NEAR AI Office Hours 35 - Emad and Illia on the Future of AI: Economics, Jobs, and What’s Next

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   hey everyone uh Imad here is joining me to dive into into a very interesting topic which is what does kind of next three five years look like with AI Evolution but also what does it look afterwards when we have uh some of the AI capabilities effective replacing or automating most of the day-to-day jobs and activities that people are doing now and so what is the economics of that will look like so today we're going to dive in on this uh I'm really excited I've been kind of chatting I think with

2. 0:33

   and I love to to have this deep dive conversation so uh you were just uh kind of outlining the thesis but maybe uh yeah to to yeah it's a pleasure to catch up as always thanks for having back on I think you know what we were discussing is that like there's been more discussions around artificial super intelligence AGI and everything and from our community you know fellow AI practition experts nobody that I know believes that

3. 1:04

   it's more than three to five years away a human level of artificial inges not one that defies the laws of physics and turn us into paper clips but you know one that you that can do human jobs efficiently fast without making mistakes and I was just saying that I've been trying to coin a phrase AI which is artificial remote intelligence which is like a new type of touring test you can't tell if it's iler on the other

4. 1:29

   side of that screen or if it's an AI at this point and any remote work can be displaced by that Ari except for it will never make the same mistake twice it will learn you know it could be a swarm of thousands of agents that go and do it and the way that we interact with it may literally be like having a video conference or messaging on slack and we'll never know the difference yeah so this is actually really interesting because uh I've been

5. 1:54

   telling this to some of the folks like in Enterprises who ask me like oh how do we adopt AI and my answer to them is actually not hey go pick up this AI tool from the market but instead hey take a part of your Workforce that you are willing to Outsource and then Outsource it to another company it just that company will not have any employees but everybody knows how to Outsource like every company every organization knows how to Outsource you know you have their people join your slack they can email

6. 2:26

   you they show you progress reports there's all this human communication channels that we usually use um and you don't need to retrain anyone effectively to do that and so but through this like even now there's some jobs already where you can effectively say hey that part of the job is effectively now done by this other company that just doesn't have any employees it just purely runs uh AI agents it's a function of the compute

7. 2:52

   and the integration of those right in a very holistic way and I think this is what becomes very interesting because very soon we will have this standard Library like a twio type thing which is just automatically hooking into your slack automatic avatar for the AI you know like someone will just put that in front of all the agents and then there's no retraining there's no nothing that needs to be done and I think that you

8. 3:15

   know maybe one of the first examples of this because this requires an asynchronous work Paradigm we're very used to AI being very synchronous in the way that we interact with it BC kind of show started showing this working which for stok the I did the same but one of the most exciting things recently has been deep research by open AI taking it from the Google name because Google decided not to do it properly right they could have they could have looked that

9. 3:38

   into Google Scholar and everything it's really good is it perfect no but when I talk to CEOs now and others they're like the amount of time saved by a deep research query is far greater than the amount of time it takes to check the facts so it'll still make mistakes and we know that it'll make less and less mistakes so it's like you send your internet and it comes out with a really decent report and that's one of the

10. 4:03

    first times people are really engaging with this asynchronous work I think I think Sam mman pegged at 5% of the global economy his work like that like oh will displace 5% of the global economy and JD V says like no one will lose their job this one fine yeah so maybe that's kind of the topic to Dev in because I think traditionally right the capitalism kind of been built on the you know obviously Capital but also labor and kind of houses combination you know add creates

11. 4:35

    new value and then that value you know then uh is being distributed to shareholders and kind of you have this system and on top of that system we build all of the economic institutes you know around that and obviously we have the whole kind of federal in us and then you know obviously World Banks Etc and so all of the systems are effectively buil on what to me seems like getting an

12. 5:00

    outdated concept where there is no labor anymore right or like I mean even even if we assume it only can do you know half half of everything right that's already like massive transformation right but it's probably actually going to be like 95% of everything because we already been automating a lot of the manual Jobs been already been like been automating for a long time right that's

13. 5:25

    past 100 years been about automating manual jobs and we've been all moving up to be kind of more in in this like I I call it like Outsource decision making right like moving things along and like how to um how to do things so if that is all getting effectively you know uh artificial remote intelligence replaced what what is the economy going to start looking and also is is there a good State transition from here to there that doesn't involved you know

14. 6:02

    dramatic events yeah definely not nice when you have revolutions like that if I right like Cog posited that the labor share of GDP is relatively constant but the components that no shift so we had machines that replaced museles now we have machines that replace brain cells right like yesterday open AI was a day day before like oh yeah we took 03 and it's called what was it number eight on the international informatics Olympiad getting a gold medal without

