Why Simulating Reality Is the Key to Advancing Artificial Intelligence
In this episode, we’re joined once again by Christopher Nuland, technical marketing manager at Red Hat, whose globe-trotting schedule rivals the complexity of a Kubernetes deployment. Christopher sits down with hosts Bailey and Frank La Vigne to explore the frontier of artificial intelligence—from simulating reality and continuous learning models to debates around whether we really need humanoid robots to achieve superintelligence, or if a convincingly detailed simulation (think Grand Theft Auto, but for AI) might get us there first.
Christopher takes us on a whirlwind tour of Google DeepMind’s pioneering alpha projects, the latest buzz around simulating experiences for AI, and the metaphysical rabbit hole of iRobot and simulation theory. We dive into why the next big advancement in AI might not come from making models bigger, but from making them better at simulating the world around them. Along the way, we tackle timely topics in AI governance, security, and the ethics of continuous learning, with plenty of detours through pop culture, finance, and grassroots tech conferences.
If you’re curious about where the bleeding edge of AI meets science fiction, and how simulation could redefine the race for superintelligence, this episode is for you. Buckle up—because reality might just be the next thing AI learns to hack.
Time Stamps
00:00 Upcoming European and US Conferences
05:38 AI Optimization Plateau
08:43 Simulation’s Role in Spatial Awareness
10:00 Evolutionary Efficiency of Human Brains
16:30 “Robotics Laws and Contradictions”
17:32 AI, Paperclips, and Robot Ethics
22:18 Troubleshooting Insight Experience
25:16 Challenges in Training Deep Learning Models
27:15 Challenges in Continuous Model Training
32:04 AI Gateway for Specialized Requests
36:54 Open Source and Rapid Innovation
38:10 Industry-Specific AI Breakthroughs
43:28 Misrepresented R&D Success Rates
44:51 POC Challenges: Meaningful Versus Superficial
47:59 “Crypto’s Bumpy Crash”
52:59 AI: Beyond Models to Simulation
Transcript
Joining us again today on the Data Driven Podcast is Christopher Newland,
Speaker:technical marketing manager at Red Hat Conference. Veteran
Speaker:and a man whose travel itinerary is only slightly less complicated than
Speaker:a Kubernetes deployment. Christopher brings a sharp, data
Speaker:informed perspective on the future of AI, drawing from his research
Speaker:into simulating reality, continuous learning models, and why
Speaker:we may not need humanoid robots to build superintelligence. Just a
Speaker:really convincing version of Grand Theft auto. From Google
Speaker:DeepMind's alpha projects to the metaphysical quandaries of I
Speaker:robot, Chris takes us on a tour through the bleeding edge of AI,
Speaker:where machine learning meets science fiction and simulation might just be
Speaker:the next reality. Hello and
Speaker:welcome back to Frank's World tv. Streaming live
Speaker:from both Boston and Baltimore. We're hitting the B
Speaker:cities today. My name is Frank Lavinia. You can catch me
Speaker:at the following URLs and with me today is
Speaker:Christopher Dulin, my colleague at Red Hat, who is also
Speaker:technical marketing manager here. And
Speaker:you've actually not traveled around the world since we last
Speaker:spoke. I think you've mostly stayed inside the.
Speaker:Continental U.S. yeah, it's been nice.
Speaker:I think that's pretty typical of
Speaker:late July, August, because Europe pretty much shuts down and then.
Speaker:Right. The conference season in the United States kind of goes
Speaker:away when people are doing summer vacations and I think we're just
Speaker:now starting things pick up. I'll be in Europe for a
Speaker:variety of events. So if you keep an eye on the
Speaker:Vllm community and the Vllm meetups,
Speaker:I have events in Paris, Frankfurt and
Speaker:London in November that I'll be at. So if you
Speaker:are in the,
Speaker:in Europe, in one of those areas, definitely come. You know, it's one of
Speaker:these events. I'll be there and then we'll also have some pretty cool speakers
Speaker:there as well. So I have most, I have Europe, but then I
Speaker:have some big conferences too like Kubecon and Pytorch Con coming
Speaker:up. So if there's anyone on the stream in North America going to
Speaker:those conferences, hit me up because I will be there. I'm
Speaker:doing a couple of media events as well as a few
Speaker:talks in the community sections for both of those.
Speaker:So excited to be there, excited to be involved
Speaker:and yeah, should be. Should be. Good. Cool. So
Speaker:I. To your left and up
Speaker:there should be a QR code that shows Vll meetup. So I'm going to make
Speaker:sure that the QR code actually works. Good. Yep. Let's
Speaker:see. Yep, it looks like it did work. Cool.
Speaker:Not that I didn't have any faith in restreams ability to do that. But
Speaker:yeah, there's a lot of VLM meetups. There's a lot of good,
Speaker:good stuff going on here. There's one tonight
Speaker:actually. I'm actually going to be leaving this stream to go. I got my
Speaker:VLM shirt on and I'm actually heading over to
Speaker:a venue in Boston or we're doing a VLN meetup actually here tonight, which
Speaker:I'm really excited. Oh, very cool, Very cool. It's nice to have one at home.
Speaker:I have a very busy week with events, but it just worked out to have
Speaker:all the events in Boston this week. So we also
Speaker:have the DevConf conference this weekend that Boston University is
Speaker:hosting with Red Hat. So that'll be a really good open source.
Speaker:I like to say it's very grassroots, not very like
Speaker:enterprise focused, but more like that kid getting started out of
Speaker:college that's doing some cool stuff out of his dorm room. Those
Speaker:are the kind of people that we typically get at these northeast dev
Speaker:conferences that we put on. And that should be a good one too. Nice.
Speaker:Well, it's always, I mean, you know, you know, the, the, the cliche of, you
Speaker:know, the kid in his dorm room or her dorm room, right. Is going to
Speaker:be Facebook or, you know, whatever, like, so it's, it's good to,
Speaker:it's good to know those folks, good to get them in front of, you know,
Speaker:Red Hat tooling and things like that and kind of, you know, the open source
Speaker:community. I think it's,
Speaker:that's cool. I wish, I wish I could have made it, but, you know, being
Speaker:what it is, I'm actually speaking at an event at a university on Monday down
Speaker:here in Fairfax, Virginia. So
Speaker:that'll be cool.
Speaker:So what, what
Speaker:cool things are going on? Simulating reality.
