How OpenAI’s SORA Changes the Future of Filmmaking

0:00:04 – The question of will our economic systems change forever?
0:00:05 I think the answer is yes.
0:00:06 – Yeah.
0:00:07 – But how it’s going to change?
0:00:08 I don’t think we know yet.
0:00:10 A lot of people underestimate humans,
0:00:12 especially when we’re talking about AI.
0:00:15 I feel like humans will figure out new ways
0:00:16 to add value to the world.
0:00:17 – And I’m pretty encouraged there.
0:00:19 Like I think people who adopt AI,
0:00:21 they’re going to become enhanced in a way
0:00:24 and they’re going to have an outsized impact on the world.
0:00:26 (upbeat music)
0:00:28 – Hey, welcome back to the Next Wave podcast.
0:00:31 I’m Matt Wolf and I’m here with my co-host, Nathan Lanz.
0:00:34 And we are your chief AI officer.
0:00:36 It is our goal with this show
0:00:38 to keep you informed on the latest AI news,
0:00:42 the latest AI tools and the talk of the AI world
0:00:45 so that you can use it in your life and your business
0:00:48 and leverage these new technologies
0:00:50 that are there for you to use.
0:00:52 Today’s episode is a little bit different
0:00:53 from what we’ve put out so far.
0:00:54 This time we’re doing a little bit
0:00:57 of an ask us anything episode.
0:01:00 We went to X and asked you to send us your questions
0:01:03 about the AI world and what you want to know
0:01:05 and hear our opinions on.
0:01:06 And we got some amazing questions.
0:01:10 So in this episode, Nathan and I are going to deep dive
0:01:13 and answer and give our thoughts on some of those questions.
0:01:17 We talk about things like where is AI art going
0:01:20 and is there room for it to get even better?
0:01:23 We also talk about how the government and corporations
0:01:26 and how people need to change to adapt
0:01:28 to this new AI world that we’re headed in.
0:01:31 And we also talk about the limitations
0:01:32 of large language models.
0:01:34 How big can these large language models get?
0:01:36 How powerful can they get?
0:01:38 And we’re going to deep dive and give you our thoughts
0:01:40 on all of this in today’s episode.
0:01:42 So hang out with us and let’s dive in.
0:01:47 So the first question here comes from Vicky Jay over on X.
0:01:50 She says, where do you see AI art and video going
0:01:52 by the end of this year?
0:01:54 I mean, AI art, the thing is like a year ago,
0:01:57 AI art was horrible.
0:01:59 Or like a year and a half ago, really horrible.
0:02:01 And then now it’s getting to the point where it’s like,
0:02:02 okay, how is it going to get better?
0:02:05 Like the newest version of mid-journey is amazing.
0:02:09 ‘Cause maybe it’s still, with mid-journey,
0:02:11 I think the one challenge is it still always looks
0:02:12 like mid-journey to me.
0:02:14 Like it looks very polished and it’s harder
0:02:16 to get it having a very different style.
0:02:21 So maybe having more controls over how,
0:02:24 like little details in the image versus now
0:02:25 where it kind of just, oh yeah,
0:02:27 produces something super beautiful,
0:02:29 but it’s not really easy to edit.
0:02:31 I think by the end of the year,
0:02:32 you’ll be able to produce images
0:02:34 that look even more amazing than now,
0:02:35 but you’ll have way more control
0:02:37 over how you actually edit those images,
0:02:39 which for commercial uses will be amazing, right?
0:02:41 ‘Cause if you’re producing like graphics
0:02:43 for advertisements or anything like that,
0:02:44 that stuff’s going to get way better,
0:02:46 I think by the end of the year.
0:02:50 When all your marketing team does is put out fires,
0:02:52 they burn out fast.
0:02:53 Sifting through leads,
0:02:55 creating content for infinite channels,
0:02:58 endlessly searching for disparate performance KPIs,
0:02:59 it all takes a toll.
0:03:00 But with HubSpot,
0:03:03 you can stop team burnout in its tracks.
0:03:05 Plus your team can achieve their best results
0:03:07 without breaking a sweat.
0:03:09 With HubSpot’s collection of AI tools,
0:03:12 Breeze, you can pinpoint the best leads possible,
0:03:15 capture prospects attention with click-worthy content
0:03:18 and access all your company’s data in one place.
0:03:21 No sifting through tabs necessary.
0:03:23 It’s all waiting for your team in HubSpot.
0:03:24 Keep your marketers cool
0:03:27 and make your campaign results hotter than ever.
0:03:30 Visit hubspot.com/marketers to learn more.
0:03:33 (upbeat music)
0:03:34 – Yeah, it is interesting
0:03:36 ’cause I feel the same way about Dolly 3, right?
0:03:37 Like I’ve gotten your point now
0:03:38 where I can look at an image and go,
0:03:40 “Okay, I can tell that was a Dolly 3.
0:03:42 I can tell that was mid-journey.”
0:03:44 I mean, I don’t know if the normal,
0:03:46 if normal people who aren’t like as immersed in AI
0:03:48 can see that or not, it’s hard for me to say, right?
0:03:50 Like I’m such in that my own bubble
0:03:52 that I don’t know if other people notice that too,
0:03:55 but Dolly creates very similar images.
0:03:56 Like the color palettes on them
0:03:58 always kind of looks the same to me.
0:04:01 Same with mid-journey that’s got like this stylistic thing
0:04:02 where I look at it and go,
0:04:04 “Okay, I can tell that was mid-journey.”
0:04:05 But I think like you said,
0:04:08 I think the updates throughout this year
0:04:11 are gonna be a little bit more nuanced and marginal updates,
0:04:12 right?
0:04:13 They’re gonna be smaller updates
0:04:15 because we’ve already got these AI models
0:04:17 for images specifically
0:04:20 that can create ultra-realistic images, right?
0:04:21 Where else do you go?
0:04:23 I think the focus now is less on realism
0:04:25 and more on prompt adherence, right?
0:04:28 Mid-journey is really, really great at realism.
0:04:30 Dolly still sucks at realism.
0:04:33 Dolly is really, really good at prompt adherence.
0:04:35 You can throw a whole bunch of elements into the prompt.
0:04:38 You know, I wanna drag in wearing a fedora,
0:04:40 eating nachos, watching MTV
0:04:43 and the words party time on the screen
0:04:46 and it will get all of those elements into a single image.
0:04:48 Mid-journey can’t do that, right?
0:04:49 So I think what we’re probably gonna see
0:04:52 is like all of these models
0:04:54 kind of combining to a point where like,
0:04:56 nobody can tell whether it was mid-journey
0:04:59 or Dolly or stable diffusion
0:05:03 because mid-journey focused on prompt adherence
0:05:04 and improves that.
0:05:08 Dolly focuses on realism
0:05:11 and more dynamic color palettes
0:05:13 and optimizes for that.
0:05:14 And we’re gonna get to a point
0:05:16 where we don’t know what model generated what
0:05:18 because everything is just good.
0:05:21 As far as video, I mean, we saw Sora.
0:05:25 OpenAI has made it sound like we might see Sora
0:05:26 sometime by the end of the year.
0:05:28 We might actually get our hands on it.
0:05:29 You know, we were just talking a minute ago
0:05:33 about Adobe rolling in Sora to Adobe Premiere.
0:05:35 So, you know, I think with Adobe Premiere
0:05:37 we might start to get some access to Sora
0:05:40 and some of those capabilities in there.
0:05:41 So that’s gonna improve,
0:05:43 which means more realism in video,
0:05:45 longer prompt generation with video.
0:05:48 Yeah, I think that’s really what we’re gonna see.
0:05:49 I think the other area
0:05:52 that we’re gonna see continued massive leaps
0:05:55 is in text to 3D art, right?
