AI transcript
0:00:01 (upbeat music)
0:00:03 – I’m still very bullish on Google
0:00:05 because I think it’s like the tip of the iceberg
0:00:07 of what you see in tech companies
0:00:09 and the submerged part is just amazing.
0:00:10 They’ve got the most complete models
0:00:12 of the physical and digital world.
0:00:15 They’ve got ubiquitous distribution
0:00:17 and an existing ecosystem of ads and sales
0:00:19 to plug monetization in.
0:00:21 So I think it’s still a magical combination.
0:00:27 – Hey, welcome to the Next Wave podcast.
0:00:28 I’m Matt Wolf.
0:00:30 I’m here with my co-host, Nathan Lans.
0:00:33 And with this podcast, it is our goal
0:00:36 to keep you looped in with all of the latest AI news,
0:00:38 the latest AI tools,
0:00:41 and just help you keep your finger on the pulse
0:00:43 so that you are prepared for the next wave of AI
0:00:44 that’s coming.
0:00:47 And today we have an amazing guest on the show.
0:00:50 We have Bilival Sidhu on the show.
0:00:52 He is the host of the TED AI podcast.
0:00:54 He’s an ex-Googler.
0:00:56 We’re gonna talk to him about what it was like
0:00:59 working on AI and visual effects over at Google.
0:01:02 We’re gonna talk about the difference
0:01:04 between whether or not we should be accelerating AI
0:01:06 or slowing down AI.
0:01:08 We’re also gonna learn about how some
0:01:11 of these AI visual effects tools work
0:01:13 because this is the field that Bilival worked in
0:01:14 for so long.
0:01:15 It’s an amazing episode.
0:01:17 You’re gonna learn a ton
0:01:18 and can’t wait to share it with you.
0:01:20 So let’s jump on in with Bilival Sidhu.
0:01:24 Thanks so much for being on today, Bilival.
0:01:25 – Thanks for having me, gentlemen.
0:01:26 Pleasure to be here.
0:01:29 – Yeah, so I wanna just kind of dive right into it.
0:01:32 Your background is Google, right?
0:01:35 So I think when you and I first connected
0:01:38 and we first started having some chats over Twitter DM,
0:01:41 you were still actually working over at Google at the time
0:01:43 and you also were kind of doing a creator business
0:01:45 on the side with your YouTube channel
0:01:46 and everything you had going on.
0:01:48 But Bilival Sidhu, what were you doing over at Google?
0:01:49 What was your role there?
0:01:52 You know, what was your experience like over there?
0:01:53 – Gosh, yeah, it was awesome.
0:01:55 So, you know, I’ve spent a decade in tech,
0:01:57 six years at Google.
0:01:58 I’ve been able to work on projects
0:02:00 that blend the physical and digital world.
0:02:03 And I started off in the AR/VR team really
0:02:06 when spatial computing, as it’s now called,
0:02:07 was first popping off.
0:02:10 This is like right after the DK2 came out,
0:02:11 Google Glass was a thing.
0:02:15 And everyone was talking about what is the next iteration
0:02:16 of computing platforms?
0:02:19 Where are we gonna go from like this mobile revolution?
0:02:22 And so I had a chance to work on a bunch of cool stuff there,
0:02:26 YouTube VR content, livestreaming Coachella,
0:02:28 Dean Choice Awards, Elton John,
0:02:30 the camera systems that we used
0:02:33 to do stereoscopic 3D capture, AR SDKs,
0:02:35 when that became popular, augmented reality,
0:02:37 sort of hit the scene.
0:02:40 And then after that, I spent four years at Google Maps,
0:02:43 basically creating a ground up 3D model of the world,
0:02:45 sort of remapping the world, if you will,
0:02:48 and then turning the world into an AR canvas
0:02:51 with the AR Core Geospatial API.
0:02:53 It’s been a lot of fun and yeah,
0:02:55 it’s been awesome to work with some really talented folks
0:02:59 to work on these projects that have been blurring this line
0:03:01 between the world of bits and atoms.
0:03:03 – So I’m curious, working on all this stuff,
0:03:05 in my mind, I can’t even imagine
0:03:07 what a day-to-day looks like at Google.
0:03:09 I’ve been on the Google campus
0:03:13 and it looks like a giant playground for tech nerds.
0:03:14 So I’m just kind of fascinated
0:03:17 by what it’s like to work at Google.
0:03:19 Like what does a day-to-day look like over there?
0:03:20 – You know, so I was a product manager
0:03:23 and so a day-to-day for me is gonna be very different
0:03:26 than if you go talk to like an engineer or designer.
0:03:27 For me, really it was a lot of meetings.
0:03:30 Let me be perfectly honest, just like a ton of time.
0:03:31 But there’s some very cool things.
0:03:34 I think like Google and big tech companies generally
0:03:37 are sort of this like interesting microcosm
0:03:39 where it’s like, you know, I’ll send out an email
0:03:43 and the guy that wrote the book on computer vision,
0:03:46 like the computer vision book that like everyone reads,
0:03:48 like a response to it and I get a bunch of pings
0:03:51 and you’re like, oh, so-and-so respond to your thing.
0:03:53 And it’s like all these Pokemon that like, you know,
0:03:55 these companies have caught that are available
0:03:58 at your back, back in call to share ideas with,
0:04:01 pull into your own projects and really just like,
0:04:05 you know, there’s such a, it’s like the tip of the iceberg
0:04:07 of what you see in tech companies
0:04:09 and the submerged part is just amazing.