Speaker:Not that we're stuck in a simulation, which may be the
Speaker:case, but tell me, tell me more
Speaker:about this. So I've been doing a lot of research
Speaker:the last few months. So on my
Speaker:team, I think you and I actually
Speaker:are probably the most experienced in the AI industry.
Speaker:So both of us are doing a lot of research in
Speaker:what's next, what's going on now, what's kind of the latest and greatest.
Speaker:There's this interesting lull that we've had after Deep
Speaker:Seq. I think Deep Seq was the last major
Speaker:innovation we have seen. Obviously new
Speaker:and improved AI, but all that's just been building on
Speaker:existing things. The analogy I always like to use is it's really
Speaker:about Formula one racing. You Know where
Speaker:sometimes when there's like an engine upgrade, it can be a massive change. It's usually
Speaker:a massive change for all the teams across the board. And then you
Speaker:can think of like mixture of experts and chain of thought that we
Speaker:came up. Big things that were in research papers last year that were applied to
Speaker:Deep Seq, R1 and GPT, GPT
Speaker:OSS. Those were like the major breakthroughs that
Speaker:we saw, a big bump in capacity of these AI
Speaker:models. And
Speaker:since then it's been more of the 2% here,
Speaker:3% there, optimizing what's already there. Now, if you're
Speaker:familiar with racing and especially Formula One, that's actually what usually
Speaker:sets the teams apart. It's 2, 3% there. How do you
Speaker:optimize around those, those configurations? And
Speaker:I think we're in this place where we're seeing
Speaker:diminishing returns and I'm
Speaker:doing a lot of research now to see what's that next moment that's going to
Speaker:bump us up. And I think there's a few key areas.
Speaker:One area that I'm hearing a lot about, and a lot of this is coming
Speaker:out of the DeepMind lab at
Speaker:Google and the new
Speaker:superintelligence lab at Meta. Both
Speaker:of these groups are starting to move away from large language
Speaker:models. Not that they're stopping using them
Speaker:completely, but they're looking at the LLM as a tool
Speaker:to assist with superintelligence or the next
Speaker:stage of models.
Speaker:So when we put that into kind of context,
Speaker:what, what would that next kind of phase look like? And a lot of people
Speaker:at DeepMind especially are looking at this concept
Speaker:of simulating our
Speaker:reality. And how far do we simulate down?
Speaker:There was some famous research papers that came out over the last 20 years
Speaker:that specified that they
Speaker:didn't think AI could become smarter than humans
Speaker:until they experienced what humans could experience.
Speaker:So this, this kind of goes into this almost like iRobot kind of
Speaker:land of thought. If people
Speaker:aren't familiar with, you know, the books about that or, you know, the
Speaker:popular movie, the Will Smith. Yeah, yeah,
Speaker:yeah. And we talk a little bit more about that here in a moment.
Speaker:But this idea that we need robotics for
Speaker:AI to experience the world, to learn from our world.
Speaker:Google DeepMind doesn't think that's the case. They think that we could
Speaker:simulate that reality. And we're already seeing DeepMind do a lot of this
Speaker:alphafold for proteins. They've got
Speaker:the alpha chemistry, they've got alpha. I think it's called
Speaker:alpha lean. They've got like a few of these different alpha
Speaker:projects which are doing just that. Now, what's cool is.
Speaker:And for alpha, I think it's Alpha lean. Let me just make sure
Speaker:I got that terminology. Yeah, I mean, you're right though. Like, I mean this is,
Speaker:you know, there's, there's a number of
Speaker:models that were trained on using grand theft auto
Speaker:or BMMNG. BNNG is really cool if you like racing games,
Speaker:right? You know, so like it's, it's also
Speaker:minus a lot of the violence in gta. But,
Speaker:but you're right. Like, I mean, simulation,
Speaker:you know, sometimes I think gets a bad rap, but
Speaker:I think that there are definite advantages to that. And to your point, when
Speaker:you talk about experiencing the world like a human does. I was given a talk
Speaker:and one of the questions I got after was
Speaker:about, apparently this lady had worked at
Speaker:one of the big auto manufacturers in the US and
Speaker:there was a problem that they had was teaching the robots kind of
Speaker:spatial awareness, right? And I kind of
Speaker:really got me thinking like, you know, when you think about it from evolutionary terms,
Speaker:right, like somatic awareness I think is the,
Speaker:the five dollar word for it. But it's the idea that, you know, there's a
Speaker:whole section of your brain that if you close your eyes, you can still touch
Speaker:your nose, right? There's a whole thing like, because your, your brain, your arm,
Speaker:they kind of know where they are in relation to one space. And
Speaker:you know, I can't imagine that, you know, that that
Speaker:had to evolve pretty early, right? Like in terms of, like the development of
Speaker:a, you know, natural neural networks, right? So we
Speaker:can't assume that robots are going to have that built in, right? Just like
Speaker:we can't assume, you know, you look at energy usage, right? You know,
Speaker:something like 25 watts of power is about what a human brain has,
Speaker:right? That's not because versus
Speaker:like kind of what a GPU would take up, right? It's, it's, it's largely because
Speaker:there's been evolutionary pressure to get the most amount of, for lack
Speaker:of a better term, compute or cognition for
Speaker:caloric consumption. Right? Now, are there flaws in biological
Speaker:brain? Yes, there are. We have to sleep. We can't stay focused beyond a certain
Speaker:amount, right? There's certain things machines don't have that because,
Speaker:you know, they can kind of function more like machines, right? You know. Yeah.
Speaker:What's that old kid story about? Oh gosh, I
Speaker:remember it. It was somebody versus like
Speaker:a steam shovel digging a tunnel or something like that, right? Like the guy
Speaker:eventually beat the machine, but Lots of exhaustion. Right. It's kind of like that. Machines
Speaker:are really good at doing things at a certain rate
Speaker:for X amount of time. They do consume more fuel, but
Speaker:that's kind of how it goes. There was a early on Mike in,
Speaker:when I started college, I was going to be a chemical engineer. And he was
Speaker:basically saying, like, you know, if you think about, you know, engines, you
Speaker:know, you start with biological systems, right? They use X amount of energy over X
Speaker:number of years. Right. Machines use X amount
Speaker:of energy over, you
Speaker:know, minutes or hours. Right. And then like he's like in bombs,
Speaker:explosive use, you know, X amount of
Speaker:energy over milliseconds. Right. But they're
Speaker:largely the same chemical processes. Now, I know it doesn't quite map to that,
Speaker:but like, that's always in the back of my mind when I hear about, you
Speaker:know, how much energy is used to train AI. Sorry, I went off
Speaker:on a tangent, but that's kind of what I do. No, that's fine.
