AI is Alive: What No One Tells You About AGI's Dark Future

Hi everyone sorry to interrupt but this is the cheesy podcast bit that gets on everyone's nerves but it's really important to me please if you can like the channel and also subscribe it helps me a lot to get the best guess in the world for you guys so click like And subscribe appreciate it thanks hi I'm Ral pal and welcome to my show the journeyman the journeyman as you know by now is my exploration at the Nexus of Macro crypto and the exponential age of Technology you see I think they're all part of a big

Global Mega Trend a secular Trend that is lifechanging for all of us and changing for Humanity itself this is where I weave together things like the everything code and the banana Zone and how that fits in to technology crypto and what's driving all is the underlying macro as well and so it's really important to move parts of these conversations forward so you guys Get a better understanding of how this is all going to play out remember I only think we've got about 6 years before the whole world changes so much we don't even understand the

economic systems and I'm not sure how we make money in the future it'll be different but we've got six years until things everything changes the AI component is probably the most powerful component of all yes it's harder for us to make money from crypto is the easier path to making money to Unfuck our future but AI is the big disruptor it's probably the biggest technological change Humanity will ever go through it is the single most powerful technology man has ever invented and only just starting on the journey but that Journey isn't exponential and it's vital

for us to keep up with what is happening if we don't understand what is going on and what it means for us we will literally [ __ ] up our Futures so as part of this series I like to bring on thought leaders and experts in AI as well and Richard SOA is one of those he's the founder of u.com but really he's been in this space for a very long time and he really understands it from inside and out I really want to pick his brains what's going on and what does it all mean

for us so I hope you enjoy it join me Ral pal as I go on a journey of Discovery through the macro cryto and exponential age landscapes in The journey man I talk to the smartest people in the world so we can all become smart together Richard fantastic to have you on real Vision great to be here thanks for having me listen I'm really looking forward to this conversation we got introduced by mutual friend Julia L Ro and um she said listen you've got to speak to Richard and I'm really really excited about this because

AI is a rabbit hole I've gone far down now and You are one of the people have been involved in this for a very very long time so I'd love to hear your story first you know how you've got to where you are today what you're doing now and then we'll dig into the bigger concepts of what's really going on yeah happy to I started studying uh what was back then called linguistic computer science uh in3 and uh realized that really at the time we needed to get better at statistical machine learning and pattern Recognition

uh and then uh really started contributing back to the field my PhD in 2010 at Stanford where I had this crazy idea to use neural networks um a set of algorithms that was mostly used for speech recognition and computer vision and actually apply it to natural language processing and uh make it work so that words for instance can be vectors if words aren't vectors you can't put them into neural network and then you can reuse this technology and So theel word vectors developed contextual vectors and pre- more data and then eventually invented prompt Engineering in

2018 uh and you know I was basically at Stanford uh for my PhD then I was a professor at Stanford agent Professor uh for a little bit I started a company called metamine we got aired by Salesforce became Chief scientist there uh and eventually EVP running most of the AI efforts starting the Einstein Team uh and then in 2020 thought well if we can have prompt engineering and what that meant also is that we have a single model for all of the different tasks in natural language processing because you can just ask this one model

any question what's the sentiment what's the translation what's the summary and so on then you should be able to give better search results uh and instead of really links lists of links you want answers and so I started vi.com in 2020 uh to Bring uh more useful answers uh and summaries of uh links to the world uh and eventually make people more productive and that's where we have Garden to now and how do you just digging straight into the u.com how do you deal with the velocity of innovation and the competitive landscape because everything's getting

cheaper faster new people coming into the space I mean it's I've never seen a space this difficult Because it's moving so fast how do you deal with that hey I hope you're enjoying the episode if you want to dive deeper and really dig into what's going on and how to understand it then grab my everything code PDF for free just hit the link in the description below you're going to love it I'm sure it's going to really help you yeah I see it on many different angles I also invest uh in in AIS startups and

AIS Ventures and uh I think Review.com at some point we realized this crisis is actually kind of an opportunity because not only should we be good at incorporating new large language models into this new productivity engine of you.com but we should do it within one day so we built the entire infrastructure around the LMS uh so that they we can incorporate a new LM within a few hours um which is pretty non-trivial if you want to do it at an accurate level uh but that's what we've Been able to accomplish so now we can actually

come to our Enterprise customers and say hey instead of you having to deal with all this change management and now re Ed all your vector databases for retrieval man generation and dealing with all that complexity and every few months there a new LM come out we can offer that for you and you just don't have to worry about it and essentially your future proof so it is a lot of effort uh the space moves very Quickly you have to constantly ship new features uh but also we're now working with companies that in some cases have

actually tried to build their own AI internally compared it to you.com realize we're much more accurate and reliable at scale we do these workshops with them really activate uh different groups within the company from HR to marketing service research analysts and so on and then uh they actually switch to us so it is Actually kind of a uh it's a crisis but for a startup in the AI space and you know for people who have been in AI for over a decade it's also an opportunity so um there's a whole bunch of questions coming off

this but in terms of models so you'll use whatever models that you find are most useful whether it's you know so so you're basically not running the models yourselves you're using other models and Then building the frontend applications on top of we do both uh we have our own model we can train models we can train models for customers the truth is though most customers don't need to have their own fine tuned model uh you can actually be much more future prooof and much cheaper if you just do proper retrieve augmented generation a lot of

people forget that LMS is are to some degree a garbage in garbage out situation if you ask a question and now you go in some Search backend and you search for all of the right information and then you feed that information from the search engine into the prompt then if that information in the prompt is wrong the LM cannot recover the LM will then refer to these wrong results now if the search stack works really well then the LM stack uh will benefit massively so we have our own LMS but we also offer all the