0:04:11 So it’s like, when I moved over to the maps team,
0:04:14 I was thinking of working on glasses at the time.
0:04:16 And the reason I went to maps is like,
0:04:18 I met this engineer is like,
0:04:21 oh yeah, we write CLs to move satellites around in the sky.
0:04:23 And I was like, wait, what, huh?
0:04:25 You move satellites around the sky.
0:04:27 And it’s like, yeah, like literally orchestrating a fleet.
0:04:29 Like, you know, like most people don’t know this,
0:04:31 like Google owns their own fleet of like,
0:04:33 not just street view cars, but airplanes.
0:04:34 Oh, wow.
0:04:36 And so like the ability to like task those.
0:04:38 So like, hey, we need to, you know,
0:04:41 we’ve got this like Sundar IO thing coming up
0:04:44 and we’re gonna be presenting immersive view.
0:04:47 We got to go capture this high resolution model of London.
0:04:48 And like, not, you know,
0:04:51 and suddenly things in the world of atoms
0:04:52 are moving to make that happen.
0:04:54 I think it’s like absolutely amazing.
0:04:56 I think people deeply underestimate the data mode
0:04:58 that Google has.
0:05:00 Obviously the most complete digital twin of the world
0:05:02 we’re talking about, but like search, right?
0:05:05 Like YouTube, oh my goodness.
0:05:07 All this stuff is available.
0:05:09 And so like, there’s cool things and products
0:05:11 you can build around it.
0:05:13 But along with it comes, you know,
0:05:15 which may not be a surprise to people in tech,
0:05:18 but like a ton of responsibility and sort of guardrails
0:05:20 for how you actually use this data.
0:05:21 So it’s like the size of the prize
0:05:24 and the data sets you get to play with are amazing.
0:05:25 But to be able to do stuff with it,
0:05:28 you really, really have to be exceedingly thoughtful.
0:05:30 And there’s a lot of process involved
0:05:32 in unlocking that innovation.
0:05:33 So yeah, that’s how I would describe it.
0:05:35 I think like it’s just like a,
0:05:38 it’s like, it’s like a Disneyland for nerds to be honest.
0:05:43 When all your marketing team does is put out fires,
0:05:45 they burn out fast.
0:05:46 Sifting through leads,
0:05:48 creating content for infinite channels,
0:05:51 endlessly searching for disparate performance KPIs.
0:05:52 It all takes a toll.
0:05:56 But with HubSpot, you can stop team burnout in its tracks.
0:05:58 Plus your team can achieve their best results
0:06:00 without breaking a sweat.
0:06:02 With HubSpot’s collection of AI tools,
0:06:05 Breeze, you can pinpoint the best leads possible.
0:06:08 Capture prospects attention with clickworthy content
0:06:11 and access all your company’s data in one place.
0:06:14 No sifting through tabs necessary.
0:06:16 It’s all waiting for your team in HubSpot.
0:06:17 Keep your marketers cool
0:06:20 and make your campaign results hotter than ever.
0:06:23 Visit hubspot.com/marketers to learn more.
0:06:28 – So it sounds like you’re still really bullish on Google.
0:06:30 Cause I know we were having like some playful banter
0:06:32 last year about like, you know, I was like,
0:06:34 maybe Google’s going to die.
0:06:36 And you were like, what the hell are you talking about?
0:06:38 – Yeah, yeah, we’ve actually had that conversation
0:06:40 on this podcast before where I’m like, you know what?
0:06:42 I think I give more credit to Google.
0:06:43 I think Nathan’s a little more,
0:06:46 I don’t know if they’re going to be like the top dogs in AI.
0:06:48 Like where do you stand on that?
0:06:49 Do you count Google out?
0:06:52 Do you think Google will pass the Microsoft AI,
0:06:54 you know, Avenger mega team?
0:06:56 – It’s hard to say anything with certainty,
0:06:59 but what I will say is like, you know, after I left Google,
0:07:00 I think I was like one of the few people,
0:07:03 I felt like I was in isolation saying good things
0:07:04 about Google.
0:07:06 Everyone was just like, oh yeah, they’re just too slow.
0:07:09 They foambolt, they came up with a transformer.
0:07:11 All the transformer folks are left.
0:07:15 I think it’s a situation where there was no real disruption
0:07:16 in site for the search base.
0:07:18 Yes, there were talks about like, hey, like kids
0:07:21 are like searching on TikTok and like YouTube like now,
0:07:24 but YouTube is owned by Google on TikTok.
0:07:26 It’s kind of short former people are really going to be like,
0:07:27 is that going to be a resilient thing?
0:07:30 And one might argue now social networks are places
0:07:31 where people do a lot of searching,
0:07:34 but traditional search sort of as like, you know,
0:07:37 a maps guide is to just give the maps analogy.
0:07:40 It’s like, you know, like maps is how you discover stuff
0:07:41 in the real world.
0:07:43 And Google is how you discover stuff in the digital world.
0:07:46 It’s literally your window to the worldwide web, right?
0:07:48 I don’t think anything had like sort of questioned
0:07:51 the strong position Google was in in that regard
0:07:55 until chatGPT came out when suddenly people
0:07:57 could start connecting the dots and see you like,
0:08:00 if you connect large language models with like, you know,
0:08:03 knowledge graph and search index, kind of like perplexity.
0:08:06 And you know, Microsoft co-pilots
0:08:09 and whatever the heck else OpenAI is going to announce
0:08:13 in order to kneecap any Google limelight next week.
0:08:14 Like I think people start saying,
0:08:17 hey, there’s a disruption inside.