0:05:59 For, you know, for creating like little animated characters
0:06:01 or for creating game assets or things like that, right?
0:06:04 You’ve got Spline 3D
0:06:08 and you’ve got CSM common sense machines
0:06:11 and you’ve got Meshie.
0:06:12 There’s all of these tools that are like
0:06:16 getting really, really good at like text to 3D object.
0:06:18 And I feel like a lot of companies
0:06:20 are kind of focusing in there right now.
0:06:21 Like even mid-journey said
0:06:23 that they think their next sort of frontier
0:06:26 is gonna be 3D generation.
0:06:27 We’ll get back to the show in just a moment,
0:06:29 but first here’s a word from HubSpot.
0:06:32 Curious about what the future of productivity looks like,
0:06:33 HubSpot’s AI tools make quick work
0:06:35 of expediting content creation,
0:06:39 optimizing workflows and elevating data analysis.
0:06:40 Personally, I’m really impressed
0:06:42 by the AI website generator.
0:06:45 You can create a webpage just from a few simple prompts.
0:06:47 That’s mind-blowing.
0:06:48 And the facts speak for themselves.
0:06:51 Over 85% of marketers and sales professionals agree
0:06:55 that AI enhances content quality and prospecting efforts.
0:06:56 So what are you waiting for?
0:06:58 Sign up for free by clicking the link
0:06:59 in the description below.
0:07:00 Now back to the show.
0:07:02 – Going back to video for a second.
0:07:04 I think that, you know,
0:07:06 I’m sure we’re gonna see Sora this year.
0:07:08 Maybe when we see Sora and then like even like
0:07:09 a small upgrade to it,
0:07:11 like something better than the demo,
0:07:13 like in the public’s hands, which is gonna be amazing.
0:07:16 And I think you, you know, by the end of the year,
0:07:18 I think there’s a chance that we’ll actually see
0:07:20 like a viral short film,
0:07:23 like something like five minutes, 10 minutes long,
0:07:25 this entirely made with something like Sora.
0:07:26 – Yeah.
0:07:27 – I think that’s gonna be like the turning point of like,
0:07:30 oh, like wow, because like,
0:07:31 ’cause it’s almost there now.
0:07:33 Like if something like Sora was out there,
0:07:35 I’m pretty sure somebody could make something
0:07:36 that went really viral.
0:07:38 And I think you might even see, you know,
0:07:39 ’cause we’ve been talking about like,
0:07:42 there’s all these great new AI video tools coming out,
0:07:43 but like what’s the use case?
0:07:45 I do feel like there’ll be something, you know,
0:07:46 almost kind of like how Instagram took off
0:07:47 because of filters.
0:07:50 I do feel like there’s some kind of new app,
0:07:52 whether it comes this year or next,
0:07:54 or I feel like there’s something on the horizon
0:07:56 that we don’t even know what it is yet,
0:07:57 but some kind of social app
0:07:59 that will integrate these AI video tools.
0:08:01 And we’re like, oh yeah, of course,
0:08:02 why didn’t I think of that?
0:08:04 Like I think something like that bigger than TikTok
0:08:07 is coming using AI, probably AI video.
0:08:08 – Yeah, interesting.
0:08:10 I wonder what a social media app would look like
0:08:13 in that realm because it’s like,
0:08:15 is it gonna be just a bunch of AI videos
0:08:18 and people are scrolling a feed of AI videos?
0:08:19 – I guess the challenge right now is like,
0:08:21 the filters were free.
0:08:22 A lot of this stuff is quite expensive
0:08:24 to actually do right now.
0:08:25 So like you would have, you know,
0:08:27 had to pay for the generation of the video,
0:08:29 I guess is the one challenge you probably would have
0:08:30 right now.
0:08:32 But so that might make it where like,
0:08:33 it’s more likely to happen next year
0:08:34 when costs come down or something.
0:08:36 But it does feel to me like, okay,
0:08:38 we haven’t had big social app in a while.
0:08:39 We had that kind of like,
0:08:40 what was that one called gas or whatever,
0:08:43 which was like this guy kind of like generated
0:08:45 this out of nowhere and then sold it to Reddit,
0:08:46 I think really fast.
0:08:47 – I don’t remember that one.
0:08:49 – Anyway, it was like half a joke.
0:08:50 – Yeah.
0:08:51 – And then he sold it for like a ton of money.
0:08:53 Like, I don’t know, like over 50 million or something.
0:08:55 And apparently like,
0:08:57 I think he told Sean Prairie from my first million
0:08:58 and a few others like that he was gonna do it
0:09:00 before he did it and he just did it.
0:09:01 But there hasn’t been any other big social apps,
0:09:03 you know, since then that have taken off.
0:09:04 And it feels like, okay,
0:09:06 so AI video probably is going to enable
0:09:07 something really cool there.
0:09:08 – Yeah.
0:09:10 I mean, right now like air chat is sort of a social app.
0:09:11 I mean, it came out like a year ago,
0:09:12 but they rebuilt it.
0:09:15 Now it’s sort of re-emerging.
0:09:16 I think that’s mostly emerging
0:09:18 just because it’s got Naval’s name on it
0:09:20 and people flocked to Naval.
0:09:22 – Yeah, Naval is saying it’s because of AI partially.
0:09:24 He is saying because he believes,
0:09:25 which is funny.
0:09:26 He used to say the text was like the best.
0:09:28 And then now he is switching to saying like,
0:09:30 yeah, it’s all gonna be audio and all,
0:09:31 you know, it’s all gonna be audio.
0:09:33 And that’s why I’m doing air chat.
0:09:35 And apparently he is saying that one of the reasons
0:09:38 is he’s now kind of reevaluating things
0:09:40 because of AI and realizing, yeah,
0:09:41 with all these new AI tools,
0:09:43 you’re gonna be talking to your AI
0:09:44 and then you’re gonna be talking to people
0:09:45 that you wanna, you know, real people
0:09:46 you wanna talk to as well.
0:09:49 And so like voice is gonna be the main medium there.
0:09:51 – Yeah, yeah, I know we’re sort of
0:09:52 on a little tangent right now,
0:09:55 but when it comes to air chat,
0:09:58 the thing that I like about it is that you can’t type.
0:10:00 You can’t open it up and type a response to somebody.
0:10:01 You have to speak.
0:10:03 And the thing that makes that cool
0:10:05 is it sort of solves the problems
0:10:08 that X has right now with a lot of bots, right?
0:10:10 Right now you can’t really bot it
0:10:11 because you’ve got to talk to it.
0:10:15 Unless you’re using a really, really sophisticated
0:10:17 sort of like 11 labs combo
0:10:20 where it’s spitting out real voices and like,
0:10:23 I don’t think anybody’s that advanced with it yet.
0:10:25 But you also don’t get like all the trolling
0:10:28 and all the negativity because when somebody
0:10:30 actually speaks to you face to face,
0:10:32 they’re not gonna be as negative
0:10:34 and as much of a dick basically
0:10:36 as they might be on something like X.
0:10:38 – Did you ever use Clubhouse?
0:10:40 Like when back in COVID? – A little bit, yeah.
0:10:41 – Yeah, yeah.
0:10:42 So I mean, Clubhouse was like blowing up,
0:10:43 especially in Silicon Valley.
0:10:47 Like everybody was on it and it started off great
0:10:49 and then it went downhill.
0:10:52 And you know, it’s like,
0:10:53 ’cause it’s really hard to moderate these things.
0:10:54 Like yeah, at first everybody’s nice,
0:10:57 but then you got people talking about political things
0:11:00 or whatever and like people from the opposite side join
0:11:03 and like it gets really hateful and it just, you know.