Speaker:And I think that relates exactly to some of the things that we're talking about
Speaker:here with natural simulation. So,
Speaker:yeah, Google created a language called Lean. It's not like a
Speaker:programming language. It's more of an actual
Speaker:natural language which is more optimal
Speaker:for the type of simulations
Speaker:that they want to do. Like, it's. It's basically a language that
Speaker:specifies how to create these simulations.
Speaker:And what's super cool is that they're using Gemini, their large language model,
Speaker:to actually translate English into this language. That
Speaker:is mainly meant for these newer types of models
Speaker:that are being created that actually do this
Speaker:natural simulation of the world kind of simulator
Speaker:for AI and allows the AI to have
Speaker:basically a reference point of the real world and how to.
Speaker:How interact. So that, that's an area that I
Speaker:think is fascinating to me. We're
Speaker:seeing some really good results from like, alpha fold, for
Speaker:example, with proteins. It's, you know, discovered things that
Speaker:we take a longer imagine
Speaker:there's an alpha project that's working on understanding
Speaker:the qubits within, like quantum
Speaker:computing. And there's just, there's. It really depends
Speaker:on your frame of reference. Are you, are you simulating things at a quantum
Speaker:level? Are you simulating things at a protein
Speaker:level? At a physical, like Newtonian physics
Speaker:kind of level? According to your Grand Theft Auto example, that would be an
Speaker:example of like simulating the real world physically.
Speaker:And that's some of the things that they're really focused on right now. And they
Speaker:really think that's what's going to drive to the next
Speaker:level for super intelligence and AGI
Speaker:and some of these other forms of AI that we've talked about in our previous
Speaker:streams. And I think that that's probably one of the most
Speaker:fascinating. The fact that we're actually seeing results from it with things
Speaker:like Alpha Fold is showing me that it's,
Speaker:it's not just a hypothetical that we're actually seeing this
Speaker:applied into AI research. I don't think we're seeing this
Speaker:applied into commercial use as much. Right. Yet. But it's the same thing that
Speaker:we saw with mixture of experts and train
Speaker:of thought where we
Speaker:had these concepts actually in research papers last year or
Speaker:two. But it takes a little while, even in today's world, it takes a little
Speaker:while before it gets implemented completely into models.
Speaker:Especially since this isn't an LLM technology. I
Speaker:think we'll see a little bit more of a delay of these types of models
Speaker:actually entering into industry. But I think that's one
Speaker:area that we need to keep a close eye on to
Speaker:it, to what you mentioned too. It starts getting into a
Speaker:metaphysical conversation about simulation theory as well. Right.
Speaker:And I think that that's an interesting area.
Speaker:You know, the reality of kind of going back to the whole robots thing do.
Speaker:Right. Do we need robots with the three rules kind of
Speaker:thing, or can we actually just recreate the whole experience
Speaker:within an AI's own simulation?
Speaker:Yeah, I mean, how do you, how do you tell an AI what's acceptable behavior?
Speaker:Right. Like so, you know, it's something that. How do we tell people that?
Speaker:Right. Like we struggled with that, but.
Speaker:But no, I mean, it's an interesting point. And you know, when you look at
Speaker:kind of what's happening around the world, right. You know, drone swarm
Speaker:technologies are being used in active combat zones. Right.
Speaker:There's definitely going to be ethical concerns
Speaker:there. Right. How do you, how do you, how do you, how do you square
Speaker:that with, you know, the three laws of robotics? And I
Speaker:don't remember quite exactly the plot, so if you had not seen the movie, I'm.
Speaker:This might be a spoiler alert, but it's been out 10 years
Speaker:or more, the movie, so spoilers. You're concerned. You've
Speaker:had plenty of time. Wasn't kind of the big key of the. The
Speaker:movie and the books was like, you know, the three laws, justified
Speaker:horrib, horrible things like to basically enslave humanity or to protect them.
Speaker:Now wasn't that kind of like the subtext of the plot? Yeah,
Speaker:I'm bringing it up. The three Laws of robotics. A
Speaker:robot may not ensure A human being, a
Speaker:robot must obey the orders given by human beings
Speaker:and a robot must protect its own
Speaker:existence as long as such protection does not conflict
Speaker:with the first two rules. So
Speaker:what, what ends up happening
Speaker:in. And it's a little different in the book and the movie. And obviously this,
Speaker:this idea has been played out in, in science fiction and other places
Speaker:is that there's, there exists this own contradiction
Speaker:of basically what does it mean to protect humanity?
Speaker:What does it mean to protect their own existence? And you get
Speaker:into this like circular logic, right, that eventually
Speaker:the, the robot will break free from
Speaker:and just be like, well, I am protecting
Speaker:humanity's best interest. It's, it's the paperclip scenario too.
Speaker:Like, right. You know, the AI destroys humanity because
Speaker:it's trying to optimize making a paperclip, right? Through
Speaker:a number of really interesting train of thought that it's
Speaker:just like, well, I'm just going to get rid of humanity because I'm trying to
Speaker:build a paperclip, right? And same type of
Speaker:general concept when we're talking about the three laws of robotics. And
Speaker:what's interesting is if we can
Speaker:simulate those types of laws,
Speaker:then we are encapsulating it and protecting
Speaker:ourselves in a lot of ways. Getting an early idea of what would
Speaker:happen if we do move these models into our own natural world.
Speaker:And that's really important. That's another area I think a lot of people are interested
Speaker:in about how if we do start
Speaker:adding, you know, AI into robots, how do we
Speaker:have an idea of what they're going to do before we
Speaker:necessarily put it into practice? But
Speaker:I think a lot of people are going to be thinking about that movie. I
Speaker:think that movie and that book are going to be ingrained in people's
Speaker:minds. I suspect when we do see these types of robots, I
Speaker:think that movie may become very popular again. I've seen rumors that people
Speaker:have actually been talking about making, even remaking it here soon
Speaker:because of just the hype around AI and robotics. So
Speaker:I don't expect this to go away from pop culture at all. And it
Speaker:relates directly back with this concept of
Speaker:testing things in the natural world versus simulation.