0:08:19 And I think combined with the fact that like,
0:08:21 the search ads business model
0:08:24 was just such a money printing machine and still is.
0:08:26 And the fact that, you know, the cost per query
0:08:29 of these generative AI models is obviously going to be higher.
0:08:32 And how do you do advertising and attribution
0:08:34 and all this stuff like that it would like,
0:08:37 kind of represents a, you know, contraction
0:08:40 in the money printing machine and the pie
0:08:41 and the business that Google created
0:08:44 had all the signs of sort of innovators dilemma.
0:08:46 And I think like Google has sort of adopted
0:08:48 the playbook of the innovator solution.
0:08:51 And, you know, initially they had some of these reorgs
0:08:54 that felt more like exec reorgs.
0:08:56 And now they’re actually bringing together
0:08:58 like the brain and the deep mind teams
0:09:01 and they’re actually shipping at a really good cadence.
0:09:04 And I think they still have some of the most unique data sets
0:09:06 that other folks are talking about, you know
0:09:07 that may or may not have been scraped, you know
0:09:11 he’s case in point, the CTO of open AI
0:09:13 being asked by Joanna Stern about,
0:09:14 what does it exactly mean
0:09:16 that you train on publicly available data?
0:09:19 So all this to say, I’m still very bullish on Google
0:09:21 because I think they’ve got the most complete models
0:09:23 of the physical and digital world.
0:09:26 They’ve got ubiquitous distribution
0:09:28 and they’ve got the right infrastructure chops
0:09:31 to basically like bring that cost per query down
0:09:34 and an existing ecosystem of ads and sales
0:09:35 to plug monetization in.
0:09:39 So I think that there’s a like monetizable
0:09:41 sort of like answer engine model.
0:09:43 I think Google is one of the few companies
0:09:44 that could crack it.
0:09:46 That isn’t to say that I think like open AI
0:09:48 and Microsoft can’t take meaningful market share
0:09:50 but let’s be honest, how many of us actually use Bing?
0:09:51 Like I don’t, right?
0:09:53 Like I used it for a little bit
0:09:55 and I probably use perplexity more now.
0:09:58 Yeah, I mean, I am using chat GBT and perplexity
0:09:59 instead of Google a lot these days.
0:10:01 I agree, me too, to be honest, yeah.
0:10:03 And then, you know, I’ve been following,
0:10:04 you know, a long time ago I used to do SEO
0:10:06 like a long, long time ago
0:10:07 and I’ve been kind of following that space
0:10:09 and like in the last two months
0:10:11 and I kind of predicted this a year ago,
0:10:14 they’re like having major changes to the algorithm
0:10:16 where they’re really focusing on authority.
0:10:18 Yeah, domain authority, yeah.
0:10:20 Yeah, the reason they’re having to do that
0:10:23 is because of all the flood of AI content, right?
0:10:26 Like they just can’t deal with the flood of AI content.
0:10:28 So it’s like, okay, how do you deal with that?
0:10:31 Let’s go back to like really valuing the big brands
0:10:34 and the big names or the famous people too.
0:10:36 That’s the other thing is maybe like they’re taking
0:10:37 social signals, like you have a lot of followers
0:10:38 on social media.
0:10:40 Now that’s a signal that you’re a, you know,
0:10:42 an author they should listen to.
0:10:44 Yeah, I mean, why do you think we signed on with HubSpot?
0:10:47 We want that backlink domain authority.
0:10:48 That’s the only reason.
0:10:48 Yeah, yeah, yeah.
0:10:50 (laughing)
0:10:54 I mean, this is like, this is gonna be a meta problem though
0:10:55 for the industry, right?
0:10:57 The just the explosion of synthetic content.
0:10:59 I mean, and some social networks
0:11:01 are almost like incentivizing it.
0:11:03 Like, like on LinkedIn, it gives you,
0:11:06 it uses GPT-4 to suggest the comments.
0:11:08 And now you have just the most cringe like,
0:11:11 like regurgitations and summarizations
0:11:12 of like the original posts from like,
0:11:15 clearly a normal human being would never write this.
0:11:18 But like, if there was a meme about how to respond
0:11:20 to somebody on LinkedIn, I mean,
0:11:22 that’s like encapsulates the style I see.
0:11:22 And so.
0:11:23 But that’s happening on Quora too.
0:11:24 So like right now,
0:11:27 Quora is starting to rank at the top of Google results
0:11:29 and Quora is being dominated
0:11:31 with a chat GPT responses now.
0:11:33 100%, I mean, like books on Amazon too.
0:11:34 And I think the way they came out with a restriction
0:11:37 as well, you can only upload X number of books per day.
0:11:40 Like, I don’t know if that’s the solution.
0:11:42 But like, this is like the deep fake,
0:11:43 shallow fake problem too.
0:11:45 It’s like, everyone’s talking about detecting deep fakes.
0:11:46 And like, how do we figure this out?
0:11:50 It’s like, well, the thing that causes maximum harm today
0:11:51 are aren’t actually deep fakes.
0:11:53 They’re like super shallow fakes
0:11:55 where you take a photo from a different time or a context
0:11:58 and like, you know, kind of put it against another context.
0:11:59 This is exactly the type of stuff.
0:12:01 You’ll see like community noted on Twitter.
0:12:04 And that stuff’s like relatively easier to detect
0:12:06 ’cause you can actually find the source imagery
0:12:08 if you do like reverse image search.