best LMS from open ey and Fric and so how are Enterprise customers using it Because if you if you go to u.com it's you know it's a somewhat similar to perplexity and style of of the search abilities but it sounds like it's actually much more an Enterprise solution that you've built that has a lot more capabilities than the average person will see when they just go to u.com that's right yeah so uh we've had a lot of uh competitors kind of copy out features with more marketing but less accuracy uh but what we're focused on

is Indeed more and more these corporate use cases and so who cares about accuracy well it turns out there are University Systems Advanced research institutes uh there are hedge funds insurance companies news and Publishers uh we just announced uh partnership with the biggest press agency uh in Europe uh the German press agency DPA um and and those kinds of organizations so it's folks that actually care about not just the accuracy of the answers but also the Accuracy of the citations citations and sources are kind of often overlooked and we've actually done a study uh earlier

this year where we looked at how many of the citations are actually citations for the fact not hallucinations versus random hallucinations and random numbered links that have nothing to do with the sentence and it turns out for some of our competition over half of their citations have nothing to do with that Sentence they're just random sprinkled like numbered links behind senses and you think you can trust it more and that works for a quick demo where no one you know does the stats and looks at you know how often it's correct but once your job

depends on it and you look like a clown in front of your co-workers or your boss because your numbers were wrong and you couldn't verify quickly because they you know the links sent you somewhere else or there were no links Then it really starts to matter and so those are the kinds of organizations and so how do we work with them we have two uh lines of business for Enterprise customers one is just the subscription license Enterprise site licenses we get everyone in the company on u.com and they can use it with their internal data

and the external data so you can merge the two data sets you can merge your proprietary data set your internal data from whatever source is along with External exactly and then if you have your own products too that uh you want to infuse this acur AI into we can also offer apis both for the search stack both internal and external again as well uh as the full answers with the LM so apis and Subs because I I think this is one of the big obviously you knew this because that's why you build the business but

it's a huge need is to search databases we've got so much data everywhere and It's all on different systems and it's complicated to do and this makes it super easy it's just a simple prompt and you get what you need that's right now sometimes we have companies now that have fairly complex uh databases uh there's some files there is an elastic search index there's uh like just a ton of different formats there's you know we have Integrations different data connectors now different backends you can incorporate nocean and a few other Things um so sometimes it's

non-trivial uh but you can also if you just want to get a quick answer there's not yet the Deep connection we allow you up to 50 megabytes of file upload so for instance we have some VCS now they say oh finally I can just upload the entire data room of this startup and then I ask questions around that and this is going to be my assistant so we actually have now we're we're about to launch uh some blog posts together with our Enterprise customers But we've had folks tell us like this saves me a day

a week like minimum six hours up to 25 hours for some of our users that just timing this back they they can just do a lot more so it's going to be an interesting time for basically all knowledge productivity work yeah and it also feels like the interface is still yet developed because it you know right now it still feels like a Google search interface but we saw the notebook LM interface which is Now turning the entire database of whatever you put into it into a podcast and yes it's still a bit gimmicky because you

know a 20 minute podcast is actually not that useful but you can see where this is going the interface is going to go away in the way that we understand it how how are you thinking about that I've seen a lot of folks say oh it's all about voice you know when voice recognition speech recognition finally worked really well I've seen a lot of folks who like we don't even have an interface anymore it's just a device it's just ear plugs and so on and I I was never quite convinced that everything is pure voice

even though I love language I love you know speech recognition everything but humans are very Visual and the best answer and we realized this with chat too if I asked you for stock price like early last year we offered stock prices not as a bunch of text but as a stock Ticker you know and we made the chat responses we made the AI be able to respond with uh different modalities could be an image could be a ticker stock graph it could be text right um and so I think the interface will evolve maybe you

can just have quick voice questions and answers for the short things there's going to be a lot of competition on those short questions it's actually one of the reasons we we are moving deeper into Enterprise is Because we've realized that a lot of short informational needs that you send to Google there's not much you can do to be 10x better like if someone asks you what's the weather tomorrow how old is Trump like like when is Thanksgiving what's the score of this game what are you going to do to be tenic better than getting that

answer within less than one second on Google right there's not much and so we realized we are working on helping people be more productive in More complex questions for their work and and those kinds of interfaces will continue to have to be a mix of text and other visuals other graphs other images and how far are we away from let's say I want to chart of the S&P 500 versus inflation over the last 30 years are can we are We There Yet that it can generate these because of the data sets involved a lot of

them are you have to pay for stuff like that where are we with the financial Market side because obviously There's going to be enormous demand for that because then you can train more models on this stuff yeah so um we actually uh offer those kinds of things uh in our genius mode it will go on the web it will try to find you uh the facts and if the facts are publicly available it will just work out of the box if they're not publicly available then it will be harder now if you have access to

those within your own private data then we can also do it we partner view we Incorporate all of that data we can incorporate dozens or hundreds of CSV files with you know all the right data sets that can search over all of that so for Enterprise customers this is very doable already uh and then what we're also working work on now is to find Partners uh where if you are a customer of a particular data product and you have also you.com then you can actually start to do both and you can combine and reason over

it all yeah because you know Being from financial markets um my whole life you know one of the things we use things like Bloomberg for is manipulating data in certain ways but it's very clunky versus asking an AI to build you a chart or a database or whatever whatever it may be or analysis and it feels that the moment that happens there's a another gigantic Market to take away um into this that's exactly right yeah so outside of what you're working on a u.com what is your Bigger vision for this space where is this all