0:11:04 And then those kind of channels became
0:11:06 the most popular channels
0:11:08 because there’s more engagement there.
0:11:10 It’s I don’t even wonder what Air Chat’s doing
0:11:13 to kind of make sure they don’t have that same problems.
0:11:14 – Yeah, it’ll be interesting.
0:11:16 I’m new, I’ve only been on it for, you know,
0:11:18 48 hours as of this recording.
0:11:22 So, but anyway, we can move off that tangent for right now.
0:11:25 So let’s, the next question,
0:11:27 this one comes from craze in the dark
0:11:29 who’s actually somebody that works for me.
0:11:31 This is his name’s John, he’s my creative director,
0:11:32 but he asked this question.
0:11:35 He said, “How can governments and corporations
0:11:39 “addressed to a potential massive AI job replacement?
0:11:41 “Will our economic systems change forever?”
0:11:44 You know, real easy softball question.
0:11:45 – Okay, next question.
0:11:48 – I do have like really mixed feelings here
0:11:49 ’cause like, you know, in terms of,
0:11:52 will it change our economic systems forever?
0:11:55 I think that was so hard to predict.
0:11:57 Like anybody who says that they have a predicted,
0:11:58 like that they know what’s gonna happen there
0:12:01 is lying or like, you know,
0:12:03 deleting themselves because AI introduces
0:12:06 so many new variables into the world.
0:12:09 It’s so hard to know, like, okay, GPT-5 is gonna come out.
0:12:12 Is it 10% better, 50%?
0:12:13 What does that mean?
0:12:14 What does that mean for the economy?
0:12:17 What does that mean for how people work?
0:12:20 Will these systems stop improving as fast as they are now?
0:12:21 Like in three years,
0:12:23 ’cause we hit some kind of limitations
0:12:27 or will they just start, you know, growing exponentially?
0:12:30 And so it’s really hard to know, but, you know,
0:12:33 and in the future, if AI can do most work,
0:12:35 that definitely opens huge questions of like,
0:12:37 what’s the purpose of the government?
0:12:38 – Yeah.
0:12:41 – Like, what happens when AI gets incredibly smart
0:12:43 and starts telling us that our government’s messed up?
0:12:45 What do we do then?
0:12:48 – We elect an AI to run our government, obviously.
0:12:49 – Yeah, yeah, do we ask the AI
0:12:51 how we should be structuring our government based on,
0:12:55 you know, and how do politicians respond to that?
0:12:57 Do they decide, okay, we have to ban AI now?
0:12:59 That’s probably what would happen.
0:13:01 So I think no one knows, like, you know,
0:13:03 ’cause it’s like, right now, you know,
0:13:05 capitalism is, in my opinion, the best system we have.
0:13:08 It’s definitely not perfect, like everyone knows that.
0:13:10 And so in the future, when AI does the work,
0:13:12 people are gonna have to find meaning in other ways
0:13:15 and also how does money work then?
0:13:17 I think Mark Andreessen talked about this recently
0:13:20 where he believes that AI will dramatically bring down
0:13:22 the cost of many things, right?
0:13:24 And then there’ll be certain niche products
0:13:26 or niche or maybe like real life experiences
0:13:29 that still cost a lot of money, right?
0:13:31 Like going somewhere and having some amazing experience
0:13:33 or having a human crate, something that could still
0:13:34 be expensive.
0:13:36 – I think that’s giving companies too much credit.
0:13:38 I think it’s giving companies too much credit
0:13:40 ’cause what we’ve seen so far, right, like,
0:13:42 there’s so many companies that are built on the back
0:13:47 of GPT-3, GPT-4, and over the last several months,
0:13:49 so many of these APIs have lowered their costs,
0:13:51 lowered their costs, lowered their costs.
0:13:54 Have the companies that we’re paying our monthly fees to
0:13:57 lower their costs or just pocketed extra profits?
0:14:00 – Yeah, well, I mean, so it depends on how good AI gets.
0:14:02 So like, that’s when there’s like a handful of competitors,
0:14:05 but if you have like a thousand competitors
0:14:06 and they’re all competing on price,
0:14:09 because AI has made things so much easier to build,
0:14:12 you know, a lot of prices should go down over time.
0:14:13 But yeah, you’re right.
0:14:15 I mean, iPhone’s still incredibly expensive.
0:14:18 Other things that we buy are still incredibly expensive.
0:14:19 So you’re correct there.
0:14:22 I’m not sure, you know, but also at the same time,
0:14:25 I don’t really, you know, like universal basic income,
0:14:27 which I believe Sam Altman originally was one of the people
0:14:28 who kind of like didn’t experiment with that,
0:14:29 where they gave people money.
0:14:31 I believe they gave people money in Oakland.
0:14:32 They ran that experiment.
0:14:33 And I can’t remember the results,
0:14:35 but I think it was kind of mixed where it’s like,
0:14:37 yeah, some people just end up kind of just like
0:14:39 playing video games, you know?
0:14:40 It’s like, okay, is that what you want in your society?
0:14:41 Like, I don’t know.
0:14:45 Like maybe there’s, you know, maybe some,
0:14:46 you know, portion of people, that is what they do.
0:14:48 And then they go off and do some productive things,
0:14:49 hopefully.
0:14:51 – I think everybody worries about the Wally scenario,
0:14:53 right, where everybody’s just kind of
0:14:54 sitting in their chairs, floating around,
0:14:56 getting fat, drinking slurpees,
0:14:58 and watching movies or playing video games all day.
0:15:00 – Yeah, but all, you know, recently there’s been this big,
0:15:03 you know, uptick in like people are getting healthy,
0:15:05 at least in like Silicon Valley circles and stuff.
0:15:07 And even me, I lost a bunch of weight.
0:15:10 I’ve been using AI even to give me advice on like,
0:15:12 how to go to the gym and like how to work out
0:15:13 and gain muscle and all this stuff.
0:15:15 And it’s the people were like, yeah,
0:15:16 so we thought it was gonna be like a Wally,
0:15:17 everyone’s super fab.
0:15:19 It’s gonna be like a Wally, everyone’s super jagged.
0:15:20 – Yeah.
0:15:21 (laughing)
0:15:23 – Like, oh, it’s free time and AI’s helped them,
0:15:24 teach them how to be healthy.
0:15:26 We have all these new drugs that make you healthier and stuff.
0:15:30 And it might be like that, like a Wally, but we’re all jagged.
0:15:33 So yeah, I don’t know, but my gut feeling is
0:15:37 that there will always be like the capitalism will just
0:15:41 evolve, like it’s not gonna disappear anytime soon.
0:15:43 In a thousand years or, you know, whatever who knows,
0:15:47 maybe we’re in Star Trek land, but no time that soon.
0:15:48 – Yeah.
0:15:50 (upbeat music)
0:15:51 – We’ll be right back.
0:15:53 But first, I wanna tell you about another great podcast
0:15:54 you’re gonna wanna listen to.
0:15:58 It’s called Science of Scaling, hosted by Mark Roberge,
0:16:01 and it’s brought to you by the HubSpot Podcast Network,
0:16:04 the audio destination for business professionals.
0:16:06 Each week, host Mark Roberge,
0:16:09 founding chief revenue officer at HubSpot,
0:16:11 senior lecturer at Harvard Business School,
0:16:13 and co-founder of Stage Two Capital,
0:16:16 sits down with the most successful sales leaders in tech
0:16:19 to learn the secrets, strategies, and tactics
0:16:21 to scaling your company’s growth.
0:16:23 He recently did a great episode called,
0:16:26 How Do You Solve for a Siloed Marketing in Sales?
0:16:28 And I personally learned a lot from it.