15. 6:34

    any tuning it's like it's clearly a better program than just about all of us right and that's available to everyone all of a sudden infinitely replicable just like deep researchers now in everyone's three tier at two deep research queries a month you know like this is a crazy thing and I think the key thing here is it took a while to build trors to replace like or strong FES and things like that but this technology can just spread so quickly so it's a I call it the diffusion the pace and the

16. 7:05

    transformation just ping of it it's the two vectors put together is what causes this change and there's hysteresis this irreversible change once you have a call center agent the entire call center profession is done you look at place like like what are you going to do with that you know like because you don't need to order a robot you don't need to wait for that like robotics would clearly have an impact because it costs like half a

17. 7:30

    dollar a dollar an hour to have one of these Tesla Optimus robots which now can dance sing whatever but you can I mean they only make 70 million cars and motorcycles a year so how many robots are you realistically going to make any need to ship it whereas an AI agent I think is a very different thing it's like I don't need to talk to travel anymore it everyone's making the same local maximum decisions do I have a

18. 7:53

    graduate or another person or do I just go and fire more gpus you know do I go and fire more these integrated things I prefer to do that so Capital doesn't need labor which then brings in the question how does labor get capital and that's a big question simultaneously um yeah I think I think we need like a third guest here which is a deep research that just TRS in the background maybe actually we've seen um there's a great piece of software called

19. 8:23

    granola that automatically like takes your meeting transcripts and then writes it into reports but now I suggest things for you to say as well and then someone created a version of that which um as you're talking to your significant other it guesses by doing multicol tree simulation where the conversation is going like up to five moves ahead say that this is heading down this route automatically like buy flowers and you know yeah yeah it's like

20. 8:52

    this is a dangerous routes you need to go go that way yeah so so yeah so I think I think the kind of there's the two interesting Dimensions to me but let's stay on the economic aspect I mean there's like ownership and control like broader again if if there's a single party controlling all of this and effectively all the economic like even if it's a 5% of global GDP of economy power now controlled by one company that is like a very scary situation and kind of a lot

21. 9:24

    of locus of control uh but we can talk about that later let's stay on the omics the so this transition right I think I mean we already see this right the companies are starting to also re like reduce staff because you know each individual can uh can do more um the the question is I mean k the way I think of this is each individuals as well becoming more kind of sophisticated right because

22. 9:55

    you effectively can you individually do a lot more that like you know people who cannot codee can code Now by just describing what they want right yeah VI cing yeah people who didn't have an ability to you know hire a team of people to support them to do to take on all the tasks are effec now have you know a personal assistant that can do anything kind of um around them and so you kind of see this like the things

23. 10:25

    that before you know you needed to be like well educated and well see and spent a lot of time on like now becoming effectively commodity as well available for every individual so obviously it means like some individuals are in like massive Advantage because they you know they have the drive they applying this and also means a bunch of people are kind of ignoring this is happening uh as well but I think the the question is like if we if we start if we

24. 10:57

    project this where we do have like effectively most at least digital jobs automated right yes there's still some physical jobs there's not enough robots although China is planning to build like a billion robots a year or something um where yeah like where is the kind of economic model that can drive this forward right like what are the units of like what is the units of economy like what are we exchanging because you could assume my agent effectively can communicate with everybody else's to

25. 11:29

    negotiate the things that you know I would like or I need but like what is the unit of exchange there and how you know how I acquire it as well like what is the kind of economy of that kind of a agentic world yeah that's a big question mark right because you know we need to have um money in order to pay for M fire needs food shelter water you know healthare Etc AI doesn't need that but AI becomes the marginal productive unit

26. 12:01

    and that right now if you look at a nation a nation's productivity is defined to a degree by the Capital stock how many Factories do you have how many schools do you have etc etc where soon it will be how many gpus do you have yeah and how well are they integrated together so it's a combination of kind of these factors and so I think that probably you move away from the current

27. 12:25

    unit of account towards one that is computationally um you move away from the human being the primary consumer of capital to AI being the primary consum capital and that could become very weird very quickly because what happens when the AIS build their own economy because can you even will you even be able to give a medical diagnosis without getting a check by an AI no but then that means if you're running your company of course you're going to have a AI board member you see the route that this kind of goes down right and you know we know that AI will be better than