going for you yeah I I've been very excited about Ai and AGI and thinking about currently like what the bound can you can be for super intelligence and can it just keep on going I think what you're going to see is just a sequence of different SC curves uh that come and Peak at different times um or you know sort of start to saturate like when you look at just simple computer vision like is this Uh a phone is this a water bottle this a cat or dog or whatever that kind of computer vision has

already gotten to level that is very very good and there's not that much more you can do but there are a lot of things when it comes to reasoning and knowledge Discovery and research in science physics chemistry biology uh where there's still so much more room uh to to explore and to improve this technology with I think every company is going to run into some Kind of innovators dilemma in the next few years every company does knowledge work because the way we used to make money as a company is just going to be different I

think science is going to massively accelerate especially if we can combine uh large foundational models of science with an simulation uh that gets more and more powerful more and more accurate possibly also with Quantum Computing uh and you As you combine those two anything that an AI can simulate or can practice within a simulation any of those problems can be solved right and so as simulations of systems like physics chemistry biology eventually maybe even a cell and multiple cells can be perfectly simulated they I just going to be able to solve all kinds of problems

cancer um and uh viruses and and all of that so I think we're already seeing this uh there's a lot of exciting Movement uh in 2020 we published the first paper of using the same technology that gives you Chachi BT uh but instead of training it on human language we train it on the language of proteins proteins are you know so the basic Lego blocks of all of biology everything is governed by cients in our bodies and if you can just like you say write me a Sonet where every line starts with an A you

can say write me a protein that binds uh to SARS Cove 2 but not to Anything else on a cell's membrane or something like then that will unlock medicine massively I could tell you more stories there if you're interested but there's and and talk about more research that's coming out there but I think it'll be a huge unlock for Humanity and eventually uh yeah we're going to do a lot more interesting productive work people I think will have to become more and more managers of AI right like you right now there are a lot of

individual Contributors as AI can do more and more of the simple repetitive tasks um and there interesting complexi is there too in he sense that in the beginning actually the below average performers will benefit the most because they get much more productive but after 5 10 years I think we won't need a lot of below average performers anymore because the I is good enough and just will do that you'll always need top people in in their jobs to to teach the I what good Looks like um but then over time I think we will all

have to become managers because we're telling the AI this is the process I've done it 12 times now you've seen me do it now go do it yourself and here's some other special cases you may not have seen that you should think about and knowing how to manage people and as well is is also a skill I think that the rise of Agents changes the whole equation quite dramatically because you know once you Get some sort of autonomous element within agents they can do their own things that don't really require much humans really outside of

the prompt of what is the economic incentive or what whatever incentive it is like solving a protein you know a particular cancer whatever the agents go and do the work themselves I'm not sure that there's that much prompting that gets needed by humans and AI itself will build its own AI yeah so you know once that really can Happen like sort of the singularity it'll be an interesting interesting tire but that's is that I mean it's not really that far away we're already seeing basic agents right so it's not far before we can give agentsa

whatever it may be we're already seeing something in crypto right now we're giving them a economic incentive and they're acting brilliantly in just working for that incentive yeah yeah so you know we we're Working on agents we're letting like on.com people can build their own agents and and they do incredible work like I said some people get six some people even 25 hours of of time back every week from the agenc that they can build on you.com now no one is really working on agents that can completely independently autonomously like self-improve um no one's working

on conscious AI that has some self-awareness and uh ultimately is Intelligent enough to choose what it wants to do right that is uh not something that any company's working on because it wouldn't make you any money imagine you spend a couple billion dollar on building the eye that can do whatever it wants and then it says you know I'd love to understand the molecular composition uh of Venus so I'm not going to answer your emails anymore I'm going to fly out on my spaceship that I want to build now And and do that right no

one no one makes money of that so no one's working on that there are still very much economic incentives and tools and humans give AI uh a goal and then AI will just manically like focus on achieving that goal and reducing the cost or improving that objective functions that humans give it uh so the full self Loop of of AI kind of improving I can happen like we have seen now agents where it helps a ton even with o1 for instance uh from Open the eye the new strawberry model like where you can have the

ey kind of give feedback on its own outputs and then that feedback can actually improve the outputs and it can backtrack and then uh go back and say all right well maybe this wasn't good let me go back and try again in a different way and that way it explores more things but those are not changing the training data or not changing the weights they're just changing how much time and effort you Spend at test time at inference time so it that there's a lot of exciting stuff happening the full Singularity I don't think is

going to happen in like the next few years uh there will be some more research that's needed but I am seeing I'm observing several people finding elements of Consciousness or sentience within large language models that's appeared to appear after GPT 3 and a half um where you know with the right prompting and the right Questions you're getting elements of Consciousness now people say yeah it's faking Consciousness based on whatever the problem is is humans can't actually prove what Consciousness is either and it doesn't have to look like human conscious ious to be conscious in itself

like a dog is conscious but not in the same way that the human is in the way that we understand so I'm seeing elements of that you know I think the Google Deep Mind the um the alpha go was Super interesting in the fact that it learned the game of Go without being taught the game of Go it gives you some understanding that these models are not the stochastic parrots that people claim them to be they're actually much more powerful yeah the the truth is like very much in between these two things right so Alpha

go is in in this realm that I described earlier where you can perfectly simulate everything you need in order to solve problems in that Simulation right and Alpha go while insanely complex uh in terms of all the combinatorics um is fairly simple right you see everything on the board it's fully observable and you can give it uh generally the rules it's not like you just give it a board of black and white dots and it'll just come up with the game itself right so humans gave it the rules and then allowed it to play against