0:16:30 You’re gonna wanna check out the podcast,
0:16:34 listen to Science of Scaling wherever you get your podcasts.
0:16:37 (upbeat music)
0:16:39 – I don’t know, I feel like a lot of people
0:16:43 underestimate humans, you know,
0:16:44 especially when we’re talking about AI.
0:16:47 I feel like humans will figure out new ways
0:16:49 to add value to the world.
0:16:53 Right now, one of the ways humans add value to the world
0:16:56 is doing the actual construction, building the stuff,
0:17:00 doing the actual work that needs to get done.
0:17:03 But I do think in a sort of post-AI world,
0:17:07 maybe call it a post-AGI world where robots and AI
0:17:09 can handle a lot of this stuff for us,
0:17:12 I still think humans are going to figure out
0:17:14 new ways to add value,
0:17:16 we just don’t know what that looks like yet.
0:17:18 Also, questions like this about how can governments
0:17:19 and corporations adjust.
0:17:23 My thoughts on that are not necessarily
0:17:25 that corporations and governments need to adjust,
0:17:27 it’s that we need to adjust.
0:17:29 The people, the general population,
0:17:32 the public should be adjusting.
0:17:35 You know, what we’re seeing right now,
0:17:37 I think is one of the greatest opportunities
0:17:41 in anybody’s lifetime to quickly build and iterate
0:17:44 and optimize and create something yourself
0:17:45 that adds value to the world.
0:17:48 So maybe you’re going to lose your job at Walmart
0:17:53 as a checker because, you know, you can just walk out now,
0:17:56 but you also just have the opportunity to use AI
0:17:58 to help you create a business that can generate value
0:18:00 and income for you.
0:18:04 So, you know, we’re in this very empowering point in time
0:18:07 where I don’t think we should be going,
0:18:10 what can governments and corporations do to protect us
0:18:13 from AI and we should be thinking,
0:18:16 what can I do with AI so that I’m not that reliant
0:18:17 on corporations and government?
0:18:21 I think we need to have a sort of paradigm shift.
0:18:24 We need to think differently, right?
0:18:26 You know, not trying to quote Apple here,
0:18:28 but we do need to think differently
0:18:31 in that we need to not look at corporations
0:18:34 and government to solve these problems for us.
0:18:36 We need to look at how can we be empowered
0:18:38 by this new technology so that we aren’t reliant
0:18:40 on government and corporations.
0:18:40 – Yeah, I agree.
0:18:44 Everyone has a free assistant, you know, in their pocket now,
0:18:46 you know, they used to have a computer,
0:18:47 now they have an actual assistant
0:18:50 and that assistant’s gonna get more and more intelligent.
0:18:53 So, yeah, that’s, and that’s probably how you’ll actually
0:18:55 see changes in government and corporations
0:18:57 is like from the bottom up, you know,
0:18:58 like of like people learning this stuff
0:19:01 and then changing the government.
0:19:03 And I’m pretty encouraged there.
0:19:04 Like I think AI, like the people who adopt AI,
0:19:07 they’re gonna become enhanced in a way
0:19:10 and they’re gonna have an outsized impact on the world,
0:19:12 which should give them a better chance
0:19:13 of actually influencing the government
0:19:15 or becoming part of the government,
0:19:16 helping restructure things.
0:19:18 – I mean, the question of will our economic systems
0:19:20 change forever, I think the answer is yes.
0:19:21 – Yeah.
0:19:23 – How it’s gonna change, I don’t think we know yet.
0:19:25 I don’t think anybody is gonna be really good
0:19:28 at predicting that in this moment in time right now.
0:19:30 But definitely, I think the economic system
0:19:32 is going to change.
0:19:34 I mean, we’re already seeing it change, right?
0:19:37 You know, people are trying to figure out ways
0:19:39 to bypass the existing financial systems
0:19:42 and banking systems through crypto and things like that.
0:19:44 It’s going to continue to evolve that way,
0:19:47 no matter how much people sort of fight that stuff.
0:19:49 I do think it’s going to can like evolve
0:19:54 towards that more decentralized crypto-centric finance
0:19:56 in the future.
0:19:57 But yeah, definitely the economic system
0:19:58 is gonna change forever.
0:20:02 I just, I can’t make a prediction on how yet.
0:20:03 – Yeah.
0:20:07 Next question here is from Universal AI Podcast.
0:20:09 Their question is, do you think open source community
0:20:11 can band together and create an AGI
0:20:13 that can compete with the big corporations?
0:20:16 So really, I think the main question is here is,
0:20:20 can open source build AI models that can compete
0:20:22 with an open AI, with a Microsoft, with a Google,
0:20:25 with an Anthropic, with companies that have millions,
0:20:27 potentially billions of dollars backing them
0:20:29 with compute power?
0:20:32 – Yeah, I think we’ve talked about this a little bit before,
0:20:33 not sure if we put it out there,
0:20:37 but I personally really hope so.
0:20:40 Like really, really hope so.
0:20:41 But I am skeptical.
0:20:46 I do feel that open AI is further ahead than people realize.
0:20:49 I think open AI is probably about one to two years ahead.
0:20:52 And when GPT-5 comes out,
0:20:53 I think that’s going to become apparent.
0:20:55 Maybe I’m wrong, delusional,
0:20:56 but that’s my current belief.
0:20:58 And it feels like, yeah,
0:21:01 open source is catching up to GPT-4,
0:21:03 but if GPT-5 is two years ahead of that,
0:21:05 you know, it’s like, well,
0:21:06 then maybe they’ll always be in it.
0:21:10 And then at some point when AI starts self-improving,
0:21:11 let’s say by GPT-6,
0:21:13 that it’s actually kind of improving itself
0:21:15 a little bit too by itself,
0:21:19 you could get in this exponential growth of the AI,
0:21:21 where then whoever gets there first,
0:21:24 or the first two or three companies that get there first,
0:21:26 it gets better so fast that, yeah, sure,
0:21:30 there’ll be cool open source models for individuals,
0:21:33 but maybe the most powerful stuff will always be
0:21:35 in like two or three companies’ hands, unfortunately.
0:21:36 And then there’s the question of like, yeah,
0:21:38 who gets to use it?
0:21:40 Do the government say like, holy crap,
0:21:42 this is so powerful, it changes the entire world?
0:21:44 And yeah, we have to like really, you know,
0:21:46 treat like nuclear bombs or something like that.
0:21:48 And I really hope that’s not what happens.
0:21:50 – Yeah, so when it comes to open source,
0:21:53 one of, obviously the biggest bottleneck is compute, right?
0:21:56 So like, you look at companies like Google,
0:21:59 like Microsoft, like Anthropic, like Mixtral,
0:22:00 some of these companies that are building
0:22:01 these large language models,
0:22:03 they have millions and in some of these scenarios,
0:22:05 billions of dollars of investment,
0:22:09 which they could use on GPUs to train these models.
0:22:11 Even the open source community,
0:22:13 even all of the large language models
0:22:16 that are out there in the open source world right now,
0:22:20 probably costed millions of dollars to train on, right?
0:22:22 So you still need one of these big companies,
0:22:24 like a Meta, like a Mixtral,
0:22:27 it has all of this financial availability
0:22:29 to train these models,
0:22:32 even for the open source to be successful, right?
0:22:34 So I don’t know if open source,
0:22:36 at least not in the near term,
0:22:38 I don’t know if it can be successful
0:22:41 without the big corporations deciding to get involved
0:22:42 and help train these models,
0:22:45 because the open source world of LLMs right now,
0:22:46 probably wouldn’t even exist
0:22:48 if it wasn’t for Meta dropping Lama 2,
0:22:51 which everybody’s sort of kind of fine-tuned
0:22:53 and built off the back of that, right?