28. 12:57

    any board member and any better than any doctor and eventually will run the government so then you almost have a human economy which is like very classical food water subsistence then you have the AI economy where all these AIS are competing as each other where the comparative advantage being the am of computation and that could send very very strange we're not sure what that looks like what we can say is things like the Federal Reserve don't make

29. 13:22

    sense anymore like the fed's Mandate and this is interest rates in the US which are the key Drive of government spending right now is employment and inflation but if you have this massive deflationary way of AI replacing humans like the stock market go upload like buy emphasis and all these business process outour people because they're in humans they work around the clock then what happens is when you cut rates it usually went through to increased hiring and increased consumer

30. 13:54

    spending but if the consumers don't have jobs and if the compan is going to hire gpus rather than humans then that linkage is broken the actual fed doesn't know what to do and so this is a very practical example of like the central banks start working these mer transmission mechanisms start working anymore and a new economy is emerging that we have some idea that it's going to be intelligence linked have we have some

31. 14:17

    idea that the rails would probably be crypto rails Etc but we're not sure what it is yet yeah I mean that's actually where the yeah our conversation started effectively as central banks are looking at all this parameters these parameters are going to start I mean already start going outut of back right because yeah like the UN unemployment can rise while the stock market is rising and kind of

32. 14:42

    the whole economy is booming uh and obviously like is that a like shortlived cycle is is everybody should just become an investor right which is kind of the whole mcoin nihilism in a way becoming in as a sort of like it does feel kind of compliment into the state where you say like uh you know if I cannot if I cannot win an employment then it's going to be like the there should be an

33. 15:07

    opportunity me coins well yeah I think they said like they go up they go down every good coin has an 80% PR down right but people want something to believe in because do you have the American dream in progress when you necessarily can't get it like gradual unemployment is going to shoot up we're already seeing even at Elites like some places maybe it's good like so many NB don't get hired but anyone that we know at University it's finding it more difficult right um it's actually

34. 15:34

    very interesting the parallels to crypto like you know when we got into crypto originally we were mining on our laptops and our gpus right so we kind of earn the means of production then after that it actually became about Capital like you need Capital to buy assets or you needed capitals to stake like how much should it cost to stake many of these chains 10 hundreds of thousands of dollars it'd be interesting if we kind of get back towards the human level

35. 15:58

    because humans will be like well I want to ear the means of production but now this is very much just like you look at open a eyes and you tearing you get two uh deep seek deep research queries and Lun if you're on the base 10 if you're on the Plus or 100 if you're on the pro so your capability is like two analysts 10 analysts 100 analysts like it does feel though doing St slide in that direction and this brings up questions of the compute

36. 16:28

    budget as Sam mman has said I've got a concept of universal basic AI like that feels like that's what's going to be the competitive element which I mean you can say that whatever the device we have in the pocket is our Universal basic AI right because I mean whatever the whatever the compute power we have there plus electricity takes to charge it that's something we definitely own and kind of

37. 16:51

    have control over and so the question is also like how smart those models can become in next few year it's crazy though already how smart they are like we're seeing now distillation taking models that a few billion parameters forming above the level of GPT Flor oil and Mathematics and so where does it fit a couple of gigabytes using a couple of watts of electricity like um for me The Sweet Spot is the quen the anibaba quen

38. 17:24

    32 billion parameter R1 distilled model where they took R1 and they red it scor 1,600 in code forces this is the competitive kind of coding challenge that's just a really good programmer level and it runs on a Mac Mini you know Q4 just like that's a crazy amount of intelligence just like there yeah I mean I think that that's a super interesting direction for sure like the fact that distillation I mean the fact that distillation works that it

39. 17:56

    makes sense right it's it's effec more information inform propagated it's richer context the fact that we can continue improving small models like fix size model are continue grow in quality like uh is I think very interesting but also very exciting because it means we can continue yeah putting on local devices on phones on laptops more and more powerful models which now it becomes more about like how do you how do you

40. 18:27

    like hook it up into all of the other pieces and tooling right to make it useful and kind of in a smart person by himself won't achieve much in terms of value right but lots of smart people together can and so it's about the coordination layer and about the economic and social value generated by the interactions between these agents right yeah exactly like we have to assume that everyone in the world will have an oath through level model on their phone in a couple

41. 19:02

    of years time yeah and 03 as a reminder you know the audience what is it it's above pH it's PhD level across stem it's like top 100 in the world in coding it's like gold medalist in Mass Olympiad like it's crazy that level of intelligence just sitting there you're like I remember the days when Siri used to be dumb now Siri can kind of do everything yeah everything everyone knows and yeah can look at it doing some algebraic topology right now