0:12:10 And so like you add in the generative problem
0:12:12 on top of that, it’s like even crazier,
0:12:14 but like most platforms haven’t really even solved
0:12:16 the shallow fake problem, right?
0:12:18 It’s really like when it reaches a certain threshold
0:12:21 of distribution, there’s sort of this retroactive,
0:12:23 let’s go throttle this thing versus like,
0:12:25 how do you get ahead of this?
0:12:26 Anyway, I could talk about that forever
0:12:27 because some of the ways to avoid that is like,
0:12:30 ubiquitous surveillance, which is also not, you know.
0:12:32 Oh yeah, sounds great.
0:12:36 It’s like somehow the solutions to things that sound 1984
0:12:38 is like 1984 technology.
0:12:40 It’s like, it’s kind of weird how that works.
0:12:42 But it’s funny, somebody shared an article with me
0:12:45 just like the other day that was an article about,
0:12:47 well, not just me, but it was like an article
0:12:51 about like these seven AI influencers are, you know,
0:12:54 changing how we see AI or something like that, right?
0:12:56 And then like one of the seven like was my name
0:12:58 and I read the blurb about myself.
0:13:00 And the blurb about myself was like,
0:13:03 I grew up in Louisiana doing real estate and-
0:13:05 – You’re my neighbor.
0:13:08 – And like transitioned into computer programming
0:13:09 and then started teaching AI.
0:13:11 And I’m like, other than the fact that they’re like,
0:13:15 Matt makes content about AI, everything else about that
0:13:17 was just completely wrong.
0:13:18 – I think that’s where also like these models
0:13:21 need to be anchored in some sort of real knowledge graph,
0:13:24 you know, and like, that’s not to say that like, you know,
0:13:27 an approach like search is only gonna give you the truth,
0:13:28 right?
0:13:29 Like there’s like, what is even the truth?
0:13:31 And there’s like differing opinions on it.
0:13:33 But I think these models to be able to like,
0:13:34 just kind of fact check themselves,
0:13:36 at least with known information and come up with,
0:13:39 oh, at least the resources that are reputable are saying,
0:13:43 this is Matt’s bio would be better than, I don’t know.
0:13:46 So I’m like, this is the equivalent of like SEO dribble
0:13:50 that sort of started bleeding into Google around like 2019.
0:13:51 I think like, you know, it’s,
0:13:53 this is gonna be a huge problem.
0:13:55 And not to mention the implications of like,
0:13:57 if we’ve run out of content on the internet
0:14:01 and we are actively disincentivizing human generated content,
0:14:03 like how are we gonna train these models?
0:14:04 Like what’s gonna happen there?
0:14:05 – Yeah.
0:14:08 We’ll be right back.
0:14:11 But first I wanna tell you about another great podcast
0:14:12 you’re gonna wanna listen to.
0:14:15 It’s called “Science of Scaling” hosted by Mark Roberge.
0:14:18 And it’s brought to you by the HubSpot Podcast Network,
0:14:21 the audio destination for business professionals.
0:14:23 Each week host Mark Roberge,
0:14:26 founding chief revenue officer at HubSpot,
0:14:28 senior lecturer at Harvard Business School
0:14:30 and co-founder of Stage Two Capital,
0:14:33 sits down with the most successful sales leaders in tech
0:14:36 to learn the secrets, strategies and tactics
0:14:38 to scaling your company’s growth.
0:14:40 He recently did a great episode called
0:14:44 “How do you solve for a siloed marketing and sales?”
0:14:46 And I personally learned a lot from it.
0:14:47 You’re gonna wanna check out the podcast,
0:14:51 listen to “Science of Scaling” wherever you get your podcasts.
0:14:57 – I wanna talk real quick about search
0:14:59 because just to peek behind the curtain
0:15:02 to anybody who might be listening to this episode,
0:15:05 we’re actually recording it right before Google I/O, right?
0:15:07 Like Google I/O is next week
0:15:09 from the time we’re recording this.
0:15:12 Billavall and I are actually gonna be out at Google I/O
0:15:14 attending in person.
0:15:16 But one of the things that you mentioned
0:15:20 is that OpenAI and Microsoft have a tendency
0:15:23 that whenever Google announces something,
0:15:26 they need to jump in and like sort of one up them.
0:15:28 So by the time this episode comes out,
0:15:32 we’ll probably already know what OpenAI did with search.
0:15:35 But the rumor right now is that OpenAI
0:15:38 is creating some sort of their own search engine
0:15:41 with maybe with Microsoft involved, maybe not.
0:15:44 There’s still a lot of rumors and speculation flying around.
0:15:45 But knowing what you know about Google,
0:15:49 do you think OpenAI and chat GPT can come in and compete?
0:15:52 – I think no, like I think the search index Google has
0:15:53 is a very strong mode.
0:15:56 The fact that they can sort of almost like map the internet
0:16:00 in like almost real time is like just a hard technical
0:16:01 and infrastructure problem.
0:16:04 And they’re really well set up for it.
0:16:05 I’m curious what it is.
0:16:08 Like obviously this is complete pontification rumor mail.
0:16:11 Like what it is OpenAI is going to roll out.
0:16:14 What I think it’s gonna be is like something at parity
0:16:18 with perplexity with maybe a better like search index involved.
0:16:20 Like, and if it’s even if it’s something like that
0:16:23 where you get sort of this like multimodal summary
0:16:25 where it looks at a bunch of links, you get, you know,
0:16:27 some images, maybe some embedded videos
0:16:29 and like a summary of whatever it is you asked for
0:16:32 with citation so you can go validate sort of the quality
0:16:35 of those like the links that were summarized.