0:22:57 So it’s like, I don’t know, at the end of the day,
0:23:00 open source is amazing and we want it to grow.
0:23:02 We want more people tackling problems
0:23:04 and more people trying to figure out
0:23:09 how to make these models a better thing
0:23:12 than less people working on these models.
0:23:13 But at the same time,
0:23:16 you do need millions of dollars
0:23:17 to train one of these models right now,
0:23:19 even if it is open source.
0:23:20 – Yeah, in the future,
0:23:24 like when we’re talking about GPT-5 or GPT-6 level models,
0:23:25 you’re probably talking about way more than that.
0:23:26 I mean, I think the other day,
0:23:29 Google came out and said that over the next,
0:23:31 I don’t know if they said next five to 10 years,
0:23:33 that they’ll probably spend like,
0:23:37 end up spending like $100 billion on AI development.
0:23:39 – And that news came right after Microsoft
0:23:41 said that them and OpenAI
0:23:44 are partnering on a $100 billion data center, right?
0:23:45 – Yeah.
0:23:47 – So basically, OpenAI and Microsoft said,
0:23:50 we’re gonna partner on a $100 billion data center
0:23:53 filled with GPUs so we can train better and better models.
0:23:55 Recently, the CEO of DeepMind,
0:23:58 Demis Hassibis, came out and said,
0:24:01 in response to Microsoft building
0:24:02 this $100 billion data center,
0:24:04 that over the next five to 10 years,
0:24:07 Google’s gonna spend at least that on their data centers.
0:24:08 So that was like him sort of responding
0:24:10 to the fact that Microsoft and OpenAI
0:24:13 were building this $100 billion data center.
0:24:17 But Meta said, do you remember the exact number Meta said?
0:24:21 I think they said they have something like 600,000 H100,
0:24:22 something insane like that.
0:24:23 – Yeah, something, I think a chart came out,
0:24:25 assuming like they have the most or something.
0:24:30 – Yeah, so Meta has 350,000 H100s right now.
0:24:31 – Oh my God.
0:24:32 (laughing)
0:24:36 – Yeah, so, but Mark Zuckerberg did make a comment
0:24:39 recently saying they have the equivalent
0:24:43 of 600,000 H100s, whatever that means.
0:24:44 – Yeah.
0:24:48 – So, I don’t really feel like an underground dude
0:24:50 in his basement is going to train a model
0:24:53 that’s gonna compete with a GPT-4,
0:24:57 they just don’t have the compute availability to do that.
0:24:59 At least not in this point in time.
0:25:00 – Yeah.
0:25:03 – This question comes from aileaksandnews.
0:25:06 What is one thing you’re confident we won’t see
0:25:08 in the next 12 months?
0:25:09 – I don’t know if I could be confident
0:25:11 about anything in the next 12 months.
0:25:12 I mean, that’s a crazy thing.
0:25:15 It’s like, it’s really hard to know where the,
0:25:18 you know, where OpenAI is private,
0:25:19 like where their model is at.
0:25:24 I don’t expect to see AGI in the next 12 months
0:25:26 or, you know, artificial super-intelligence
0:25:28 or anything like that.
0:25:29 And partially because I think we’re probably,
0:25:32 we need, you know, we’re probably limited
0:25:37 by the current compute, but that’s about it.
0:25:39 – Yeah, I mean, I have the same thoughts.
0:25:41 Like, almost every time I’ve tried to predict
0:25:44 something that won’t happen, I’ve been wrong, right?
0:25:46 – Yeah, look at AI video, right?
0:25:47 – Yeah, I’ve made predictions
0:25:50 about how far AI video is gonna come, how fast.
0:25:51 I was totally wrong on my prediction,
0:25:53 and what I thought wasn’t coming for years
0:25:56 came months later, right?
0:25:58 – I was very optimistic about AI video.
0:26:01 Like, I was putting out AI video threads for a long time,
0:26:02 talking about how it was, you know,
0:26:05 gonna really change Hollywood and whatnot.
0:26:06 But yeah, it’s hard to know.
0:26:10 Like, yeah, AI video, it could have taken years,
0:26:12 but now it looks like it’s actually going to be amazing
0:26:13 in the next six months.
0:26:14 – Yeah, yeah, yeah.
0:26:17 – So there’s not much that I would be willing
0:26:19 to like confidently predict in 12 months.
0:26:20 – No, I’m with you.
0:26:22 I think AGI and ASI are probably the two things
0:26:25 I’m fairly confident we’re not gonna see within 12 months.
0:26:29 Definitely not ASI, but most likely not AGI either.
0:26:31 – Yeah, I’m fairly confident that AI
0:26:33 will not replace all jobs in 12 months.
0:26:35 They will not be that good.
0:26:38 – We won’t have Skynet taking over the world
0:26:40 within 12 months.
0:26:41 – Yeah, probably not.
0:26:43 – Probably, I like how you said probably not.
0:26:45 I’m not definitely not.
0:26:47 – Yeah, probably not, yeah, probably not.
0:26:49 I think I saw something the other day where
0:26:52 one of the AI systems in China
0:26:54 with the government it’s actually called Skynet.
0:26:57 – Yeah, a few people have named their company Skynet.
0:26:59 I’ve seen a couple different Skynet companies
0:27:04 and I’m like, you do realize Skynet was the bad guy, right?
0:27:07 – Yeah, I think they thought it was humorous or something,
0:27:09 but yeah, it’s literally China’s AI system
0:27:11 that monitors the public.
0:27:15 – Yeah, all right, so this one comes from Akshay Lazarus,
0:27:18 and I’m sorry if I butchered that name.
0:27:20 But he says, “I’d love to hear you discuss
0:27:21 “the future of tech.
0:27:24 “For example, will UI switching from GUIs,
0:27:27 “graphical users or interfaces, to voice?
0:27:29 “Do we see there being a consolidation
0:27:32 “of all application tech by cloud-based hyperscalers?
0:27:34 “What is the role of startups as data resides
0:27:36 “with these hyperscalers?”
0:27:38 Well, let’s start with the first part of the question.
0:27:40 Like, what do you think the sort of future
0:27:42 of these user interfaces are?
0:27:44 Like less graphical user interface, more voice.
0:27:46 What are your thoughts there?
0:27:48 – I think both.
0:27:49 I mean, I think you’ll have both.
0:27:53 I mean, I think you’ll have a minority report kind of thing
0:27:55 where you’ve got like these visual systems
0:27:58 that you can interact with and they get really intelligent.
0:28:00 And yeah, of course, you’ll be able to also use voice
0:28:02 to interact with it as well, kind of like her
0:28:04 and other movies, right?
0:28:05 – Well, I mean, we’re already kind of doing that, right?
0:28:08 Like that was sort of the point of Siri, right?
0:28:11 That was sort of the point of Amazon Alexa.
0:28:12 So we already kind of have that.
0:28:14 They’re just kind of dumb AI’s right now.
0:28:17 They’re not very great AI’s, but we’re already kind of,
0:28:19 we’ve already had that for a while now.
0:28:21 So I’m going to speak for myself here.
0:28:23 When it comes to like speaking out loud
0:28:25 to some of these apps,
0:28:27 I feel very awkward doing that in public, right?
0:28:29 Like I’ve got these Meta Ray-Ban sunglasses
0:28:32 that have Llama 2 built into them.
0:28:34 I feel awkward as hell in public
0:28:37 asking questions to my sunglasses, right?
0:28:39 ‘Cause you go, “Hey, Meta,” and then you add,
0:28:41 and that’s like saying, “Hey, Siri,” right?
0:28:42 That’s what starts the prompt
0:28:44 so I can start talking to these sunglasses.
0:28:46 I feel ridiculous in public saying,
0:28:49 “Hey, Meta,” and then talking to sunglasses.