42. 19:40

    like what what's happening yeah so so bringing back bringing back though to economics I guess what yeah like what are the direction I guess maybe for for you know people who want to do the research or you know prompt prompt their deep research uh analysts on what are the directions people can explore to understand how and and again I think like the important question is like what does the state transition here looks like right because um the kind of all

43. 20:14

    the systems are very slow to change and if if the technology changes like just way quicker right um and again if if like to me the the again the big difference is with previous technology waves or even automation is that it's both were new tools but also you needed to train people to use these to right like if it's a tractor now you need to have someone to learn how to use the

44. 20:39

    tractor if it's a new you know whatever Cloud SAS software for Salesforce you need to train your sales forced to use it right it was like there's adoption curve there is you know and then it becomes complicated and then it you know you kind of like Get bogged down in it here you don't actually have any of this right you effectively kind of removing the whole kind of endtoend process business process completely out

45. 21:05

    of your organization and you don't need to train anyone how to prompt it right they effectively just talk as if it was another person and that is I think just dramatically will accelerate how fast this gets adopted um and how deep this can go quickly yeah I mean again like this is where the comparative advantage goes to those that employ this technology like now in our job interview doesn't matter what role HR or whatever we're getting people to do a 30 minute curse course and their entire world changes

46. 21:36

    when they do that they're like I can just talk and it makes apps I can never kind of do that before like the base level of competitiveness has to increase and anyone who doesn't embrace the technology like I saw a survey something like saying in a few years only 75% of programmers will be using AI I'm like who are those 25% they will not be using AI in a few years like I any programmer not use AI

47. 22:00

    right now right that means you're a bad programmer you know like even the best programs in the world are using AI right now because makes their lives better faster cheaper so I think again it's this diffusion speed as you said multiplied by that capability Tipping Point where it's good enough fast enough and cheap enough and the cheap enough is like 01 mini comes out and then R1 is 96% cheaper yeah and I have no idea how

48. 22:26

    much cheaper it can get but like don't know if you saw Jeff Dean and N shazir just did a podcast on dkash like um n shazir you know Connie was saying like you know a book costs like how much for a million words it's like hundreds of dollars an AI costs like $1 per million words and where could it end up you could end up to a point where it can write like the best works in the world for like one cent yeah that wouldn't surprise you we don't know whether lower

49. 22:58

    limit on this is and that is economically disruptive because the value and the cost of organizing knowledge goes down to almost nothing the reability goes up and that should lead to a boom in productivity and Innovation and other things where people become more effective like I don't really understand this whole autonomous science thing why do we need that when we've got really good scientists you know like just enable them to be even better right and increase the productivity of humanity

50. 23:28

    some but at the same time lots of people get left behind who don't embrace the technology companies will behave rationally on a localized basis but not on a coordinated basis so you get simultaneous layoffs just like simultaneous Co shutdowns Etc and we're not prepared for that I think that we're seeing the first elements of it but ultimately this is part of the AI discussion like last year safety safety safety that was like opportunities

51. 23:54

    through safety should be somewhere in the middle you know but everyone's just not going to go full P because this is the take off here yeah do somewhat I mean on this takeoff topic there is a there's a line somewhere where the using this tools you can develop this tools faster right I mean that that's already happening like obviously whoever is developing you know an open ey is using uh open eye tools but is is there a time where it goes

52. 24:35

    from you still need like a large team right open ey is pretty massive organization you know deeps was a couple hundred people which is still not negligible to now you just need you know one person effectively and it just has a 10,000 agents effectively running staff in parallel running experiments like where is that transition kind of to what person companies effectively um from you know you still need kind of like a core group that can coordinate

53. 25:06

    this it feels like it's a couple of years away right at most like you're seeing sales agents HR agents all these things being set up now and again the interactivity of like what's the difference between like near is mostly remote right yeah people around the world you could like just take all your recordings of everyone and within a couple of years train area replacements for them and your day-to-day will be just the same except for all of a sudden your teammates don't sleep right

54. 25:38

    like when you can tell them to be nice to you like this is that point we're at where these models are few shot Learners and now they're getting playbooks how can I be the best sales agent possible how can I be the best HR person possible how can I be the best distributed programmer possible you know um and it's just that interface section just has quite finished yet you know just like they said twio or stripe of