0:16:37 I think that would be a huge step up, right?
0:16:40 Like just getting, just being able to invoke like,
0:16:43 like search inside of chat GPT right now is clunky.
0:16:44 You have to be like, hey, well, and look this up,
0:16:48 research this and like explicitly prompted to do that
0:16:49 and being able to do that in a fashion
0:16:54 that is like really about leaning on like sort of a real time
0:16:58 and like sort of, you know, content that has real provenance
0:17:02 along with, you know, like sort of the distilled wisdom,
0:17:03 you know, the wisdom is debatable.
0:17:05 The distilled wisdom that is in,
0:17:07 in these large language models to summarize that,
0:17:09 I think it’s still a magical combination.
0:17:11 And I don’t know if y’all feel like this,
0:17:12 but I think the vibes have been shifting
0:17:14 with regards to just the conversation
0:17:18 and sort of Matt, I know you had this like post about like,
0:17:19 we think things are slowing down,
0:17:22 but here’s a bunch of announcement events coming up,
0:17:25 but doesn’t it feel like just like those leaps
0:17:28 that we hoped for haven’t quite come?
0:17:31 I guess Sora was kind of a leap.
0:17:34 I think the trend from like a million to 10 million
0:17:37 to maybe infinite context is like an interesting leap too,
0:17:40 but maybe we’re just getting too used to it, you know,
0:17:43 like versus the technologies,
0:17:45 like the pace of advancements slowing down.
0:17:46 Curious, which y’all think?
0:17:47 Yeah, I don’t know.
0:17:49 I feel like it’s kind of like it’s on social media,
0:17:52 the perception is, you know, the vibe has shifted,
0:17:53 but I feel like a lot of people
0:17:56 who actually are closer to Sam Altman,
0:17:57 they don’t feel that way.
0:18:00 And so I still believe that, you know,
0:18:01 they have something amazing coming.
0:18:02 – I follow Gary Marcus.
0:18:04 I know I think you actually have Gary Marcus
0:18:06 coming on the TED AI podcast,
0:18:08 so we’ll probably get some more insights from him
0:18:10 in the near future, but, you know,
0:18:12 the sort of Gary Marcus thing
0:18:15 that he’s been sort of all over Twitter about
0:18:20 is that we’re not seeing that same leap from GPT-4 to GPT,
0:18:25 or the same leap from GPT-2 that we saw the GPT-3.
0:18:27 He’s sort of arguing that that exponential curve
0:18:28 that everybody’s talking about
0:18:31 that we’re on with AI is not true, right?
0:18:32 We’re not on this exponential curve.
0:18:36 Otherwise, like, why didn’t we get from GPT-4 to GPT-5?
0:18:40 And half the time we got from GPT-2 to GPT-3, right?
0:18:42 Like, why is it not showing that?
0:18:45 But I also, like, I think my counter argument to that
0:18:46 is just ’cause we’re not seeing it,
0:18:47 doesn’t mean it’s not there, right?
0:18:49 – Yeah, totally.
0:18:51 – I think there’s a lot of stuff happening,
0:18:53 like Nathan sort of alluding to
0:18:57 behind the scenes over at OpenAI that we’re not seeing.
0:19:00 – I think it’s probably more related to, like,
0:19:02 compute requirements to actually use
0:19:05 some of these newer, more advanced models.
0:19:06 If they were to release it right now
0:19:08 with, like, the compute that’s available,
0:19:10 it would be, like, really expensive.
0:19:12 And, you know, the 20 bucks a month
0:19:14 that people are paying to use Jet GPT
0:19:16 is probably not gonna cover the cost of inference to run,
0:19:19 you know, these newer models, same thing with Sora.
0:19:21 – I mean, that’s kind of the leading theory.
0:19:24 I know that I believe with the GPT-2,
0:19:25 like, the mysterious model that came out
0:19:27 is, like, what the hell is that?
0:19:29 Like, maybe that’s what it is, is, like,
0:19:30 they’ve actually, maybe this is something
0:19:32 they actually developed, like, a year or two ago.
0:19:35 And it’s like a, it’s a more efficient architecture
0:19:36 or something like this.
0:19:38 And possibly that’s what GPT-5’s built upon,
0:19:40 then they, you know, they have less issues
0:19:42 with the cost issues in theory, but, yeah.
0:19:45 – I think it’d be hilarious if the GPT-2 stuff
0:19:46 has actually opened AI.
0:19:49 I mean, like, what an interesting way to sort of
0:19:51 test a model in the wild versus, I don’t know,
0:19:54 running some sort of A/B where, like,
0:19:57 there’s some sort of experiments on the chat GPT website
0:19:59 where a subset of users get a certain model
0:20:01 versus another set of users.
0:20:03 Maybe, like, making it explicit that there’s
0:20:04 these two versions of the models
0:20:07 and having people respond to it separately is interesting.
0:20:09 Maybe there’s an intention to create a bit of a PR,
0:20:11 like, sort of, like, you know, kind of,
0:20:14 a seed the conversation and kind of, like,
0:20:16 grease the wheels before the reel,
0:20:17 like, sort of, I don’t know,
0:20:19 race car jump moment happens or whatever.
0:20:22 I don’t know, but it’s like, yeah, like,
0:20:24 I mean, clients agree, like, there’s the compute stuff.
0:20:26 Certainly, Sora hasn’t been rolled out widely
0:20:29 because, like, it’s just so compute-intensive, right?