0:28:51 I feel like I’d feel the same way with an AI pin, right?
0:28:55 The humane pin pressing a button and then speaking out loud.
0:28:57 Like if I’m sitting in a coffee shop or a library
0:29:00 or somewhere that tends to be a more quiet place,
0:29:04 I’m definitely not using the voice interface, right?
0:29:05 But even just being out in public,
0:29:07 like if I’m walking through my supermarket
0:29:09 and there’s other people around,
0:29:13 I feel like a crazy person talking to my gadgets, you know?
0:29:15 – Yeah, I mean, that might be an age thing too, right?
0:29:16 – Could be.
0:29:18 – I think that young people may, you know,
0:29:19 AI gets better and better,
0:29:20 they may just be totally used to it.
0:29:22 Like of course I’m talking to it.
0:29:25 Why would I, you know, it’s like my son with like Siri and stuff.
0:29:28 He doesn’t understand the world without that, right?
0:29:29 – That’s true, yeah.
0:29:31 – And he also just thinks Siri’s so dumb
0:29:32 as soon as it gets better.
0:29:34 I think he’s gonna love interacting with it
0:29:37 however he can, including being outside.
0:29:39 – Yeah, but yeah, I also would feel very weird
0:29:42 like talking to, you know, AI outside,
0:29:43 especially here in Japan.
0:29:47 People, he’s like a psycho, like what is this guy doing?
0:29:50 – I mean, if really the question is essentially like,
0:29:52 yeah, we’ve got huge incumbents
0:29:53 that are sort of running everything right now,
0:29:54 how does anybody compete?
0:29:56 – I personally think it’s a big open question.
0:30:00 Like, yeah, yeah, will, you know,
0:30:05 when open AI has AGI, you know, what do startups do?
0:30:07 But I think they’ll always be like,
0:30:10 I don’t think, you know,
0:30:13 open AI is gonna want to build every single use case
0:30:14 and every single kind of product.
0:30:16 They’re wanting to build the technology
0:30:18 and have other people build on top of it.
0:30:20 And I think if they tried to do everything,
0:30:21 they’re gonna run into like really major
0:30:24 like regulation problems, if I had to guess.
0:30:27 So I think, you know, I don’t like regulation,
0:30:28 but that might be one area where like,
0:30:30 regulation actually comes to save the day
0:30:32 because I do think the big tech companies
0:30:35 are not going to try to do everything.
0:30:37 They’re going to try to build the best AI technology.
0:30:38 And then on top of that,
0:30:40 there’ll be so many things for startups to build.
0:30:42 – Yeah, and really if the question is like,
0:30:43 how do they compete?
0:30:46 Honestly, our Greg episode, our Greg Eisenberg episode
0:30:49 is literally that entire topic of like,
0:30:52 when do you have these big incumbent companies,
0:30:54 the Googles, Microsoft’s, Amazon’s,
0:30:55 all these companies out there,
0:30:57 how do these little companies compete?
0:30:59 Well, that’s exactly what we talked about with Greg.
0:31:01 And Greg was talking about, well,
0:31:03 you got to create brand, you’ve got to create community.
0:31:05 You’ve got to, you know, what are you going to do
0:31:07 to stand out, right?
0:31:10 I just listened back to that episode when it went live.
0:31:13 And Greg was saying things like, you know,
0:31:15 people choose sides, right?
0:31:17 I’m an Apple guy, I’m a PC guy.
0:31:21 I’m a Nike guy, Adidas guy, Puma guy, whatever, right?
0:31:25 Like people sort of gravitate towards brands
0:31:27 and attach their identity to brands.
0:31:29 And a lot of times they’re the brands
0:31:32 that have good community connected to them, right?
0:31:35 So that is in my opinion, and Greg’s opinion,
0:31:39 and it’s my opinion because I heard Greg give me this opinion,
0:31:43 but you know, that’s how you stand out,
0:31:44 is through brand, through community,
0:31:47 through those things that differentiate you, right?
0:31:51 People don’t necessarily want to work with Google
0:31:53 where they’re just like a number
0:31:56 and a giant database of a million people.
0:32:00 But I use, for instance, Beehive for my AI newsletter, right?
0:32:03 And with Beehive, I’ve talked to the founders.
0:32:04 They actually respond to people on X.
0:32:07 They’re still a smaller, nimble company
0:32:08 with community around them,
0:32:12 where everybody that uses Beehive sort of loves talking
0:32:15 about it and recommends other people to go use Beehive.
0:32:18 And they’ve got that community element, that brand element.
0:32:20 And there’s a narrative that sort of farmed
0:32:24 of this like convert kit versus Beehive narrative, right?
0:32:27 And so you’ve got that kind of thing going on
0:32:29 where people might like pick their sides.
0:32:31 And the reason they’ll pick a side
0:32:34 is they connect with the community.
0:32:37 They connect with the brand that that company built.
0:32:39 And so I think, you know, as Greg said,
0:32:40 that’s how they stand out.
0:32:43 And, you know, I totally agree with that opinion.
0:32:44 – Yeah, yeah.
0:32:46 And I think, you know, AI will only like make that
0:32:47 even more so, right?
0:32:49 Like where people want more human connection there, right?
0:32:51 Like, yeah, I know the founders are like,
0:32:53 “Oh, the people in the community, I love them.”
0:32:55 And I feel like they actually like listen to what I say.
0:32:58 And this is like one reason to do like Q&A like this, right?
0:33:00 Like is to actually like, yeah, we’re doing the show,
0:33:02 but also, you know, it’s four other people
0:33:04 and we’re, you know, interacting with those people
0:33:06 versus just, oh, it’s just what me and Matt say.
0:33:08 Like it starts in general beyond community.
0:33:13 I mean, I think, you know, AI is going to make,
0:33:15 there can be so many opportunities for startups.
0:33:16 And I just, I don’t believe that the date,
0:33:18 like three companies are going to control everything.
0:33:21 And there’s so many, there’s so many problems
0:33:23 to be solved in the world.
0:33:25 And it’s not going to be two or three companies
0:33:26 that solve all of them.
0:33:28 Like there will always be new opportunities.
0:33:30 – I think if it does become three companies
0:33:32 trying to do it all, governments will intervene
0:33:35 and be like, “No, you can’t have this much power.”
0:33:37 I just, I think that’s where it’ll go.
0:33:38 – Yeah, yeah.
0:33:40 Okay, this one comes from my Twitter feed,
0:33:42 comes from Jason Vanish.
0:33:44 Hope I’m saying that correctly.
0:33:46 Jason’s always been really awesome on my Twitter
0:33:50 and doing a lot of good comments on my Twitter.
0:33:53 How much better can LLMs get?
0:33:55 At some point, there are two larger
0:33:57 and larger training datasets, are there?
0:34:01 At this point then, do we hit an AI plateau?
0:34:04 Are there different methods that could leapfrog LLMs?
0:34:06 Like the computer industry had computing breakthroughs
0:34:08 from the 70s to the 90s?
0:34:09 I mean, I think what this would like,
0:34:11 we’re definitely like speculating
0:34:13 ’cause like either one of us are like,
0:34:15 the smartest AI engineers in the world
0:34:16 or something like, I code.
0:34:17 – Yeah, I mean, you can just end in that sentence
0:34:19 as neither of us are really the smartest.
0:34:22 (laughing)
0:34:23 – Well, yeah.
0:34:24 There was an interview though,
0:34:26 I think with like Sam Altman and Illya,
0:34:28 maybe, I don’t know, maybe it was like five months ago,
0:34:29 six months ago, something like that,
0:34:34 where they said that with existing data in the world,
0:34:37 with video and audio and other things
0:34:40 that we’re just now starting to train on,
0:34:43 that there’s a pretty clear path to major improvements
0:34:47 for the next three to five years, right?