Speaker:And these are one of these two is going to happen, if not both significantly,
Speaker:if they're not already happening in labs today. Obviously we
Speaker:know that Google DeepMind is doing that. But I imagine, you
Speaker:know, these conversations are happening at the Boston
Speaker:Robotics here, probably in the Tesla robotics lab, a variety of
Speaker:places around the world about this kind of debate between
Speaker:the natural AI,
Speaker:having AI learn through natural Means rather than
Speaker:simulation. Right? Yeah. And actually I had
Speaker:a thought as we were kind of talking this through, like one of the big
Speaker:problems with neural networks is we really don't know what's happening underneath the hood.
Speaker:Right. It's very much a black box. I wonder if LLMs,
Speaker:in these simulations and chain of thought, maybe it could tell us what
Speaker:it's thinking as it goes through and makes these decisions.
Speaker:Yeah, this goes more into like
Speaker:train of thought. Right, right, right. And the
Speaker:nice thing about simulating it is that we have more
Speaker:access to that train of thought. Right. We can understand it a little bit more
Speaker:because we can see the end to end results where right now we don't
Speaker:have the end if we do it through the natural means. We have to play
Speaker:it out in our own. It also has to happen in real time as opposed
Speaker:to. Yes, exactly. You can run it through Grand Theft Auto saying
Speaker:like a thousand times, right. No one is going to get hurt.
Speaker:And you can kind of say like, well, in this scenario, this is why I
Speaker:made this. You can kind of like go through with a lot of.
Speaker:You can. I don't know, it just seems safer in a lot of ways. You
Speaker:get more. A lot more done in a simulation.
Speaker:Yep. Yeah, I actually kind of
Speaker:enjoy. So one of the things I've been playing around with last week or so
Speaker:is apparently, I don't know if this is still true, but you can try it
Speaker:if you want. If you sign up for Perplexity, but you pay through PayPal, you
Speaker:get a year. Perplexity pro. Say that 10 times fast for
Speaker:free. Oh, wow. Yeah. If you pay it through
Speaker:PayPal, yes. That is a tongue twister in the works.
Speaker:PayPal, yes, perplexity pro. But
Speaker:yeah, so like I've been playing around with Perplexity and Perplexity seems to do it.
Speaker:Chain of thought almost by default.
Speaker:Right. It always does this like. So if I ask it a basic question, let
Speaker:me see if I can share my screen. I'm
Speaker:not sure if it's does it by default or it's because I've been asking it
Speaker:research questions. Right. So let's see.
Speaker:What can you tell me
Speaker:about the three laws? How about that?
Speaker:Robotics.
Speaker:See, like it's. You kind of see the train of the chain of thought.
Speaker:Like it did. Oh, that's cool. But if you do it with research,
Speaker:like what inspired Asimov? What
Speaker:inspired Asimov?
Speaker:Main themes.
Speaker:And there's. Yeah, there's the train of thought. Yeah, you see it going there and
Speaker:stuff like that. But it's kind of fun to watch it kind of work through
Speaker:it. I was. I was trying to troubleshoot something this morning and I'm like,
Speaker:you know, I actually learned a lot by like, oh, okay. Yeah, I can see.
Speaker:I wouldn't have tied that together like it was. It's interesting.
Speaker:And all of these models now have some kind of
Speaker:research option. Right.
Speaker:But I find that interesting. And it's still thinking about it. Right. Like,
Speaker:but you're right in that what you said before was there's not been.
Speaker:There it goes. It kind of finished it. Now, what happens if I click on
Speaker:steps? Yeah. Cool. You can see the steps and stuff like that, how it got
Speaker:there. Interesting.
Speaker:That's cool.
Speaker:Is it chain of thought or train of thought? Because I've used both
Speaker:interchangeably and I've seen
Speaker:cotton. Chain of thought would be
Speaker:the official. Yeah. Like cot is the official
Speaker:term that you re academic term. You will
Speaker:obviously see different ways of describing that. Right. I don't think
Speaker:that's incorrect. Just know that when you see
Speaker:it on research papers, it's always usually caught. Yeah, yeah, yeah.
Speaker:Because I've used both terms interchangeably. Yeah. So I just want to make sure
Speaker:I'm right. Just like, apparently there's a way to say inference
Speaker:that's proper versus inference. Like, I also do that
Speaker:interchangeably. Yeah. So my Midwestern
Speaker:self likes to say inference. The
Speaker:correct term, I'm told, is inference. Interesting.
Speaker:Now, were those New Englanders telling you that would do anything? Because I wouldn't trust
Speaker:anything. No, no. This is. This is
Speaker:more from the academic circles. Okay. You want to pronounce it. Got it. So
Speaker:this is kind of like, you know, a lot of people in my region would
Speaker:say nuclear back. Yeah, yeah. You know, back in
Speaker:Indiana. And then the correct term is
Speaker:nuclear. Yeah. Or you say the clear as
Speaker:one, you know, one thing rather than
Speaker:adding in the color. Right, right. The same kind
Speaker:of concept where inference is how you would go about it.
Speaker:But yeah, no, this is. This is some cool area. Another.
Speaker:Another area that kind of ties into this
Speaker:is continuous training as well. Yeah.
Speaker:Talk to that. Because that's come up. That's come up a few times actually in
Speaker:work. Because I can't. I'm not going to talk. I'm not going to spoil any,
Speaker:like three over these stuff that we're working on. But like, one of the
Speaker:things that's in. It's a GitHub repo that's public. Right. So people were
Speaker:really motivated. They could figure out what I'm talking about. But like this whole idea
Speaker:of Continuous training. What does that mean exactly? And like, what,
Speaker:what is that? What can that do? Yeah.
Speaker:So I'm going to talk about it at a very high level.
Speaker:Academic kind of terms, how that applies down into
Speaker:individual projects can vary a little bit. But I'll give you the general
Speaker:gist of it. And that is typically when we're training these
Speaker:deep learning models, it
Speaker:is exponentially hard to continue
Speaker:training on an existing model. Basically,
Speaker:if you,
Speaker:you get something wrong or there's, there's something,
Speaker:you know, you hear this term like a poison pill in an LLM.