0:20:31 Like, you need to come up with a completely different,
0:20:34 like, pricing model way beyond the $20 one.
0:20:37 Maybe, hence, them talking to studios and stuff like that.
0:20:39 But all these models will get optimized, too.
0:20:40 To the Gary Marcus point, it’s interesting,
0:20:43 I had a conversation with him last week and it was like,
0:20:46 it’s like, he’s been very consistent about, like,
0:20:48 this not being the right paradigm.
0:20:51 And, you know, I think people liked it,
0:20:53 like, for the lack of, let’s just put it blunt.
0:20:55 Like, people like to shit on Gary a lot.
0:20:57 But, you know, if there’s one thing Gary’s been,
0:20:59 it’s like, it’s been exceedingly consistent.
0:21:01 And so, I don’t know, like, I would like to see
0:21:06 this sort of agentic co-pilot that feels more like an employee.
0:21:08 We certainly haven’t seen it yet, right?
0:21:11 The converse of everything else is the expectations
0:21:13 around AI, like, a year ago,
0:21:15 if we go back to when GPT-4 came out,
0:21:19 we’re just so fricking high that people just thought,
0:21:20 like, whether you were a knowledge worker
0:21:23 or a visual creator, you would look at the narrative
0:21:24 and you’d be like, holy crap,
0:21:26 like, this thing’s gonna take my job.
0:21:27 And then I don’t know if y’all saw the tweet,
0:21:29 I was like, and then you go use the tech
0:21:32 and it feels less like this, like, fricking kaiju Godzilla
0:21:34 that’s gonna stomp you and more like this, like,
0:21:36 chaotic golden retriever that you can kind of coax
0:21:39 to do cool stuff with you.
0:21:41 And I think that delta between expectations
0:21:44 and reality is so stark and, you know,
0:21:46 like, there was all of these,
0:21:48 I think even in a recent Sam Altman interview where he was
0:21:50 asking one of his biggest regrets is like,
0:21:53 GPT-4 didn’t have that, like, economic impact.
0:21:55 Everyone thought it would.
0:21:56 And he’s worried that the pendulum
0:21:58 will swing sort of the other way now,
0:22:00 where if expectations were so high,
0:22:02 people were like, oh yeah, whatever.
0:22:03 And so, I don’t know,
0:22:04 I think the answer is always gonna be in the middle,
0:22:07 but I can’t help but feel like we’ve gone past
0:22:09 the, like, peak of inflated expectations
0:22:12 and we’re going into the trough of disillusionment.
0:22:12 – Well, we’ll see.
0:22:14 I mean, like, Sam Altman also said that, like,
0:22:17 he was surprised that GPT-4 was so successful
0:22:19 and that, you know, it kind of sucks.
0:22:21 – I don’t know, my feeling on that is that Sam Altman
0:22:23 is one of the, like, greatest marketers
0:22:24 of our time right now.
0:22:27 And he’s, you know, he’s really, really, really good
0:22:29 at getting that hype wheel spinning.
0:22:30 – Oh, baby.
0:22:32 – Yeah, I swear, I think Sam Altman
0:22:34 is just really, really smart.
0:22:36 And I think, you know, when you see him speak, right,
0:22:39 when he does interviews, he’s very calculated, right?
0:22:41 He’ll ask a question and he’ll sit there
0:22:43 and he’ll usually pause for a good few seconds
0:22:45 before he responds.
0:22:48 And I think that he’s got that marketer brain.
0:22:49 Like, what can I say?
0:22:52 It’s going to sort of spread the flames
0:22:53 to hype this up a little bit more.
0:22:56 And I think that’s kind of how his brain operates.
0:22:58 So I think, you know, him saying like,
0:23:01 this is going to be the dumbest model
0:23:02 you’ve ever used by a lot.
0:23:04 And we’re going to, you know, we’re going to look at back
0:23:06 at this and be embarrassed by what we put out.
0:23:08 I think that’s all marketing.
0:23:09 – Well, I’m certainly excited.
0:23:11 But, you know, if I could have one request right now,
0:23:15 it’s like just give me GPT-4 from April, last year.
0:23:17 Give me that vintage of GPT-4.
0:23:18 It was better.
0:23:19 It was better, damn it.
0:23:21 – Can you still go into like the open AI playground
0:23:23 and like select the older models?
0:23:24 – I think you can, yeah.
0:23:25 And I feel like just all of these models,
0:23:27 especially in the consumer interface,
0:23:29 follow this trajectory of like when they launched,
0:23:30 they’re really good.
0:23:32 And then over time as like, you know,
0:23:35 various efforts to make sure the output is, you know,
0:23:38 on rails and not harmful kick in,
0:23:40 you see this deterioration happen.
0:23:43 But hey, that’s why we also have open source, right?
0:23:43 – Right.
0:23:44 – Yeah.
0:23:46 – Well, let’s talk about visual effects too,
0:23:48 because that’s like, that’s really your background
0:23:49 over at Google.
0:23:53 I want to go back to like the sort of, you know,
0:23:57 3D imaging 101 for a second here.
0:23:59 Can you sort of like break down the difference
0:24:01 between things like photogrammetry,
0:24:05 LiDAR, NERFs and, you know, Gaussian splats?
0:24:06 – I would love to.
0:24:07 In fact, the thing I was talking about is like,
0:24:10 everyone talks about like generative AI a lot,
0:24:12 but I think like the part that’s getting
0:24:16 not that much attention is this like visual spatial AI space.
0:24:19 And so think about spatial intelligence as like,
0:24:21 really just like a reality capture.