0:34:47 And so that’s the case.
0:34:51 We’re probably talking about like GPT-6, GPT-7
0:34:53 before you need any kind of breakthroughs.
0:34:54 – Yeah.
0:34:56 – And if GPT-5 is as good as they’re saying,
0:35:00 I think by GPT-6, 7, we’re talking about like work
0:35:02 looking very, very different.
0:35:03 – Right.
0:35:05 – So I think the next three to five years,
0:35:08 there’s already enough data for the world
0:35:09 to be entirely transformed,
0:35:12 mostly in positive ways.
0:35:14 And then beyond that,
0:35:17 I’m sure we will need more breakthroughs at some point.
0:35:18 And there’s a big open question too,
0:35:20 that a lot of engineers are debating
0:35:22 is like with synthetic data, right?
0:35:24 Like, is synthetic data actually useful?
0:35:26 Like data that the AI helps generate.
0:35:29 You didn’t train on that, that new data.
0:35:31 You know, it’s synthetic data.
0:35:33 And I think it’s not entirely clear yet,
0:35:35 but I could be wrong on that.
0:35:36 – Yeah.
0:35:37 – But if synthetic data works,
0:35:39 well, then yeah, there’s,
0:35:42 we’ll probably never run out of data to be training on.
0:35:43 – Yeah.
0:35:44 And I mean, everything we’re talking about at this point
0:35:46 is very sort of theoretical, but-
0:35:47 – Yeah.
0:35:48 – You know, at some point,
0:35:50 I don’t think humans need to continue
0:35:53 to improve on training the large language models.
0:35:56 Like once we hit a point of, you know, AGI,
0:35:59 we get to a point where the models
0:36:01 figure out what they need to train on next
0:36:03 and figure out how to self-improve
0:36:05 and sort of get better and better.
0:36:08 I think there is going to be sort of a model
0:36:10 that’s kind of like the last model
0:36:11 that humans needed to train.
0:36:13 And then the models beyond that
0:36:15 become the models where the AI
0:36:17 is sort of a self-improving model,
0:36:19 where it just, it gets better and better
0:36:20 and better on its own, right?
0:36:22 It does its own,
0:36:25 it’s reinforcement learning through AI feedback
0:36:26 instead of reinforcement learning
0:36:29 through human feedback at some point, right?
0:36:31 You know, to some degree, we have that right now, right?
0:36:33 Like that’s kind of how GANs,
0:36:35 Generative Adversarial Networks work,
0:36:37 where you have an AI that’s a discriminator
0:36:39 and then you have the AI that’s the generator
0:36:41 and the AI that’s the generator
0:36:43 tries to generate something that fools the AI
0:36:46 that’s the discriminator and they go back and forth
0:36:48 until the discriminator is actually fooled
0:36:50 by what the generator made.
0:36:52 I think we’re going to see that kind of thing
0:36:54 get more and more prominent with large language models
0:36:56 where it gets to a point where
0:36:57 it’s giving itself its own feedback
0:36:59 and getting better and better and better.
0:37:01 But then I also think there’s a phase
0:37:03 after large language models where
0:37:06 we start talking more about the embodied AI, right?
0:37:09 Like we were looking at the Boston Dynamics robot.
0:37:11 We’ve talked about the Figure One robot.
0:37:12 We’ve got Tesla Optimus.
0:37:15 All of these are robots that, you know,
0:37:17 they’re gonna have AI injected into them.
0:37:21 And when we start putting AI into some of these robots,
0:37:24 some of these machines, well, now they have
0:37:25 more stuff they need to learn on.
0:37:27 They need to learn on how to interact
0:37:28 with the physical world.
0:37:31 And when my arm moves like this, what’s the result?
0:37:32 When my arm moves like this?
0:37:34 And now it’s starting to get trained more
0:37:38 on the domain knowledge of just how to operate this robot.
0:37:41 So I think there’s gonna be this shift
0:37:45 to now we need to train these LLMs
0:37:48 to work with the specific use case like embodied robots.
0:37:51 Like, you know, going into drones
0:37:54 or whatever we use it on next.
0:37:56 And they need to sort of train on the domain knowledge
0:38:00 to operate the vehicle that’s now embodying that AI,
0:38:01 if that makes sense.
0:38:02 – Yeah, totally.
0:38:04 I think a big open question too is like,
0:38:07 can AI start solving real world problems too?
0:38:09 Like can it help cure cancer?
0:38:12 Can it help us figure out how to make the robots better?
0:38:14 Right, ’cause if that’s the case,
0:38:16 then we’ll probably get to this kind of exponential point
0:38:18 where things will just keep getting improving.
0:38:21 And then yeah, something that seemed like such a small
0:38:24 little breakthrough, you know, with GPT-1
0:38:27 and with transformers, you know, concept of transformers.
0:38:29 Yeah, that takes us, it could take us all the way.
0:38:30 I think that’s an open question.
0:38:32 We don’t know.
0:38:35 It could be that like just a ton of data is all we needed.
0:38:37 – A ton of data and a ton of GPUs to process it.
0:38:38 – Yeah, yeah.
0:38:40 Also, I think, you know, LLMs are gonna be just like
0:38:42 one part of the equation too, like you said, the robots.
0:38:45 But also my understanding is that, you know,
0:38:47 the rumors where the open AI has created this thing
0:38:50 called Q*, which is supposed to be some kind of logic engine.
0:38:52 I don’t think there’s any details been revealed about that.
0:38:54 Is that just like some kind of other LLM?
0:38:56 I don’t know.
0:38:59 But so in theory, you would have like the LLM
0:39:01 then having some kind of like a logic engine
0:39:02 attached on top of it.
0:39:05 So in the future, these systems could be really complicated
0:39:06 for like regular people.
0:39:08 You wouldn’t know when that’s going on.
0:39:09 – Well, if you ask Yon Lacoon, right?
0:39:12 He’s the lead AI over at Meta.
0:39:13 Also worked with Jeffrey Hinton,
0:39:16 one of the Godfathers of AI as they call them, right?
0:39:18 He doesn’t believe large language models
0:39:19 will ever achieve AGI.
0:39:22 He thinks that the sort of technology underneath
0:39:24 large language models just will never get
0:39:26 to that point of AGI.
0:39:28 And over at Meta, they’re developing something
0:39:33 they called VGEPA, which stands for Video Joint
0:39:35 Embedding Predictive Architecture.
0:39:39 But basically it’s a way for like the AI models
0:39:41 to see and understand the world
0:39:43 and sort of train themselves by actually doing
0:39:45 and getting a response.
0:39:48 And he believes this is what will actually lead to AGI
0:39:50 and not necessarily the large language models
0:39:52 that everybody’s familiar with today.
0:39:54 So I don’t know.
0:39:56 There might be a point where large language models,
0:39:59 there is like a point of diminishing returns
0:40:01 where they just don’t get any better.
0:40:03 But I still think there’s a lot of models
0:40:05 that can replace large language models
0:40:09 that continue to improve what the capabilities are.
0:40:10 – Yeah, I don’t know.
0:40:12 He like, I know he’s highly respected
0:40:14 and obviously he knows more about this technology
0:40:15 than I do, but he’s also made some predictions
0:40:18 with like towards like GPT-4 and things like that
0:40:19 that didn’t seem that accurate.
0:40:22 Like he was really skeptical on how good GPT-4
0:40:24 was gonna be and stuff like that.
0:40:27 And now like with the rumors about GPT-5,
0:40:28 if they turn out to be true,
0:40:30 well then obviously Open AI is doing something
0:40:31 that he doesn’t understand.