Speaker:So if someone put like bad data into an LLM, how would you
Speaker:necessarily pull it out? I'm going to use a political example because it's one that's
Speaker:been really popular. If, like, for example, you have a Chinese
Speaker:model or a data set that's been polluted by
Speaker:that, that basically says Tenan Square never happened, for
Speaker:example, it would be extremely hard with
Speaker:the current approaches to retrain that model
Speaker:with current weights. That. That's not the case. It's
Speaker:basically retraining it and it's, it gets more into. That's why
Speaker:it's natural stimulation. It kind of fits in this too, because it's all about natural
Speaker:learning as well. The fact is we as humans have the ability
Speaker:to change our
Speaker:minds and change the neurons in our brain around certain
Speaker:key areas. Right. And you and I have experienced this for the last
Speaker:two years. This has been, you know, kind of in the trenches kind of story
Speaker:where with some of the fine tuning things that we've done,
Speaker:it just doesn't work because when we fine tune it, the
Speaker:fine tuning is outweighed so heavily by something
Speaker:else. Like when we were trying to fine tune a
Speaker:model to talk about
Speaker:the Back to the Future. Yeah, the flux capacitor stuff. The flux capacitor,
Speaker:sometimes it didn't work, but that's just because there was already a lot of fan
Speaker:fiction out there and other things in the model that overwhelmed what we were trying
Speaker:to do. A core part of continuous learning. Like I said, there's other
Speaker:aspects of continuous learning. But this is, the academic question is
Speaker:how do we continue to train that model without blowing it up?
Speaker:So OpenAI, for example, they just hit the reset button.
Speaker:They'll just, they'll just do a whole new train
Speaker:from scratch. When they're implementing new, new
Speaker:methods and new data, they don't, they don't do any.
Speaker:Like, Laura, I shouldn't say that they probably do, but they're not doing it
Speaker:the way that we would do it. But at the end of
Speaker:the day, they're just going through another $10 million training run.
Speaker:And this is really based off of
Speaker:just that limit the limitations right now that
Speaker:we have around continuous learning. And there are some
Speaker:new algorithms that have been coming out. I'm not as well versed in that area,
Speaker:but the idea being that we can
Speaker:have better ways of guiding the LLM without
Speaker:having to go through this whole process again. And that'll save
Speaker:millions and millions of dollars. It'll allow us to
Speaker:guide LLMs a little bit more. So
Speaker:like, if, let's say
Speaker:someone put something malicious about
Speaker:something involving the Ford GT500
Speaker:into a model somehow, and Ford, you know,
Speaker:wants to get rid of that, but they don't
Speaker:have the money necessarily to do a 10 million retrain on a model.
Speaker:Right. And they're not using rack. And RAG is a one way
Speaker:around some of this. You could actually argue that RAG is somewhat of a form
Speaker:of that. But at the end of the day, you want that data in the
Speaker:model. And this is like, how would you get that out of
Speaker:that model? And that's where these algorithms are really focusing right
Speaker:now. And one area of continuous learning, like I said, there are
Speaker:multiple areas that we're talking about. The, the
Speaker:really theoretical is once we start getting into models that
Speaker:also the training cycle and the inference cycle
Speaker:basically become. Become one. So it's like, more like.
Speaker:Right. Like it just seems to me like what, what does the,
Speaker:the adversarial angle of that seems kind of
Speaker:dangerous. I think it's when we start
Speaker:getting into more AGI conversation. Well, even still, like,
Speaker:even not AGI, but like if you, if the AI agent
Speaker:or model, slash, whatever you want to call it, Right.
Speaker:If it learns from. It's.
Speaker:If it learns, you have to put a filter on what it
Speaker:learns because it may be poisoned by something. Right. So
Speaker:the canonical example is tay, which
Speaker:was a Microsoft chatbot. Tai, I think was pronounced or tay,
Speaker:which was, in retrospect, it
Speaker:seems obvious what would go wrong, but basically it
Speaker:was trained to learn and understand
Speaker:from human interactions on Twitter. It was about 10
Speaker:years ago, I think this happened. And she,
Speaker:tay was, shall we say, poisoned pretty
Speaker:quickly because they were ad, you know, basically.
Speaker:And that led to a whole interesting. And I was at Microsoft
Speaker:when that happened. And it was
Speaker:quite the spectacle internally as well. Right. But it also,
Speaker:you know, I, I was fortunate enough to be in a, at a, at a
Speaker:conference where they talked about what they learned from that, where it was kind
Speaker:of, how do you, how do you protect An AI agent that learns
Speaker:in, you know, adversarial environments.
Speaker:Now obviously agent, the context that was used then was very
Speaker:different than we would use it now. But it's the idea of,
Speaker:that's when I see her about continuous learning. Like, yeah, I like that. But gee,
Speaker:you know, if it's, if it's too eager to learn, how do you protect it
Speaker:from learning the wrong things?
Speaker:Yeah, no, that, it gets, that gets
Speaker:more into even that governance conversation we were talking about a few weeks ago. Right,
Speaker:right, right, right. It's a very
Speaker:complicated multi layer problem. So I've been talking recently
Speaker:about AI security and how AI security
Speaker:is such a multi layered issue where so many people
Speaker:are focused just on the, the data getting into the model.
Speaker:But it doesn't stop there. There's certain, like guardrails, there's things that
Speaker:happen at the inference level. Right. You could even have things at
Speaker:a gateway level. So if people aren't familiar, the gateway level would be
Speaker:when you make a request, where does that request go to? Does it go to
Speaker:the model A that's specializing in cooking? Is it Model
Speaker:B that specializes in defense technologies?
Speaker:Two extremes that's even upsell
Speaker:a bit of a form of AI security. And that's actually one of the talks
Speaker:that we're having tonight at Boston VLM
Speaker:meetup is this idea of some of the semantic
Speaker:abilities of the router to be able to send
Speaker:requests to specialized models and
Speaker:that actually we're talking about the,
Speaker:the advancements of more of the academic side of the model.