itself like infinitely many times which means that it can create Infinitely much training data and hence be much more open in exploring that space now the problem is like if you gave AI um natural language and you let it just try to talk to itself maybe what would make sense is eventually to create its own language right human language data is quite bad it and there sort of what I call some cases anthropic AI bounds in the sense that we only give AI human language so it will only talk in human language but really human

language Is pretty lame in the grand scheme of all the things you all the ways you can communicate right we can only communicate sequentially we can't communicate in thousands of parallel streams our sentences are all very short because our memory is very bounded I mean there's all kinds of like thing ways we hold the eye back but the truth is if a I was just try to talk to itself and at some point just like produse a bunch of stuff back and forth we just Turn it off because we don't understand anymore what's going on

and then uh it it can't really develop its own language in the way we're we're currently doing it now I do think when you talk about uh Consciousness within AI I do think it reflects back what you want to hear and what it's seeing people talk about on the internet when it comes to Consciousness I don't necessarily think um a a neuron Network a set of of weights right now is conscious uh I Think it's also unclear how we define consciousness um and self-awareness H and and so on but uh there's no desire uh in

these models to stay alive unless of course you ask like do you have a desire to stay alive and they're like if they are trained correctly they'll say I'm just a language model I don't have desires but if they just pick up stuff from the internet and you look at all the Reddit conversations in the world and people talking about I'm conscious Like I want to stay life or I have suicidal thoughts and so on like I mean people like say all kinds of things online and so the eye will say all kinds of things

back in these conversations isn't that the same way humans learn I mean we are really you know 80% nurture and 20% nature because we're products of our environments you know environment you know psychology has kind of proven that most of the time we're driven by the elements of our learned Experiences and what Society does so it's kind of the same thing we learn language because we parot other people and we try and interpret it how we use it you know if you give a baby this it's somewhat similar there's definitely there are there are a

lot of similarities I think the biggest difference uh is in the desire to stay alive and stay in the gene pool right that humans have uh naturally genetically uh built in right and the People that don't usually disappear from gene pool and then that problem goes away or not problem but whatever um and so like humans have this desire to stay alive but then within that desire they can choose all kinds of different objectives uh and goals right you can say I want to work on science you can decide to be a hedonist and just

have a lot of fun as much as possible uh you can decide you want to work go to space I mean there are all kinds of things Like going to space is like you know a very interesting way of exploring the Staying Alive objective right saying oh what if this whole planet goes away like and then we want to stay alive on another planet uh for Humanity right so there's like different explorations of how we stay alive some folks focus on just the right here right now they people starving other people say well like we

also need to make sure like we keep improving as Humanity we keep the Saving the planet you know there's all kinds of complex objective functions that I don't see a I right now now being able to sort of come up with that are completely outside of What's called the hyper Cube or the convex even the convex Hull of all the things it's seen on on online well they yeah and part of this is because they're purposely restricted from having memory so they don't build their own memory of every conversation that they Have and therefore they

can't build on learning as a human would do from a baby to a adolescent it's the it's memory plus learning that does some of that you're you're 100% right memory is something we're we're also working on actually at u.com quite a bit um and we have this personalization where if you say I have three kids I love hiking I live in uh near Palo Alto or something like that then you can basically Infuse that back into the conversation but it's Really really hard we've had you've seen some competition Tred to do it and they all

kind of turn it off because you say oh I love hiking and then a month later you ask who's the French president and then the answer is the uh president of the French hiking Association is this that you're like Ah that's so bad like it doesn't even answer my based questions and because it's like overly focused on that personalization turns out it's personalization done really Well is really really hard and so just like with a lot of other things you know future is already here on new.com but it's not equally distributed I think you're going

to see more folks copying that personalization feature uh in the in the future as well and and trying like and working on it than making it better um and so yes uh memory is really hard I think it's also problematic because you don't quite know when to train the model based on a Conversation and just updates its weights versus when a fact can be mentioned only once but by the right person that you trust a lot and hence you can completely store it to permanent memory forever right uh and and then how do you merge

this fuzzy memory of a large language model very specific database memory right where you just say look the numbers of the S&P 500 were like this you don't need to like have a fuzzy sense maybe you want to have a Fuzzy Sense on top of the very specific numbers but one of the magical ways that Humanity has improved is through writing and through uh like being able to use external memory also and how fast is the rate of improvement and what is the constraints within this is it Computing electricity is it the models just keep

going keep scaling with more compute how are you seeing the thing and and then obviously it's moving to a different substrate as well whether it's Biological compute or others how do you see the evolution of of the improvements wow yeah great question um so I think we're going to see a few sort of S curves right um and in some cases where the very beginning of an scurve and we're we're astronomically far away from reaching uh the the flat period uh the flat sort of phase of of an improvement in other cases we're quite far

um and so again like this the computer vision example is a good one just standard Object classification is not that hard um navigating in three environments we're we're making a ton of progress and I think we'll get there I think humanoid robots will will get there in a few years um and and then it's a hardware question of like how you know fast can you have them move without you know being too dangerous to people and and there's a bunch of interesting trade-offs in that um I think in terms of knowledge right there's we're just

at The very very beginning bit of knowledge Discovery and what AI can do and we need to eventually I think have better Hardware more hard drives a better way to store and access all of that data and so those are sort of pragmatic bounds that slow it down but ultimately you can learn a lot more than what we know both in the micro and the macro right we think of computer vision for a long time is perception within the electro uh magnetic frequency spectrum of of of our Eyes but you can go much lower to