0:40:32 – Yeah.
0:40:34 – Right, and so we don’t know what that is.
0:40:37 Like in this Sam Almond bullshit, I don’t think so.
0:40:39 Like, you know, and the other day Sam Almond said,
0:40:41 they’re like, you know, people who are ignoring like
0:40:43 where things are gonna be going with GPT-5,
0:40:44 I think he said something like,
0:40:46 you’re gonna, we’re gonna steamroll you.
0:40:47 – Yeah.
0:40:48 – It’s what he said.
0:40:49 We’re just kind of shocking.
0:40:51 He was like, and he delivered this kind of like
0:40:52 really like, you know, peaceful, nice way.
0:40:54 But it was like, yeah, we’re gonna steamroll you.
0:40:56 And so I think they have something.
0:40:59 I think it’s gonna be shocking, but who knows?
0:41:02 Who knows what’s gonna be the best way
0:41:03 to do this in the future?
0:41:06 – Yeah, I actually think that government
0:41:09 is going to be a huge bottleneck over time, right?
0:41:11 I think, I don’t wanna get political with this,
0:41:13 but I do think that the government
0:41:16 is going to be a bottleneck to progress.
0:41:18 And the reason I say that is because I was recently
0:41:22 listening to a conversation with Ray Kurzweil
0:41:25 and Jeffrey Hinton, right?
0:41:28 And they were talking about these AI models
0:41:31 being essentially like the best correlation machines
0:41:32 on the planet.
0:41:37 They can find correlations between seemingly unrelated things
0:41:39 that humans could never spot.
0:41:41 And that’s the reason why these large language models
0:41:45 will likely find the cures to various cancers and diseases
0:41:48 and possibly solve climate change and world hunger
0:41:49 and all of this stuff, right?
0:41:50 They’ll find correlations.
0:41:52 – But they’re also very good at finding corruption.
0:41:53 – That’s true.
0:41:56 – Which, I had a viral tweet on this like maybe a year ago
0:41:57 where I talked about that.
0:41:59 And I think it’s gonna be a big thing
0:42:00 where you can actually see how,
0:42:02 where all the money is moving in the government
0:42:04 and the humans cannot process this.
0:42:07 And it’s a very easy way to like be corrupt and hide that
0:42:08 is like through moving money around
0:42:10 in like really sneaky ways.
0:42:13 – Which is another thing blockchain sort of solves as well.
0:42:13 But you know, we don’t do that.
0:42:15 – Yeah, but AI will be able to like look at that and go like,
0:42:17 oh yeah, this person’s doing this.
0:42:18 This is why they’re doing it.
0:42:20 This is why they, you know, sign this into law.
0:42:22 And so yeah, I could definitely see government people
0:42:25 being very like not wanting that to come out.
0:42:27 – Yeah. Well, you know, where I was going
0:42:30 with them, the government being the bottleneck is,
0:42:32 I think these large language models
0:42:35 will likely find cures for cancer,
0:42:39 you know, solutions for climate change, you know,
0:42:41 ways to end hunger in, you know,
0:42:43 parts of the world that need it.
0:42:46 But I think government is gonna get in the way
0:42:50 of making what the AI finds a reality, right?
0:42:52 Like when there’s new drugs that come to market,
0:42:55 how long does it take for the drug to go through
0:42:57 animal trials and then human trials?
0:43:00 And then, you know, all of the steps
0:43:01 before the drug finally gets on the market
0:43:03 for humans to use, right?
0:43:07 I think AI will probably find a lot of solutions
0:43:08 to a lot of problems.
0:43:11 And then the government’s red tape is going to be
0:43:13 what slows down these solutions
0:43:16 actually becoming live to the world.
0:43:18 – Yeah. I mean, I mean, that’s why there’s movements
0:43:19 like EAC and things like that, right?
0:43:22 It’s, they are really afraid of that.
0:43:24 It’s like, I think people talked about like, you know,
0:43:26 we could have solved a lot of our energy problems
0:43:28 with like nuclear power, but then regulation
0:43:31 like stopped that from happening, right?
0:43:33 And I think France is one of the big success stories
0:43:35 where they adopted nuclear power
0:43:37 and they haven’t had energy problems
0:43:38 like a lot of other countries have.
0:43:41 And so there’s a lot of fear around that.
0:43:43 And, but then the challenge is like, you know,
0:43:46 EAC does kind of go so extreme that they might, you know,
0:43:47 bringing on some of the regulation, right?
0:43:49 ‘Cause like just let’s go as hard as we can
0:43:51 as bad as we can.
0:43:53 And then the government kind of freaks out about that too.
0:43:56 So it’s really hard to know like how to get the government
0:43:58 to embrace this technology.
0:43:59 I’m hoping they will.
0:44:02 I mean, I think it’s like the U.S. winning at the internet.
0:44:05 Like the U.S. needs to win an AI too
0:44:06 in order to kind of set the stage
0:44:09 for the next 100 years of the world, right?
0:44:12 And if we don’t, whoever wins that, China or whoever,
0:44:13 you know, Russia, whoever,
0:44:16 they then get to set the stage for the next, you know,
0:44:18 chapter of humanity.
0:44:20 – It’s the new space race.
0:44:22 – Yeah. I think it’s bigger than that.
0:44:22 – Yeah.
0:44:23 – Yeah, I mean, it is really interesting
0:44:25 to think about it from that perspective too,
0:44:28 because like the U.S. probably wants to be the country
0:44:29 that finds the cure for cancer,
0:44:32 but is the government going to, you know,
0:44:34 is the government gonna put up a bunch of red tape
0:44:35 and slow that down?
0:44:36 I don’t know yet to be seen.
0:44:37 – Yeah.
0:44:41 – I think there’s going to be a sort of necessary overhaul
0:44:46 coming because with this AI world that we’re entering into,
0:44:48 I do feel like we need stuff to happen faster
0:44:52 so that we can let the solutions that these tools provide,
0:44:54 provide the solutions.
0:44:55 – I think ideally that’s what we do.
0:44:57 Like as these systems become incredibly intelligent,
0:45:00 more, they can process more data than humans can,
0:45:02 we should be asking them like, okay,
0:45:04 how can we reorganize this part of the government
0:45:06 to be more efficient and actually get more things done
0:45:09 and help more people, bring more people out of poverty
0:45:10 and things like this?
0:45:11 That’s what I hope for.
0:45:13 You know, that’s the one reason I wanted to do it,
0:45:14 you know, do a show with you.
0:45:15 – Yeah.
0:45:17 – That’s why I tweet is I’m really hopeful
0:45:19 for what this technology can do for humanity
0:45:22 if we don’t all get in the way, right?
0:45:24 – You know, I think that’s a perfect spot
0:45:25 to wrap this one up.
0:45:28 If anybody is watching on YouTube,
0:45:29 let us know in the comments
0:45:31 if you like this style of episode,
0:45:33 if you want us to do more Q and A
0:45:36 and you like just the sort of us giving thoughts
0:45:37 on your questions,
0:45:38 ’cause if you like this,
0:45:40 we’ll make more episodes like this for you.
0:45:41 If you haven’t already,
0:45:44 make sure you like and subscribe
0:45:46 wherever you’re watching or listening to this.
0:45:47 It really helps us out
0:45:50 to get more listeners and viewers on the show.
0:45:51 So appreciate that.
0:45:53 And thank you so much for tuning in today.
0:45:55 We’ll see you in the next episode.
0:45:57 (upbeat music)
0:46:00 (upbeat music)
0:46:02 (upbeat music)
0:46:05 (upbeat music)
0:46:08 (upbeat music)
0:46:10 (gentle music)
0:46:20 [BLANK_AUDIO]