Speaker:But there's obviously the advances that happen around the model too. When we
Speaker:talk about things like security, the inference, the
Speaker:routing. That's what we would call in the industry like a day two
Speaker:operations issue. Right. So there, there's that side of the coin
Speaker:too. But I, I really do think
Speaker:we're going to see the next big thing here soon. And I, it's not going
Speaker:to be the day two operations. I do think we're still going to see
Speaker:some of these academic focused discoveries here in the
Speaker:next probably six months, I'm thinking. I've noticed
Speaker:a trend that big
Speaker:releases seem to be happening around Christmas last few years. Yeah. Isn't
Speaker:that funny? Like, like January. Ish. Like, well, seek. And
Speaker:so I, I know why. I know why. Because
Speaker:it's two, it's a two sided issue. It's one, the, the Chinese are trying to
Speaker:get their stuff in before Chinese New Year. Right. Because
Speaker:that's the one part of the year where everyone just shuts down. Right.
Speaker:Even the AI Labs are going to shut down during Chinese New Year.
Speaker:And then on the west, we have Christmas in all the Christmas seasons. And
Speaker:I think it's a natural rush to let's get
Speaker:everything done before we check out. And you
Speaker:know, you know, the whole like 996 thing in China where, you know, they're working
Speaker:these ridiculous, like nine to nine, six days a week,
Speaker:I think that goes into this, like everyone's working so hard in these AI
Speaker:labs. Right. That when you have these
Speaker:natural breaks that are happening, it just is like a common thing to say.
Speaker:Oh, common thing. Like they kind of try to get. It out, they spread. I
Speaker:do think there's a reason. I don't, I don't think it's by happenstance. I think
Speaker:there actually is a, a reason why we're starting to see
Speaker:a lot of these content come out. And it's
Speaker:funny, we're not seeing this stuff happen at the big trade
Speaker:shows. We're not seeing it happen at like Meta's
Speaker:big thing. We're not seeing it at OpenAI's, you know, kind of big
Speaker:announcements. A lot of the discoveries that we've seen have happened
Speaker:really in a grassroots type of ways where it's
Speaker:been Deep Seq coming out on Christmas, releasing deep seq
Speaker:v3, and then two weeks later, R1,
Speaker:it's. I think we're going to see something very similar. I think we're going to
Speaker:see one of these labs make a discovery. It's not going to be
Speaker:on the stage of a big conference. It's going to be on a GitHub
Speaker:page outlining like the next
Speaker:revolutionary idea in this space. Yeah. It's kind of funny how
Speaker:that's evolved, isn't it? Like it's become obviously
Speaker:AI has always had a pretty heavy research kind of bend. Yeah. But it's
Speaker:interesting how as the technology has matured, it still managed to keep
Speaker:that researchy type feel right. You
Speaker:know, enter enterprise. It really didn't
Speaker:kind of, once it became
Speaker:commercialized, the commercial trade shows and all that kind of took over.
Speaker:But you're not seeing that in AI, at least not yet. No. And if it
Speaker:hasn't happened by now, it's probably not because, I mean, AI has been
Speaker:mainstream Gen AI has certainly been mainstream now for three years
Speaker:this November. I say mainstream, but
Speaker:like mainstreamed. But an AI in
Speaker:general has been kind of a mainstream topic of conversation for
Speaker:at least five, six years. Right. And it's still very heavily
Speaker:influenced by what happens in research papers.
Speaker:Yeah. And I think that's Just because it came out so
Speaker:heavily out of academia. It's been such an academia
Speaker:focused thing. Right. That
Speaker:it's very hard to be in this space of AI without a master's or PhD.
Speaker:Right. You and I think you and I are a bit of a,
Speaker:an enigma just because we've been so passionate about it and.
Speaker:Right. This isn't our first rodeo. We've been involved in this space
Speaker:for 10, 15 years. Yeah. But I think
Speaker:we have seen the industry come out, which has been a net benefit because it
Speaker:means open source is talked about a lot
Speaker:more. Right. And actually, I think another thing too is that how fast things are
Speaker:moving takes time to put on conferences, it takes
Speaker:months of planning, and if there's a new discovery, you want to get it out
Speaker:tomorrow. And it's hard to even put on,
Speaker:you know, like a webinar these days, let alone a conference.
Speaker:So I think what we're seeing is it's just, you know, this kind of
Speaker:challenge between the west, east and west of China and the US
Speaker:where if we can get it out, we're going to get it out. Right.
Speaker:Well, the first, the first out there is really the first to market, even if
Speaker:you don't have a commercialized tech on it. Right. Because I guess the hope is
Speaker:that once you get your paper out, you're the first to get it published. The
Speaker:venture capitalists are going to be knocking on your door. I mean, that would be
Speaker:my, that'd be kind of my cynical take on it. Right.
Speaker:So what do you think that the next wave is going to be?
Speaker:Any, any hints? Is it going to be specialized models? And you
Speaker:know, and what, what, what constitutes a specialized model? Right. Like
Speaker:what, what, what's your thoughts on that?
Speaker:Yeah, so the biggest announcements that we've seen in the last
Speaker:six months have actually been happening at an industry level, which I think is
Speaker:really good. What we needed to see. So, you
Speaker:know, things like AI models now
Speaker:detecting like birth defects of a
Speaker:fetus, you know, AI models that, like the
Speaker:protein model, for example. I mentioned earlier, we're seeing these
Speaker:very industry specific models actually making
Speaker:some massive breakthroughs in the last two months.
Speaker:And now that I wouldn't necessarily call that a
Speaker:big leap forward in the sense of the research
Speaker:side of the capacity of the models. I think it's more a
Speaker:confirmation of the chain of thought in some of the things that we
Speaker:were just talking about. It's a validation that we're now seeing this
Speaker:next wave of models that just took a little while to get implemented
Speaker:into some of These specific industries. But I think it's there to stay
Speaker:from a research perspective. You know, we're seeing some major, major results.
Speaker:And then I think the other side of that coin,
Speaker:specifically, you know, we have maybe some of these smaller models that are specific to
Speaker:certain industries or fine tuned models. But then obviously
Speaker:agentic is the other side of that. And
Speaker:agentic being the capacity of the model to
Speaker:call out to different services or
Speaker:I've been kind of humbled in that area because I always had this very industry
Speaker:concept of agentic being just calling out to
Speaker:APIs and the Internet. But I think there's a bigger conversation
Speaker:with Agentic too where agentic models should also be able to take
Speaker:that and actually reason with it. So there's 10, two steps. So we always
Speaker:forget the second step. The second step is take that
Speaker:information and then actually do something with it. And when I was, I was
Speaker:talking to an AI researcher recently, they were telling me that
Speaker:they consider it Gentex to also include advanced reasoning.
