just really need to know where I’m going to. But what what happens with um a lot of computer systems is they know everything about the town. They’re holding all of that information. They know this the room plans of all of the other rooms. There’s a lot of not of information that they don’t need. And that’s really what we’re moving towards, isn’t it? To actually try to find the date the data that we want and not
00:28:28
actually get all of that data moving at the same time. >> Yes. And that’s a great metaphor to move from A to B. You have to know where you are and you have to know where you’re going. And then you have to understand the pathways to get there, the potential options, and find the fastest path from A to B or the most scenic one or the one that has the most restaurants. And so there’s a lot going on in our human brains in terms of uh framing and referencing. And so that’s what’s
00:28:59
missing from the AI. Just knowing uh language probabilities is a good way to understand the output of thinking, but it’s not the actual thinking. >> Well, we’re not modeling the thinking. We’re modeling the output of thinking, which is a good start. And we need the inferencing. We need logic. We need the knowledge represented in the computers to take it to the next level that starts to approach how humans actually reason and think. And of course, we learn. And the more we learn, the more knowledge we
00:29:29
put into our brains. So we’re learning as we go. And we’re also learning when we make mistakes. So if you’re moving from A to B and you move to a dead end in London, you’re going to turn around and go back and retrace your path and then continue to move on. So it’s a very good s simple way of understanding um >> Jim where we are. I think all of this from the um knowledge graphs to the uh AI chips to all the different dimensions are going to accelerate and move AI forward is something that we’re going to
00:30:03
be exploring in great depth and in great detail in a number of the programs that we have planned. But obviously the the the whole hire thing was uh oriented around some of the talent and there is a massive um battle for talent at the moment. Um and interestingly one of the tactics used to attract talent at the moment has been soup. Uh, and I don’t know if you saw this, but um, uh, Mark Zuckerberg has actually uh, used a a tactic all his own to uh, uh, seek to approach a number of the key characters
00:30:44
that he’s targeted for recruitment in the AI space by making his own soup by hand and then handd delivering the soup to some of these candidates. Uh guys, I don’t know. Have you been approached with soup by Mr. Zuckerbergs? How would you react if uh he approached you with some of his homemade soup? >> Well, um that is an interesting uh and uh if you look at it from a different perspective, fairly insightful metaphor to be used to be using in respect to Mark Zuckerberg, God bless his little
00:31:23
heart. the um the sum total of Facebook uh on my um on the side of what is happening uh in the way we’ve um reduced critical thinking and how social media has actually um reduced uh the ability for public discourse to be meaningful and to be fact-based has resulted in actually um some political views that are causing a rise of soup kitchens in America. And I think that the the side of me that is more cynical is that is actually one of the worst metaphors to use to approach people to say, “Hey, you
00:32:06
can use some of my soup and come work for me.” Given that Facebook has um really caused a lot of the political chaos that’s in our country at the moment. So um as political decisions are being made and as people are unable to afford the basics in our nation and the and including food and the subsidies including the elimination of the SLA SNAP program which they refuse to fund uh it’s actually a form of projection. I believe it seems like a great idea, but in the end I I think analysts are going to tear it apart
00:32:45
given the amount of people that who are not in the tech industry that that are in soup lines this uh season. So >> but Pete, but but Jim, it is a bit of a recognition though, isn’t it? Apparently there the talent pool that people are after is microscopic. Apparently, there are only about a thousand people in Silicon Valley who are top AI caliber people. Those are the ones you’re after. So, yeah. And I I suppose you could say, yeah, that that that soup is a starter. You could say it’s an entree. You’re
00:33:21
saying, “Hey, this is how I want to begin my my my relationship with you.” I mean, Bill has been thinking that probably the soup of the day at the moment would be turkey soup as everybody tries to get rid of all of those leftovers. But I mean, this is you were talking about it being necessary to create these relationships with with people and particularly people who are working in in these areas. That’s what the soup’s intended for, isn’t it? It’s it’s not it’s it’s not a light-hearted soup.
00:33:54
Well, there’s um I I wouldn’t want to spend too much time thinking about the different ways that this is going to be viewed, but I appreciate the sentiment that by giving you some food um appealing to you to consider uh talking to us. Yeah. AI uh talent AI talent and talented AI um scientists and algorithmic u developers are very few and hard between. They’re extremely rare and I appreciate the sentiment. However, the um political overlap of uh the realities in today’s society outside of
00:34:31
Silicon Valley, I think it just reflects the bubble nature of Silicon Valley. We still operate in this uh world where everything is fun and games. Um, and it require it requires us being a lot more responsible about how we treat others and how we treat the effects of AI and society. And certainly it’s going to end up filtering down to things as basic as food, shelter, and uh, income. Mm-m. >> But Jim, Jim, one of the things that we want to do and we’re very keen on doing is hearing more from you and and and of
00:35:09
you becoming our Silicon Valley correspondent. So, are you saying to me that I’m not going to be welcome on your doorstep with a bowl of Stilton and broccoli soup? >> Well, I guess I’ll depends on the soup. Is it How good is it? >> Well, I’m Italian. I’m Italian, so it’s gonna I have a high bar for how soup is going to be. >> Ministroni. Minestroni. Jim. I love ministr. I’ll come around with some ministr. >> Well, guys, thank you so much for joining us and uh maybe offline we can
00:35:43
exchange soup recipes. And we look forward to following up this because as this acquisition and number of potential other acquisitions unfold as we go into the new year, I think there’s going to be plenty of foder for us to discuss here and onwards. Um I I’m I’m going to have to rush now cuz somebody’s at the door. It might be Zuckerberg delivering soup or something. So um I’m going to have to leave you here guys. Thank you for your contributions in 2025 and we look forward to a great year ahead and
00:36:16
discussing some of these stories and how they unfold as we move ahead. >> Yeah. And we can compare soups. Hopefully all of us will get a small paint. >> Exactly. >> Thank you guys. >> Thank you so much. Happy holidays and happy new year.