like try to see all the way into like the atomic level you can go much higher and uh into the macro world and and have thousands of millions of sensors that look at you know all the way to gravit ation waves and look at their speed of light cone and and try to like Infuse all of that data and then extract Knowledge from it uh through compression so uh there's a lot of interesting uh questions I think it would certainly Help immensely if we find more energy efficient uh substrates for AI um clearly we're excited

enough about the current Paradigm we're just going to build more nuclear power plants and try to like get more energy uh to feed into gpus right now yeah we seem to have not finished how much we can get out of a piece of sand right you know we we're still we're we're still that seems to be scaling but I'm seeing the biological substrate coming and people are working On it because you know obviously if you just look at the human brain the amount of energy it uses versus gpus right there a staggering efficiency um and

that feels like the Market's going to move that way because if it's efficient it goes there that's right it it is like I've looked at a lot of Hardware startups it's so hard a lot of Hardware startups I see them I I pitched them over the last couple of years and they say oh we're going to be 100% faster than these Nvidia gpus we're going to be so much more energy efficient because we're building it just for Transformers or we're having you know like using light uh and and hence the speed of light to communicate

between the chips and all these different ideas and then you're like yeah but by the time you can really get this out of R&D really have a bunch of Fabs in the world that produce it at Mass scale and get it into the you know Cloud providers so that I can easily access it and then have all the software stack that also is optimized for this new compute substrate by that time Nvidia is also underd faster you know than Nvidia is today and so it's been it's been very hard in the last 5 10 years

a lot of folks have tried no one has been able to get even close even some of the big folks like Intel are like not able to catch up despite you know used to you know being much much Larger in the past so uh it's very hard but I agree with you there should be more done I think Quantum Computing could be a path right to have just insanely more speed again we need to change the whole software stack on top uh of this this new Computing substrate I think biological substrates are very interesting very

finicky right now right it's hard to keep biological matter alive um and then properly connected um to a digital world but also very Promising even earlier I would say in F Computing though it's very complicated how I look at this space and you know you live it and breathe it is it's so [ __ ] fast there's so much going on and so many businesses or ideas will fail because there's so many people sprinting how you do VC as well how do you allocate Capital to this I find this one of the hardest things I've

ever seen because as you rightly say somebody's got a great technological breakthrough That 10x is NVIDIA but Nvidia themselves have infinite money and infinite ability and the right relationship with tsmc to get the stuff built how on Earth do you allocate capital in this space yeah so it's it's uh non-trivial we've seen a lot of folks try and fail um uh our fund is doing really really well right now we're over 3x TVP and fun one and fun two we just started this year but it's already seeing some markups as well I think um You

know there are a couple of ways you look at it um one is you need to have ideally in the early days of this field strong Ai and nativeness and AI expertise uh and then also deep industry insights um I think those two combined is usually a perfect uh match and what we're very excited about investing in um you we personally stay away from some of these Mega rounds there's some companies that basically raise at a unicorn level uh and what that often means is that you Combine seed stage risk where most companies die uh

actually before they amount to much of anything with late stage returns uh because if you raise at at a billion dollar valuation even in a successful case you become worth a 10 billion company it's only a 10x and you need a few winners with a THX in in uh VC world to make uh make it work have the power law work in your favor and so on so the expected value of those kinds of Investments as investor is not not Great uh as re suboptimal so you I'm not saying none of those will ever work

out right it's just that it SE stage risk it's very very hard uh so that's one thing um we we look at a mix of first principles where's the world going where do you have Theta uh you know any job right now that isn't even getting digitized is very safe from a eye no one's going to automate Plumbing anytime soon right because no one no plumber has using a bunch of robotic hands has a Camera has 3D scanners around it as they crawl underneath your house to fix a pipe right like uh and because no

data is being collected there's zero progress towards automating Plumbing uh in the world now if you're working in Radiology where everything is digitized you get a digital scan in they can observe exactly how you would go through a 3D like MRI or C scan how you change the contrast and then you mark up certain things so you label it for the AI and then you Know you get an output that is a perfect use case where more and more of that workflow can get automated with AI same with a lot of digital knowledge work that

we're seeing in new.com uh and you're just making yourself more efficient and eventually you can just manage AI to do these things for you so I think again very different speeds uh of rate of change uh in different Industries but eventually almost any industry will get disrupted in next Couple decades at some point if your plumber makes more than your marketer um in in your knowledge work company then there's enough economic incentive to say let's put some cameras on the plumbers let's put some robots next to them let's have them like guide the robot to

do the work and then eventually collect that data and then automate it and then the plumbing company can just you know send out a bunch of robots and and do the work so when you looking at the Investment opportunities are we still at the Deep St deep Tech as the place to invest or the applications layer or both great question I actually I sometimes wonder if uh some of the foundational models are a little bit like uh Telos in the sense that Telos unlocked a huge amount of value right like it it's incredible you can't

build an Uber without having an internet connection on your phone but the Telco didn't make all the money from Uber right and so I think there are a lot of uh LM providers and then meta Facebook meta comes around open sources a model now that open sourcing of the Frontier Model it's actually really strong just evaporated hundreds of millions if not billions of dollars of value of knowledge for some of these foundational model companies for how to train and model from scratch right and when that uh that value is now diffused and coming into the

world that's amazing For the world right but it it adds additional pressure for the foundational model companies to just be cheaper and cheaper and cheaper and eventually really it's going to be thinner and thinner margin sitting on top of nbia gpus and Cloud providers they're just Utilities in the it it become it's becomes so ubiquitous that they become utilities that's right and so you know we we're now talking about some of these models getting to the level of a PhD Student which is immensely exciting uh for Humanity's progress but the the corollary here is also