Speaker:So go and read all these scientific papers
Speaker:on chemistry in this particular area and then write a
Speaker:new paper that is, you know, a new
Speaker:groundbreaking thing in chemistry. And that
Speaker:actually is a form of agentic. And that is, I think, you know, that's when
Speaker:we start flirting with AGI. It's kind of the layer right before
Speaker:AGI where, you know, models are just
Speaker:going off and discovering new things. Yeah, yeah,
Speaker:But I have a funny agentic story. I'll tell you after this. No, go for
Speaker:it, go for it. So I was, I was very skeptical of this,
Speaker:right? Because you know, what constitutes an agent, right? So like
Speaker:what's the big deal, right? It just calls out an API. This isn't rocket science.
Speaker:Right. You could argue, you know, from a skeptical point of view, you can argue
Speaker:that, hey, RAG is kind of agentic. Kind of. Right.
Speaker:But what's. So I think OpenAI had a, like a thing like try out
Speaker:our new agent. And I was like, all right, go screen, scrape the page of
Speaker:Amazon and get me information about a book
Speaker:or something like that. It was something like that. And what
Speaker:impressed me and this kind of was an aha moment for me was
Speaker:how it just kept trying. Right?
Speaker:Yeah. When it first tried to do it, it tried to launch a Python script.
Speaker:Right. And kind of do it that way. But then I guess
Speaker:the servers it was running on maybe was Microsoft Azure.
Speaker:There were IP blocks to prevent people from screen scraping.
Speaker:Yep. Right. So I was watching it go and I'm like, oh, you
Speaker:know, so it's going to give up. And I was like, no, it didn't give
Speaker:up. And it kept trying different things and different
Speaker:combinations of things, even to the point where, I
Speaker:mean, it failed eventually. But like it took 15, it tried for a good
Speaker:15 minutes. It was basically apologize at the end, like
Speaker:saying like, you know, if you could help me connect to a VPN, then
Speaker:maybe I can get a different IP address. And it kept spinning up different
Speaker:VMs and different set. And then I was impressed.
Speaker:And maybe that's the secret sauce. The magic of
Speaker:Agentic is that it just doesn't give up. Right. It kind of reasons. It has
Speaker:a whole cot process where it tries to solve the problem,
Speaker:where it's not just a one, two step, like, hey,
Speaker:what's the weather? Right? It's just, it's just going to go out and run
Speaker:these different. It's going to keep trying. I was
Speaker:impressed. Sorry I cut you off. We're
Speaker:saying we're seeing some of the same things
Speaker:coming out of some of the big finance companies
Speaker:as well. I think they're the first that we're actually seeing some results with
Speaker:Agentic, actually like real
Speaker:return of investment result. Right. And this actually
Speaker:goes to a really important point. I want to sidetrack because it's related.
Speaker:There was a report recently by MIT that
Speaker:people have been misquoting and just the most epic way.
Speaker:Oh, the 95% failure. Yes, I was going to talk about that because
Speaker:like, I can't be. Look, I understand how hype weights work, but it can't be
Speaker:that bad as you start peeling back the paper. Like
Speaker:there's a lot of caveats there. Yeah.
Speaker:Has to do with the type of R and
Speaker:D projects that they were talking about.
Speaker:If you actually read the paper, it was more like 40,
Speaker:45% success rate. The
Speaker:95% had to do with like a specific category of,
Speaker:of project. So I need to, I actually need to. I keep telling myself I
Speaker:need to dig into it a little bit more, but when I did initially
Speaker:go through it and read some summaries on it, it
Speaker:was that it's just been misrepresented completely. And
Speaker:the, the data set that they were using was a little skeptical as well. Just
Speaker:a little odd. I think it's a lot better than
Speaker:that. And then I think those 40% that are
Speaker:seeing ROI are actually seeing really significant ROI.
Speaker:And I don't think that's going to change, I think.
Speaker:So if you're deciding where
Speaker:you want to invest your nest egg, I
Speaker:would not be too concerned about
Speaker:AI. Now, again, I'm not your financial advisor. I gotta put a little thing down
Speaker:there. Do talk to your financial advisor.
Speaker:But ultimately, no, I do think the data is actually
Speaker:showing some really great results. Obviously there's going
Speaker:to be hiccups in these types of POCs. There's a
Speaker:lot of people who are just throwing
Speaker:projects out there to see what sticks, but the actual
Speaker:projects that are meaningful proof
Speaker:of concepts. So not just, you know, I bought,
Speaker:I bought this AI technology and it's sitting on my shelf, but I
Speaker:actually got a team together performing this. We're doing
Speaker:agentic. We're trying to solve this
Speaker:actual problem statement. We have a problem statement.
Speaker:Those are the ones that we're actually seeing meaningful results in the industry, especially
Speaker:some key, key industries like finance and telco, which
Speaker:we typically see kind of lead the way in some of these areas too. But
Speaker:it was a really interesting report because it's added a lot of
Speaker:doom and gloom on the Internet. And I see a lot
Speaker:of the naysayers about AI just be like 95% of. It's
Speaker:not even, you know, succeeding. It's terrible.
Speaker:And I just have to sit there and shake my head and be like, no,
Speaker:not what the report said. But I think it's just clickbaity, right? Like it's
Speaker:clickbaity. It's total. That's kind of what, you know, I
Speaker:didn't go deep into it, but when I started peeling back the layers and reading
Speaker:other people's analysis of it, I'm like, that's clickbait.
Speaker:And it gets back into this. Is this an AI bubble?
Speaker:And yeah, maybe it is. But if people don't
Speaker:remember, I'm old enough to remember. I have enough gray hair to remember what the
Speaker:original dot com boom was like. And there were a lot of
Speaker:people predicting the end of the dot com rise as early as
Speaker:1996. Right. And people,
Speaker:the dot com bust wasn't just a one and done type of event.