0:24:23 Like the world is in fricking 3D, right?
0:24:25 And so, you know, Matt, you nailed it.
0:24:28 It’s like basically photogrammetry is the art and science
0:24:31 of taking 2D images and other sensor data like LiDAR
0:24:34 and turning it into these 3D representations
0:24:35 of the real world.
0:24:37 So photogrammetry has been around
0:24:39 since like before computers were invented even.
0:24:43 Like this is a way of basically using like math and images
0:24:45 and observations of the world really
0:24:48 to like extract 3D structure from it.
0:24:50 But you should also think of spatial intelligence
0:24:53 as the ability for machines to sort of interpret
0:24:55 the spatial data like maps, 3D models,
0:24:57 like the world as we see it, right?
0:25:00 And so like, to me like photogrammetry
0:25:03 or like reality capture, all these other techniques
0:25:05 are all about recreating reality.
0:25:07 And so photogrammetry isn’t new as I alluded to, right?
0:25:09 But I think what’s gotten a huge boost
0:25:10 in why you hear about all these things
0:25:12 is like, thanks to machine learning,
0:25:15 basically like these learned approaches
0:25:17 to modeling the complexity of reality, right?
0:25:20 Like basically like, how do I take a bunch
0:25:23 of 2D images of the world like this
0:25:25 and essentially have a model do this
0:25:26 inverse rendering problem?
0:25:28 Or it’s like, oh, here’s where these hundred photos
0:25:29 are located in 3D space.
0:25:34 Based on this, I’m literally going to like eyeball ray tracing
0:25:37 like and create a 3D representation that makes sense.
0:25:41 And since you know exactly what the model looks like
0:25:42 at the photos that you’ve taken,
0:25:44 the representation that you get eventually
0:25:46 is like good enough from all viewpoints.
0:25:49 And so like, this basically the first nerve paper
0:25:52 dropped in 2021 called neural radiance fields.
0:25:54 And then there’s just been insane progress.
0:25:56 Like we talked about from like data centers
0:25:59 to like the GPU and your fricking like, you know
0:26:03 Nvidia workstation to like the iPhone in your pocket.
0:26:04 But even this wasn’t new.
0:26:06 There was like spiritual successures
0:26:09 to these like ML based learned representations
0:26:12 to sort of encapsulate the complexity of reality.
0:26:13 Enter radiance fields, right?
0:26:17 Like the way like, think about radiance fields generally
0:26:22 sort of like imagine like a voxel grid, a cube of cubes
0:26:25 where every single cube has like a color value
0:26:28 and an alpha like transparency value.
0:26:30 And like that’s kind of what you end up getting with a nerve.
0:26:32 And then when you do volume rendering,
0:26:35 you can basically like, you know
0:26:37 end up getting these photo realistic renditions
0:26:38 of the world.
0:26:41 And so like the cool part about neural radiance field is
0:26:43 instead of photogrammetry where you get this
0:26:46 like 3D mesh model, this like with surfaces
0:26:48 with textures plastered on it.
0:26:51 Think of it like crappy GTA looking models.
0:26:54 What you get with nerfs is like a radiance field
0:26:57 this voxel grid of like all these voxels
0:26:59 and their various values they’re in
0:27:02 that change based on how the camera is looking at it.
0:27:05 And because of that, you get all these things
0:27:06 that photogrammetry couldn’t do
0:27:10 which is like modeling transparency, translucency
0:27:14 like fricking like glass, like shiny objects
0:27:17 all this stuff can be done fricking fire,
0:27:19 volumetric effects, all the stuff that photogrammetry
0:27:21 can’t do cause imagine needing to come up with like
0:27:24 a cardboard paper mache model of that thing.
0:27:25 It’s gonna look like crap.
0:27:28 How do you model hair, fire, fog, all these things.
0:27:30 And you can do all of that
0:27:32 with these implicit representations.
0:27:35 Now the problem with nerfs were the rendering speed
0:27:37 because you’ve got this voxel grid
0:27:38 and you’re doing this volume rendering
0:27:40 where you’re like first doing like
0:27:41 the training process takes forever
0:27:43 but then when you want to render an image
0:27:44 you got to do volume rendering
0:27:47 and like trace like these rays through that voxel grid
0:27:51 and add up these like values like that takes a lot of time.
0:27:54 And basically it’s like think of it like one frame per second
0:27:57 to render out some of these videos, right?
0:27:58 Along comes Gaussian splatting
0:28:01 which is like, hey, do we even need the neural part
0:28:02 of radiance fields?
0:28:04 Like do we need like ML?
0:28:08 Can we just do this with like all school statistical techniques?
0:28:10 And like, which is kind of wild, right?
0:28:13 And so instead of having this implicit black box representation
0:28:17 where like realities modeled in the weights of this like MLP
0:28:19 this like multi-layer perceptron,
0:28:22 you’ve got this explicit representation
0:28:25 of these like ellipsoidal splat looking things
0:28:26 called Gaussians.
0:28:28 Just think of them like super stretchy like fricking spheres
0:28:33 like turns out you can get like a huge jump in quality
0:28:37 while also being able to render way, way faster.
0:28:40 And so like it’s like from one FPS
0:28:42 you’re getting like a hundred frames per second.
0:28:44 And since it’s an explicit representation
0:28:46 it’s in this like most formats
0:28:48 like all these apps that I’m showing on the screen
0:28:52 it’s in this format called PLY the Stanford PLY file
0:28:53 you can basically bring it
0:28:56 into any industry standard game engine.