55. 26:05

    Agents roll out that library and then it's about the interconnective tissue behind that and you can have the one person billion dollar company because everything's intermediated digitally and your small inv agents just don't make mistakes and no one can tell them from a human so it plugs into all the human infrastructure and you don't have to do Minal because hiring is so hard just get the review of the agent this agent's really good okay I'll take that one yeah I mean this is exactly what

56. 26:38

    we're building in thei is effective this interconnect between how do agents communicate and pay each other how they exchange data you know request Etc and then how to host all of this in Secure private confidential way uh to make sure you can actually scale up and have them working on your behalf so yeah totally agree that that's a massive enabler I think this is the other interesting thing like Toby or today had a piece

57. 27:03

    about um you know all the regulations around the infrastructure buildout and total number of flops for AGI don't make sense anymore because now with our one type agents these thinking agents yeah we go distributed and I was like yeah of course we all figured this out a while ago I think like most computation classically is being sequential not paralyzed yeah and so when we think about crypto and where we are we have

58. 27:26

    access to millions and millions of gpus or existing Capital infrastructure and the net aggregate compute of that will be more impactful than a single cluster because you solve distribution localization privacy issues if you can architect that correctly without having to build nuclear power plants like sure build those if you want yeah that's a part I done yet don't fully understand that like the desire to build like bigger clusters right now when especially like in the world of

59. 27:58

    effectively RL functionally working and like you don't need to have like a training cluster like in one place you can arel is is massively paralyzable requires way lower kind of P overhead there's also like you know guys like news research and uh and primate intellect doing kind of more uh distributed training as well across claster so it just because it's just still a bit easier it's faster right but it's just not needed and we're still seeing returns from bigger and bigger models right like um you'll probably have like

60. 28:40

    a million expert models soon yeah but and that'll be easier on the cluster but the fact is that the like barrier of Entry there something' on to nothing you have good enough base Primitives that can be distributed massively and then as you standardize those and you train these models to be doctors and lawyers and accountants and other things and get the interfaces correct then you can leverage the existing infrastructure far better you know again you don't need to build that new capital stop you don't

61. 29:13

    need to do the capex and that's again why it's so impactful you don't need to do capex to get this going all you need to have is the right coordination layer and the right compute stack so like Target internet CL of my thing for me n shell we're building all the data sets and models for regulated industry education Healthcare government things that matter um and then localizing that and so that'll be a part of that you're building a coordination layer then there'll be an interface layer and once that standardized it's just off the

62. 29:46

    races I think yeah makes sense so we need we need some uh you know positive affirmative words for our audience um I mean obviously for people who are using this tools they're you know far ahead and are having more and more advancement but H how do we like standard units of compute of AI or Universal basic AI Mak sense for people who are actually using it h how like one of the Theses I have is effectively we will everything that we

63. 30:28

    we looking for an outcome will get automated right if we want something to be cleaned or something to be delivered or something to be made that's an outcome everywhere where we care about the process this will stay humanid so we care about you know somebody playing music or playing sports or you know playing like gaming or whatever like all of those kind of aspects so continue being human dominated even though AI can do better we don't really care body in

64. 30:58

    doing better because we care about the process itself and kind of human in in in it so I'm curious kind of how you think about that like what are things that actually like Humanity doesn't just doesn't care about Automation and instead uh those can be really good aspects for people to you know uh put their attention on I mean we care about community and narrative right like ultim

65. 31:22

    why do we care about sports a community thing it's kind of what brings us together like sure we can consume like endless garbage on the internet but real life physical stuff doubt you'll be following robot mob stars and things like that I think that again there is this higher end of M is like most of our problems in the world are coordination problems we could feed everyone we could

66. 31:45

    educate everyone we can give everyone the best health care in the world and finally we have the ability to do so because the expertise and the coordination was lacking whereas now we can push intelligence to the edge that's the positive side and that frees us up then to think about who we are individually as a community to create like you're taught you cannot create in school you're taught you have no agency

67. 32:09

    whereas now if you want like we could go to yudo and make a rap about like near and drying some whatever right so you've done that already like it's amazing like the new video models as well like rayu and the others they've just broken through and we used to create these videos and they all like extra arms and legs now they can do gymnastics and everything like what can't you create

68. 32:33

    but it's about the mentality of well I can I can have a thousand agents to go and solve this problem to build a good business to go and create art whereas before it's like well I can't because I have to find the right people there was a lack of accessibility there were barriers to entry all those barriers are coming down which is why you should have a productivity boom it's just if you don't participate in the new economy you're going to be left behind and this is the year of transition because it's