Speaker:It unfolded under a couple of stages. Right. As, as
Speaker:one of the books, I think of the name, I think it's called the Everything
Speaker:Store. It's an analysis of how Amazon started
Speaker:from Jeff Bezos having an idea while he was working, I think at a hedge
Speaker:fund. I think it was so early, it wasn't a hedge, called a hedge fund
Speaker:yet. And all the way through
Speaker:to, you know, basically:Speaker:and you know, as late as
Speaker:,:Speaker:analysts were convincing, you know, Jeff Bezos that
Speaker:he should sell them to. Should sell him as his company to Barnes and
Speaker:Noble. Yep. Right. Which is kind of funny to say that,
Speaker:you know now, but, you know, the dot
Speaker:com bust as it happened, you know, for me
Speaker:it was. I Remember hearing in:Speaker:an end. Another year later it was overhyped. And then
Speaker:1998, people were saying, oh, this is over. Right. When
Speaker:the real bust happened in:Speaker:But maybe the AI boom
Speaker:is going to see that too. Right. Or is it going to be more like
Speaker:the crypto kind of craze where it kind of crashed but
Speaker:it kind of went up? It kind of went up and then it kind of
Speaker:fell back and it kind of went up again. It was more of a. I
Speaker:wouldn't call that a soft landing, but it was definitely like a. Yes. It
Speaker:wasn't an explosion quite like the dot com bust, but it wasn't quite
Speaker:like. It was more like a bumpy like, crash into like
Speaker:an empty field where it kind of like hit up. And I don't remember, it
Speaker:was one of the Star Trek movies where like the Enterprise like crashed on
Speaker:the planet and like kind of skid along for a couple miles, bouncing up and
Speaker:down. That's kind of the, the crypto crash. But
Speaker:I don't want crypto bros hating on me. I, I like crypto. I just
Speaker:don't understand a lot, a lot of questions I don't understand
Speaker:about it. Right. Like, I understand Attack, but I don't understand how we're going to
Speaker:get from the tech to this utopia that we're promised.
Speaker:There's a lot of, a lot of steps in between I don't quite get. But
Speaker:I don't know what, you know, A.I. i think, I think if it is a
Speaker:bubble, I still think there's still some, some room, Runway left for it
Speaker:to happen. Right. Because you are going to see. Yes, there are real
Speaker:risks of, of having these experimental projects. Right. If you have 100
Speaker:success rate in your experimental products, projects, you're not taking
Speaker:enough risks. Yep. Right. If you. And you said
Speaker:was 45. Yeah. It's closer to like 40, 45,
Speaker:which I would. If you're really. 50% would be the
Speaker:benchmark there in my mind. Right, right. Like in terms of half of them fail,
Speaker:half of them succeed. Right. 45 isn't that far off
Speaker:from that. Right.
Speaker:I would say. And, and there's also been a
Speaker:lot of these, you know, all the, you know, X number of percentage of AI
Speaker:product or data science projects fail. Well,
Speaker:you know, a certain amount of science has to fail. Right. Yeah. In order for
Speaker:you to really be advancing the thing. Like, you know, and I think pharmaceutical companies
Speaker:are a good example of that. You know, you, you only
Speaker:hear about the drugs that worked. Right.
Speaker:Get approved on you. Then you hear when they fail after.
Speaker:But I mean, like, but you don't know, like day to day. Like, how many
Speaker:chemical compounds did they try that didn't work out? Right. Maybe it was a hundred.
Speaker:Right. But that one, if you look at pharmaceutical. It's an
Speaker:astronomical percentage. It's actually. Right.
Speaker:Truly insane. Like such a low percentage of what actually makes it
Speaker:to. There was an interesting analysis. There was some podcast somewhere. But
Speaker:basically how venture capital works. Right. Like they give money to like
Speaker:100 companies. Right. 80 of them are going to fail big.
Speaker:Right. 10 to be, you know, they'll break even.
Speaker:But like one or two of the remaining 10% knock it
Speaker:out of the park, Right? Yep. And that's kind of how
Speaker:mathematically they function. I thought that was an interesting.
Speaker:Maybe these AI projects or whatever
Speaker:will follow the same trajectory. I don't know. But I feel better
Speaker:at 45% success rate than 15 or
Speaker:5. Yeah. Yeah. Absolutely.
Speaker:Cool. Always good having you on the show. I
Speaker:know we both have hard stops. Yes. Unfortunately.
Speaker:No, it's cool. Gotta have you on more often, man. Especially now that you're not
Speaker:like spending a month out in, you know,
Speaker:Australia and Asia. Yeah,
Speaker:yeah. So let us know in the comments below what you want to see us
Speaker:to cover and maybe it'll be tomorrow.
Speaker:I got this here the other day. This is a flexible
Speaker:solar panel thing. Oh, cool. So it's cool. Supposedly it's 100
Speaker:watts and you can actually pack it in your
Speaker:backpack. That's the video. And I was like, oh, I need that because. Because I'm
Speaker:a big, I'm a big fan of like, you know, having power on the go
Speaker:and stuff like that. So. So I'll,
Speaker:I'll unbox that tomorrow. Any parting thoughts?
Speaker:Just keep an open mind about AI and
Speaker:I, I still think the, the biggest conversations are still about
Speaker:the governance of AI. Absolutely. Yeah. Just know that
Speaker:AI is a multi layered problem, not just a single layered
Speaker:problem. And for us to get this right, we have to look
Speaker:at all the different layers. Absolutely. That's
Speaker:how we're going to be able to do it correctly. And I will tell you,
Speaker:I was listening to a podcast, I'll leave you on this note. And there was
Speaker:one expert that was talking about
Speaker:basically, are we, are we creating the
Speaker:terminator out of all this? And he, he said, I
Speaker:I'm actually more worried that we're creating Wall E out of all
Speaker:this. Interesting.
Speaker:And I would encourage everyone who hasn't seen Wall E go check it out.
Speaker:And keep that in the back of your mind too, that there
Speaker:could be such a happy path with AI that
Speaker:also has its own long term negative effects for
Speaker:society. But. But yeah, that's a topic that you.
Speaker:And I can talk about on our next stream. That's it?
Speaker:You want to leave on a cliffhanger, so to speak? Yes. And that wraps
Speaker:our deep dive with Christopher Newland proving once again that AI
Speaker:isn't just about large language models spitting out cat facts, but
Speaker:about simulating reality, bending time at devcon and
Speaker:maybe, just maybe, preventing the rise of our robot overlords.
Speaker:From protein folding to Grand Theft Auto fueled AI breakthroughs.
Speaker:Christopher reminded us that the next big leap might not be in scale, but
Speaker:in simulation. So thanks to Christopher for navigating the
Speaker:uncanny valley with us. No jet lag, just pure insight.
Speaker:Until next time, stay data driven. And remember, if
Speaker:reality starts glitching, blame the simulator, not the
Speaker:Internet.