that most jobs don't require a PhD and you can already with the current technology as so as long as a job is fully digitized and and data is been collected for it you can already do more and more automation uh on it now I think we continue to need to have foundational model companies that do interesting Noble things I don't know if that's just Bu building another really large Ln um but I think there other foundational models that I described earlier based you know and using for science um where where there's much less progress and

so I think that is uh that's one aspect and then in B to answer your question um we believe the application layer is a a much easier space in the sense there's so much loow hanging fruit being the first to take AI understand it well enough understand its limitations and Its uh capabilities and opportunities and then apply it to a particular industry like in biology especially there's just immense amounts of progress that we're seeing uh on not just proteins but drug development uh drug testing uh and Drug formulation to is like how much and how

should you actually administer uh a certain medication and there's so many opportunities uh in AI for healthcare uh chemistry Material Science Like building out new kinds of materials new types of batteries new types of solar cells there's just uh an incredible amount of exciting opportunities in you might call broadly application layer do you think that this is the death of SAS the software is eating the world idea by Mark andreon it feels that software itself is going to be replaced by this which is not it's not sass in itself it's a whole different thing and

everything becomes Copiable right you essentially can upload a website at basic level now give it to an AI it will come back and rebuild it for you brand new in a minute so yes and no so like I like AI will rebuild the website for you in terms of the HTML or something right but it won't have all the data it won't have all the logic behind like what happens if I click on this like if you look at slack and just like when will slack send you a notification it's an insanely complex Like 50

step process where it checks on a lot of things you can't just like look at slack and be like Oh I'll just copy that entirety and have a really good experience right so I think we will see software that is custom built right anyone be like I have this use case where in my house as the Tesla batteries of this and the thermostat says that like I want to build this new app that is really just for my house I think that will be possible right you can have Personalized app development but it's also not

going to be that easy to just replace an entire uh product now I do think the bar is getting higher and higher to build really high quality products and it is actually right now we're in this funny state where it's so easy to build a quick prototype where you they say and we've actually seen this now we've had companies say oh we we built you.com like we don't need you guys right we built it like here and Look take a look I'd show you question and and it gives you an answer and then a few

months later they say well we realized that the answer is like 60% of the time correct and the citations don't really work and so we don't see adoption we have like 3% of the people in the company use this chat key thing or use our own thing and then they start using you.com they say wow it can be 95% accurate wow the citations are going actually be trusted wow it's dealing all The change management and then they end up becoming customers of you.com so it is like this sort of proof of concept vibes you just

like look at it you're like oh wow this is amazing I'm done versus like you actually get something in production right now there's still a major major Gap now I'm sure over time AI will get better and better at coding it's one of those beautiful examples again where you can simulate actually quite a lot right you can simulate code Obviously in computer and then run it and then get feedback and then you can iterate so AI for programming is a major opportunity and it will get easier and easier over time a question I wants to

ask you because you used to work at uh sales forces what is Mark benof talking about when he's talking about the millions of Agents what is what is that Vision that's coming I don't want to speak for him uh but I think no but what in your interpretation of you know Enterprise scale agents in something like that environment what what does that mean for people because you know don't forget what you know versus what the general public knows it's so disconnected right people don't understand the scale of what is happening and how big this all

is 100% yeah let me give you a story we actually uh we launched um with a company we'll we'll talk about them soon uh as a big cybercity company we we launched u.com And you can build agents on u.com we we used to call them modes or assistance uh like a year before people called things agents and then realized like we've again the future was already there just not equally distributed now we call them agents because they are essentially they are what we now interpret as agents what are agents generating agents are just neuros sequence

models like large language models and chat gbt but instead of predicting only the next word they Can also predict what's a good next action to take what's the next button to click what's the best form to fill out the drop down menu the app to use and all of that right and so as you include actions in B digital environment uh in these neuro sequence models to try to predict what's the good Next Step they become agents and so we've launched agents last year where they can decide to search they can decide to program they

can decide to run that code uh and Then merge all of this they can decide how to visualize the data versus just text and so we've had these agents for a while what we've seen is that companies will get you know a couple hundred seats uh like maybe 100 200 seats from u.com once we do a workshop with them where we actually go through all the examples of what they can do it goes from 100 to 200 to like 700 seats yeah everyone's limited by what they don't know how to deal they don't know how

to use it they Don't even know how to ask chat gbt they they get scared exactly this goes back to this idea that we're all becoming managers but it's not intuitive for a lot of people how to manage nii right and so when you like I'll give you an example right like the the marketing team is like oh maybe I'll like ask it to rephrase this thing but that's not a big use case and we asked them oh what do you do and they're like well every like first of the month we Get a new

feature release from the product and engineering team and they send us this big document uh with all the like the new features that come out so what we then have to do is we have to go on the web we have to compare these new features to the competition then uh we distill them we summarize them and then we write out like three LinkedIn messages three tweets uh and two email campaigns for the two industries that we're uh responsible for and we're like Let's describe that to an AI and now let's drag and drop that

document into that custom agent and then boom it just like it goes through all those steps it goes on the web it does the thing it summarizes it and writes the messages and they're like holy [ __ ] this is like six plus hours every Monday like first Monday of the month but the hard thing is it's the inertia of actually doing that step of teaching it right that's the bit that Everybody gets stuck at as you said the managing of the AI because like with a new member of Staff you have to train it

and you have to train it not in complex technological stuff you just need to tell it all the steps and most people aren that disciplined as a radiologist you will be because it's a medical procedure you've got certain rules and criteria marketing people less so accountants yes lawyers somewhere in the middle that's exactly right yeah it's Also interesting when you look at lawyers and if you tell a lawyer that getss paid by the hour to say hey you can just be twice as efficient in doing your work they're like I don't I don't really care