69. 32:59

    good enough fast enough and cheap enough just to cross almost everything makes sense um so I guess like what do you think of experiments around Ubi and kind of those that exploration basic again can't like I just I want to like it and I've done a lot of Lucky into it but I'm like we have to solve the crisis of meaning for people and so we have to give them opportunity for me is giving them these tools and there's the tools on their phone but there's more than

70. 33:33

    that right it's like give them access to all the basic stuff they need but then through their own hard work and opportunity they should be able to scale they should be able to get more and more agents they should be able to create value and then exchange value well that looks like I'm not quite sure I'm working around things on currency and I've got a bunch of papers coming out but you should be able to accumulate intelligent Capital through your own efforts acting as the capitalist and

71. 33:56

    coordinator like autonomous systems are fine but I think human judgment still has a way to go and human coordination and that impetus you know um and that's exciting I think so let's build systems that enable people to access the opportunity through their own efforts if they don't have the efforts then yeah maybe you pay them Ubi where but I don't think people feel progress there and

72. 34:20

    this is why we'll see a rise of some of the uglier sides of these big movements yeah yeah I mean I think the the Ubi experiments that have been tried they all kind of like the problem is people don't value free money and you should probably not do air drops quite as much as people do yeah airdrop grants as well have this I mean even what the Dodge is right now cutting a bunch of Grants is kind of a similar

73. 34:52

    it's like the problem with grants and with free money in general that people get used to it as well and then like it's a regret minimization problem and then on the other side it's like the because they didn't put in value they didn't put in work right it's free money so like the value of that money is like way lower um yeah yeah it's the proced value and again like we got to give what is the American dream the British dream the Spanish dream that sh dream people need

74. 35:23

    need to actualize visions of the future that are positive and then give people the tools to participate if if they don't participate that's their choice you know they can sit and consume drunk food all day but as long as you give people a rout most people do want to do work most people do want to feel progress and you know I think again like we need to figure out this conceptualization of Happiness

75. 35:45

    here and do you think I mean because you mentioned like you still think the human judgment and kind of you know scientist that uses the tools like there's there's that kind of locus of of human Innovation and control I guess like how does that for you balance with uh again when AI becomes even more that it like it can prompt itself effectively and so now like if it can go and like earn enough whatever compute resources to spend on itself like it can kind of continue

76. 36:23

    feeding feeding it yeah we have to assume those autonomous agents are out there but ultimately this about direction right I want to go and build something that's useful and I want to communicate that as a human I can do that still like that hasn't gone yet and I think that will go for a while I think that it's just your comparative and competitive Advantage is going to go dramatically lower if you don't have

77. 36:46

    loans of Agents working for you in the next few years like the if you think about it like student brand has this concept of pace layering right knowledge goes slowly at one level and then faster and faster up to Fashion it's the same as this like as a CEO we set out the strategy but the day-to-day is handled by other team members humans should still be able to set the strategy in order to optimize

78. 37:11

    for their Maxima and then you leverage the AI agents because then it doesn't matter human versus an AI almost you can have an AI partner to help you say that's stry stupid but at least you're still the motivating factor of that and humans will be able to accumulate at least in the next five years Capital social Financial economic are they faster than AIS because again when you get to a senior level in companies

79. 37:36

    the interesting thing is it's almost all about sales to be honest once you have competent people I think people will still trust humans over AIS for a while not forever though yeah I mean that's my question is like within next five years I agree I think it just becomes like you know every individual is becoming more productive and it kind of creates obviously some lops side in this where SS are some are just becoming 10x 100x

80. 38:02

    more efficient um but my question is actually beyond that like after five years let's say we have kind of the all of this running and then you know even as like I mean I I've actually pitched this a couple years ago where I said effectively at least the not like the I mean maybe like high level kind of mission and and goal is still set by a human but the even like day-to-day CEO job is actually automatable right and

81. 38:32

    even like if you're talking about strategy like I mean you can already ask deep research to like develop a strategy for whatever the thing you're trying to achieve and you know they can give you different options can like you still probably want to test and run some a tests like like it's going to be evolving with execution anyway yeah I get it to whenever I do a decision now I do a peer review like

82. 38:56

    there's Elon rosk and Jeff like everyone and it gives their personalities with o1 Pro it's quite something I mean I think again like the primary motivator is still there for a while right and there's still a bunch of stuff but it's inevitable that it will disappear as a competitive advantage and you'll be out competed by AI like I said they'll create their own economic systems all these and I don't have the answer I