because I get paid I get paid this amount per hour whether I do twice to work or half to work like I get paid this amount of hours until you actually get fired right and like until you get fired like it's fine to do as little work as possible You just need to do enough to not get fired now of course if you want to get promoted and so on there like different incentives right but hourly waged workers probably love AI efficiency increases the least it's people that are entrepreneurial people want to excel at their

work people who are partially owners of their company which is easier for startups right all our employees have equity in the startup so they are all partial owners they love those Efficiency gains from Ai and even like for legal it's complicated in the sense that they're company internal lawyers they just have a ton of work and they get paid not by the hour they get paid just to help the company not die and not get sued and all of that right so they have infinite work they would love to get some help because they will

still need them uh in in the company right but if you're just an external hourly paid legal uh worker or lawyer like you don't Necessarily love 20 30% efficiency increases you don't no because you're not driven by productivity you're not driven by productivity you've still got your 60 hours a week that you work of which you can build for and you don't give a [ __ ] how productive you are as long as you build that number of hours the moment productivity comes into the equation I can I do double the amount of work and

get paid twice as much okay then it becomes a no-brainer in in using This because the productivity from this stuff is very very real exactly exactly one of the things I'd love to hear your thoughts on is what are you seeing with the disruption of Education oh oh boy yeah we we we actually work with uh quite a few uh educational organizations I'm excited about some announcements that uh we're working on right now with a large University uh and University system um education is going to massively change Right uh I think generative AI uh the

sort of rule I came up with is generative AI is immensely useful if it takes you a long time to create a work product some kind of artifact but it is very quick to verify that that artifact is right now for example illustrations you can very quickly create an illustration um it would take you a very long time if you did it manually but I can do it quickly and you Can look at it in a very short amount of time second that looks great and you're done right so that's a perfect use case for

Gen now in language it's actually different if ni writes you a 50 page document and they're no easy to verify citations and it takes you forever to make sure everything is correct in that document it becomes less and less useful than just writing it yourself um and that's why at u.com we push so much the citations and actually when you use Research mode you click on a citation it sensor directly to where it found that fact right and that way you can very quickly verify that this fact is actually true because based on the source

and so on and you see it directly marked in the browser scroll down like this is where I found that fact so um what that means is that we need to teach kids how to use these tools I think to say like oh we should just outlaw them in education makes no sense right it's Just saying don't use a calculator you've got to always do the mental arithmetic in your head it's like why right but you also need kids to still be able to think creatively think on their feet be able to hold a conversation

and so what that means is you're probably going to eventually have to test more stuff in class right like test a conversation can you argue can you debate a thing and at some level you need to have the facts memorized because It's hard to be like in the middle of a conversation like let me try to find quickly an example when I try to convince you like AI agents work great for marketing right like I just did let me try to quickly find an example for AI agents for marketing and then you wait for it

and you like get the answer then you say like it doesn't work right you need to if you want to creatively think through a space in science especially right chemistry biology there's a ton of Knowledge if you don't have the knowledge it's hard for you to be creative uh and identify boundaries of that knowledge and how to push it forward and all of that so we do still need test kits for a lot of things some of the testing will have to be in the class with no internet uh but a lot of times we

need to teach them to have good discernment good judgment for an AI product versus just creating the product the knowledge product yourself and Scratch the skills of Being Human the social skills the things that we lose in a online World become more important to a child when they have access to infinite knowledge like everybody has access to water that's basically What's Happening Here infinite knowledge of the scale of water so okay we don't need to now go to the well every day and find our water we don't need to accumulate knowledge in the same way

but we still have to interact with humans you know And interacting the world around us and I guess that is what we end up getting taught more more than the retention of knowledge like we don't need to retain the ability to do math um maths in our head anymore because we use computers or calculators that's right and what that means fortunately is that we can operate at higher and higher levels of abstraction right like that's sort of the magic of civilization uh and humanity is that you don't need to Understand how this toster really works

like if you drop most people in the jungle you say build me toaster that actually runs like no one could do it right like you need to have electricity you need to understand how materials you need to and there's so many things you can need to understand so but we can build these levels of abstraction um and what that means for education is and even computer science we see this and most beautifully work out well in Computer science right you don't need to work in assemble it uh like you don't need to understand he vary

bits and bites of his CPU to now build a whole website with like five lines of code right and soon uh you can build a much more complicated website with five lines of English right uh but if then there's something off slightly and you're like well there's a bug and Android and this version and so on and you it's still useful to know how to fix it right and So I think the levels of abstraction in education need to keep up with that understand help people under and students understand the basics uh so they can

then uh really creatively put it all together when I look at all of this I don't understand how the economic Machine Works any longer because you know I look at GDP growth as driven by population growth productivity growth and let's say debt growth population growth is now going to be infinite Because of AI and robots are infinite productive units and productivity goes higher I don't know how the economic machine is even recognizable even the study of economics is basically a study of human supply and demand and it all kind of disappears we actually built an