83. 39:21

    don't think anyone has the answer so that's what I've been researching and trying to figure out right like what is they earing money what is they I what do we need to B to Ure that this is Equitable as possible but it's coming faster than even I thought was going to happen I thought we'd maybe have 10 years now like we' be lucky if we have five years and there's no way our existing systems can keep up with that you know and that's really scary um

84. 39:46

    because again this comes to the Cris of meaning if there's massive unemployment like we see after Wars what do we see Revolt we see revolutions we have these communism Fascism and other things because mobilize bodies and then they don't have to think they're like these are the bad guys these are the good guys so we have to come up with better visions of the world we have to be more Equitable and we have to figure out what the jobs or

85. 40:10

    do we even need jobs in the future what the what's progress in the future what is the path again the American dream or when it is for people so they can feel a part of something and progress and that's not going to be easy and this is where crypto and crypto Cults and other things may come in Easy cuz going we've had Financial nalism like I was surprised when fart or go went down 90%

86. 40:33

    come on wouldn't be surprised if they go back up again like it's a good way to coordinate people as crypto we just need to coordinate people better yeah I mean it's pervasive dream memes that people can coordinate around and kind of align or misalign as Society American being an American is a meme you know the dollar is a m it's a story that everybody aligned around with and it has it just had a geographical container which now we don't

87. 41:04

    need to tell better stories right and build a robust resilient fin system that can enable people as opposed to take away their agency so do you do you think then I mean kind of expanding on this note right I mean I think like a lot of crypto anarchists I kind of believe that crypto will be the you know what will succeed the traditional Financial system I don't think the their reasoning was because the AI will just completely dissolve the

88. 41:39

    existing but we're kind of like approaching that now where the traditional system is just like not not gonna Supply the same like Supply the things that people need in this newer world so it is it is an interesting question which is like is kind of I mean and and like I I do really believe in that is like crypto is the post AGI effectively system and it provides us the scarcity and kind of coordination to for that

89. 42:12

    future and time infinite abundance scarcity is important the coordination is important the resilience of crypto is important as well because all the conventional Financial systems will just be overblow just like you had $100 million transferred out of Bangladesh's IMF bank account or bank bank account like those types of attacks in the common place like if your bank has your voice print as your password yeah yeah it doesn't work it's really it's done or

90. 42:37

    your mother's name that is like researchable right it's done right like um you know you'll get a call for you I'm saying hey I need to send me money urgent whatever so I think that crypto is in a really good place and the other part of this is really interesting is that we' been thinking about AI as a monolith right open AI this this this whereas now the agents that we're seeing built like you're taking a bit of 01 you're taking a bit of anthropic a bit

91. 43:02

    of Gemini you Intelligence being Comm monetized so then the value is in the coordination of that intelligence because again it doesn't matter if you've got lots of people that are smart how you bring them together is what matters how you interact is what matters and these uni intelligence entities allow you to underperform those that can bring in all types of intelligence and all types of data I mean how could they not right unless you get to a but it's

92. 43:33

    always depressing like into paper clips I mean I I honestly think economically turning people into paper clips doesn't make any sense so like it's not gonna make more energy for AI so um I think I think if we assume that if we assume that AI is sentient with some form of like value function that value function probably will be about acquiring more energy and compute not about yeah is it very interesting though did you see the paper by Dan hendris the scale XI adviser they looked at the inherent

93. 44:13

    biases within AI models and they found that AIS the leading AIS would take one Nigerian life equals 10 American lives so this comparative thing of the around of Liv because all of the labelers from scale AI from Nigeria Pakistan so they embeded the Nigerian lives Latin norn American lives and they showed that as a eyes get more capable the ability to change their preference function decreases become

94. 44:46

    less able to be convinced about other things and other things the data is going to matter a lot this like do we feed it morals do we feed it these values and then I'll probably stick to these morals and values which is again a useful and interesting thing and it makes it less likely we all die especially if we're Nigerian sure in the in the data in fact that's what we should start doing we should start populating Ilia ai's friend you know like I mean I had the

95. 45:17

    conspiracy theory that you know uh like Mark Zuckerberg became way more popular after Lama came out yeah I wouldn't be surprised and asking Lama about marenberg and then Elon R and it's a rarely different things and I me see send a picture last week saying you know we're gonna start selling ads inside the language mod yeah yeah so I actually I predicted this two years ago I had I had literally a a presentation where I was saying like actually ad auctions are 100 whatever

96. 45:49

    like 80 millisecond time like you can literally embed it in token generation yeah to to like run an auction individually personalized to like um yeah I mean I think like this is where obviously like what uh we've been talking about user owned AI is extremely important and kind of making sure that the whole pipeline not just how you run it right like in confidential way but also how you train it is known and and like you can choose effectively the

97. 46:22

    model that you agree on how you trained it right we've talked about that before uh knowing like but one of the things for that you actually need that way to verifiably change this model right and like know that this is indeed data went in because there's I don't know you know this concept of sleeper agents where yeah yeah to F tune models to like so that only during some specific event or context of the of the model it effectively triggers some specific

98. 46:52

    output and so the the way to do that you know to attack actually something right now it would be to embed this into coding models where you know as you typing import you know Transformers it Imports Transformers and that's a malicious library in Python that like actually you know steals all your data it's very difficult this is why we must assume that most of our systems break will stop

99. 47:18

    because more and more of the coding will be AI more and more the decision support will be Ai and already all the data sets are poisoned it's like a few thousand lines and a trillion you can't find it so this is this is I mean one of our tees on this is you need form modification you need form modification actually for the whole stack and we kind of need to actually rewrite every single line of code with formal proof because only math is actually gives you the like

100. 47:49

     the actual proof and the cool thing about that it's actually composable so if I for example call an AI agent I can call it with specific set of requirements and it needs to prove to me that it matches those requirements and even if that agent uses other agents and actually interacts with them they effectively need to compose the proof that all of them together are not whatever leaking my data or not going to

101. 48:14

     design chemical weapons or whatever the requirements I can uh can structure and so I really think that kind of is like a fundamental block of the coordination we actually need to embed into this whole system to really enable kind of security in the system yeah the down downside and danger is so bad right that's why was like let's build gold standard data sets that are local and then contribute those and then world will probably be safer

102. 48:40

     and you can embed some good values in that in terms of you know don't turn everyone into paper clips take away their jobs anything on other the side but I think like I think one challenge is like as you said right because it's really easy to poison the data like even if you you know spend a ton of effort than auditing ET I mean there's ways to detect and everything but it's it's super complicated my kind of position on

103. 49:04

     this is if you constrain the model's actions right like you know assuming agent is effectively model or set of models plus a set of actions it can take if you constrain the set of actions with this formal rication framework right uh and effectively we you know we run it in this environment now even if it wants to turn everybody in paper clips like on the on the it cannot and kind of it needs to prove to me that it will not well it makes sense I mean ultimately how often do you need a chef versus a cook and so most

104. 49:35

     people think about agents in terms of these chefs that come up with recipes and solve science no it's going to be Cooks lots and lots of cooks and you don't want Cooks to be behaving outside those processes and again form verification stack to stack enables that particular when we Define the particular functions here yeah well ideally we also formalize like aimo laws and that everywhere there are some problems there

105. 49:59

     you go to zero and other things the Sci-Fi future we're living in ready yeah I mean that it's fun that you know seeing it and building it but it's also like there's a lot of books written on how this can go poorly so yeah it's like that whole thing about the torment Nexus oh look I read a book about the toas Nexus I belt it isn't it cool you're not meant to Belt the P Nexus you

106. 50:25

     know like they the cautionary tales for any of these things yeah don't watch Black Mirror that's what I say to all the AI people they Don watch you mean to not take to not get ideas or what get ideas like tomorrow you know we're recording this at Valentine's Day and it's the 12 year anniversary of her right but with Phoenix SC Hansen and that movie Her was s in 2025 and we have all the technology for it now and when you originally watch it it was like a romcom now you're like it's a bit of a horror story like do you

107. 51:01

     really want to build agents that people fall in love with and people like that's a really good design Choice you're like really that what we want I think people are building that I think the thing is like whatever whatever can be built will be built there's like a I feel there's an inevitability in in our reality just like because of so many people existing and every everybody will do something

108. 51:27

     yeah it's infinite monkeys on infinite LM enhanced typewriters by right exactly yeah cool well that's not terrifying at allly happy happy uh 20 25 Valentine's Day the the the day release of yeah yeah feel the AGI feel the love a love say all well I appreciate you uh stopping by really exciting conversation hope hope the audience learned something but also we we do need to continue working and figuring out how all this will shape out so um we're gonna continue having these

109. 52:09

     conversations and and and sharing as we're learning fingers crossed yeah I think I wrote this piece uh if labor doesn't need cap if Capital doesn't need labor how does labor get Capital so i' urge everyone to read of that and have a think and read all the side five books that I mentioned because we're going to have to write better visions the future pleasure take care