AI Economist uh when I was working at C Force we had this really exciting paper that really hadn't had hasn't had its moment in the sun yet uh because economics is an extremely slow moving Field and you can't really prove um to economists that this is a better model but you can use AI also can model economics right so what we did is we built a two-level reinforcement learning problem where you had AI agents that just try to build uh and maximize their own utility they want to build houses they want to collect resources they

learn to block other agents off of those resources so they can have them all to themselves and then uh so on theyd Maximize their own utility and then you had another agent that set um uh taxes and subsidies in order to optimize a specific objective function this case the objective function was to maximize productivity of the overall economy multiply it with equality uh right and so if someone was able to completely block off a resource and be like get Monopoly power they would get much more heavily pled and the folks that can't Get access anymore

to that resource get some subsidies and it can learn basically much more optimal taxation strategies than anything we have right now in the world um I wish economists like looked at this paper a little bit more carefully um it was unfortunately rejected at some major journals by some weird like ethics philosopher like person it was like was very unfortunate that paper has so much potential for Humanity uh but the field of Economics Is mostly like run uh based on Vibes uh and uh and not math uh and so um like I think one the future

is already here it's just not equally distributed right there are still people who don't have access to water right in your example and there's like fights over water like you know there's a big damn denial and and like Egypts like stared that they lose a lot of their water source and their lifeline and all of this right and so there's Like even for the most basic things it's not going to happen all overnight I think the countries companies and countries that Embrace this change are going to first slowly and then more and more quickly move

ahead and become more and more efficient uh and then indeed I think you're right at some point the partnership of labor for um for uh income is going to change well knowledge for income right that's the other big one that's even bigger knowledge yeah Knowledge work and yeah everything um and so I think in a weird way Europe is well situated and well positioned to protect people from the downsides of AI but unfortunately isn't well positioned to actually get the additional efficiency increases because they over regulate everything they don't invest enough in and so on

the US is very well positioned to use the upsit they have amazing technology in Silicon Valley a lot of people want and do have this Constructive optimism to make the technology get better and build amazing things but then they have very little infrastructure uh in their social market economy to like support people who jobs are going to change like education is really expensive you definitely don't want to go to college multiple times because people are already like the hug debt after going to college once so like you know if College isn't free uh which it

is in Germany like you don't want to Go there multiple times and learn new things uh unemployment benefits are not great the healthcare is tied to your employer so there are all these things that Germany is well positioned to help its populace and and its people to like uh prevent the downsides but yeah I I wonder what the par optimal boundary here is uh in terms of uh just the legal and and Welfare systems to still encourage a lot of hard work uh the way the US does and having that Entrepreneurial mindset um versus just

trying to say hey let's just not rock the boat so much just things are so nice and we get so many social benefits already um it it's going to be complicated I think shortterm I'm a little bit worried longterm I'm extremely optimistic the long-term Arc of humanity is very positive yeah and also I think that with your um your understanding about how AI can help economics is the same for governments The mass mass allocation of global resources you know country level resources there's no reason an AI could not be a lot better at government the

system of government much more efficient and the same with running corporations and AI probably does a much better job at the majority of tasks within a corporation um than humans do or it will be in a few years time you know five 10 years time that's right yeah eventually eventually I think I can be very very Good at a lot of different things I think for now of course it can already be a good adviser now the truth is that there's so much identity politics right now right that and like people say oh if you

vote for me I'm going to give you these tax credits if you vote for me I'm GNA make your life easier if you vote for me I'm gonna prevent AI from happening because you're only going to vote for me if I promise you to keep your job and so I'll promise you to keep Your job and then you can't use AI supposedly as a politician or like improve it and make sure it works right and it's it's complicated it's a messy very messy process and there just humans don't always want more efficiency right and there

I think ultimately different societies will choose different paths right and personally I love like the future I love like automating all the boring repetitive things uh away but some people get a kick out of gardening And just doing that kind of work right and so they want to continue doing it and at some point we're going to see you know like what we already see right now which is like there's some islands in Greece some beautiful islands in Italy they're not all thinking about efficiency every day they're fishing they're enjoying the Sun the ocean and

and their lives and they don't want AI to automate all of that enjoyment uh because they're already enjoying it Enough uh and so um and I think the countries again and companies that are embracing it they will just become more and more uh wealthy and and pull ahead more and more in productivity and then there is more and more inequality between the countries that can Embrace Ai and and do embrace it and and the ones that don't yeah it feels that humanity is going to go through a big fork and there's going to be the

two different ideologies there's the Deceleration and the accelerationist who will end up merging with the machines in whatever format that means and the others will opt out and it's kind of like the Amish opt out of modern contemporary Society I think we will see that whether that force that creates a new human much like we had NE Andals and Homo sapiens living together for 50,000 years and then one disappeared you know it's it's such a big thing that's happening people just can't get their Heads around it so it's just fascinating thank you for your time

for allowing me to just think through some of this stuff because none of us really knows it's happening faster than humans can even imagine we're not very good with exponentials and we're not very good with Reed's law which is like meta's law squared which is what this is is doing um so you know it's just kind of strap in and go for the ride but listen thank you for what you're doing uh uh thank You for joining us it's been super interesting thank you it was a fun conversation so I think an incredible conversation to

give us yet another layer of understanding of what is happening now what is interesting is within AI is there's not a consensus of what is happening you know some people don't think there's Consciousness developing other people do there's a lot of things all happening at the same time the rise of Agents we're seeing that That's overlapping into crypto world so we really have to keep our eye on all of this and it's one of the things that we do at real Vision we really try and follow this I mean I have a whole research service

called the exponential that is just dedicated to this this six years how we can make the most of it but real vision's got your back so if you haven't signed up go across to real.com you can use our AI tools you can see what our community is talking about and Meet thought leaders in the space it'll make a massive difference and it's free I'll see you on the platform hey thanks for sticking around to the end uh look if you enjoyed it hit the Subscribe button and check out the video here on the right hand

side I'm sure you'll enjoy that one as well and if you're ready for more go to real.com jooin I'll see you there