0:28:57 Like you can bring it into Blender
0:29:00 you can bring it into Unreal into Unity.
0:29:03 And since it’s not this like like black box
0:29:05 like this neural network that you have to deal with
0:29:07 and it’s explicit you can go and delete
0:29:09 and edit things far more easily.
0:29:13 And so it’s like super crazy to see like what’s happened there
0:29:16 but basically between nerfs and Gaussian splatting
0:29:19 think of Gaussian splatting basically as radiance fields
0:29:21 without that neural rendering part.
0:29:23 And the paper uses terms like training or whatever
0:29:27 but there’s no neural networks involved at all in 3DGS.
0:29:29 So yeah, like how crazy is that?
0:29:30 We went from like cool.
0:29:32 Yeah, you could do like cool fly through videos
0:29:34 if you remember that was what the early days
0:29:36 of like the Luma app was do the scan
0:29:38 and now you can reanimate the camera
0:29:40 and you left this thing render for 20 minutes
0:29:41 and you got back something.
0:29:43 Now you can literally take your scans
0:29:45 and drop them into these real time environments.
0:29:47 And it’s like fricking amazing.
0:29:50 Like I think on the left I’m getting like 400 FPS
0:29:52 on an NVIDIA GPU and on the right
0:29:55 I’ve got this thing in Unreal Engine.
0:29:59 And what the cool part of it is like unlike photogrammetry
0:30:02 like and very similar to, you know, neural radiance fields
0:30:05 they still model these light transport effects.
0:30:07 So like again, like imagine if this was like
0:30:08 a cardboard cut out model
0:30:10 you wouldn’t have had all these light transport effects
0:30:13 of the light going through the tree, et cetera.
0:30:17 And so the way Gaussian splatting does this
0:30:20 is like by using this OG physics concept
0:30:23 called spherical harmonics to model it.
0:30:25 And so like if you’re trying to optimize stuff
0:30:27 you can get rid of some of these view dependent effects
0:30:29 as they’re called view dependent meaning
0:30:30 as you change your view
0:30:32 that like materials look slightly different
0:30:35 but you basically get it all with Gaussian splatting.
0:30:37 So I think it’s super exciting.
0:30:40 And yeah, like you could do this stuff in the cloud
0:30:43 you can do this stuff on your fricking desktop now.
0:30:45 Like I think post shot is a tool
0:30:46 that not many people have used
0:30:48 but like if you’re working on a commercial thing
0:30:50 and you don’t wanna upload your data
0:30:52 with Luma’s terms of service
0:30:53 or Polycam’s terms of service
0:30:58 like you can train this all locally on your desktop
0:31:00 with post shot, with Nerf Studio
0:31:03 the sum of the models Nerf Studio aren’t commercial friendly
0:31:04 and then even in the phone in your pocket, right?
0:31:06 Like so if you’ve got an iPhone like a modern iPhone
0:31:09 like and you just wanna know what 3D
0:31:12 like what radiance fields and reality captures all about
0:31:15 just download the Scaniverse app and like have at it.
0:31:17 So this is maybe a dumb question
0:31:19 but like so with like Nerfs and all this new tech
0:31:22 are you able to make like a really realistic 3D model
0:31:23 of like a city like San Francisco?
0:31:25 I mean, is that what you’re showing me earlier?
0:31:27 Or like, or is it only like a certain scene?
0:31:28 Like how hard is that?
0:31:30 Yeah, I mean, there’s a bunch of new papers outright
0:31:31 the initial radiance fields.
0:31:33 Like so there’s a paper called Block Nerf
0:31:35 that tries to scale Nerfs up to city scale
0:31:37 using Waymo data sets.
0:31:39 And similarly you’re seeing in the Gaussian splatting world
0:31:42 different papers about basically having like
0:31:46 in a kind of like nested hierarchies of splats
0:31:48 that have really good transitions to model
0:31:51 an entire like city and eventually the globe.
0:31:54 So I think like that’s the path that academia
0:31:55 and industry is on.
0:31:58 And I think like already you’re seeing city scale data sets
0:32:00 that are very plausible in research.
0:32:02 And I think it’s only matter of time
0:32:05 before that stuff gets into production.
0:32:06 You think that’s like the future of Google Maps?
0:32:08 I think it’s the future of maps for sure.
0:32:11 Like, you know, in immersive view
0:32:13 there are certain indoor locations
0:32:15 where you get a pre-rendered neural radiance field
0:32:18 that you can kind of like walk around and see.
0:32:20 This is just the evolution of that.
0:32:21 I think those data sets exist
0:32:23 and there’s like a handful of companies
0:32:25 in the world that have it.
0:32:27 So I think like that is the future of geospatial
0:32:29 and like maps in general.
0:32:31 But on the other hand, I think what’s interesting is like,
0:32:33 you know, this technology like building a map
0:32:35 of the world is easy.
0:32:36 Updating it is way harder, right?
0:32:39 Like when people talk about this like
0:32:42 one to one digital twin of reality, it’s like,
0:32:45 oh yeah, well like, by the way, new stuff
0:32:47 has built all the fricking time.
0:32:48 Things change all the time.
0:32:51 Seasonality is a fricking thing, right?
0:32:54 Like so, I think with this technology we’ve got
0:32:56 and since we’ve commoditized capture
0:32:59 because sensors are cheaper, computers cheaper.
0:33:02 And now we’ve got access to the same sort of algorithms
0:33:05 and approaches to model reality.
0:33:07 I think like updating this model of the world
0:33:08 is gonna get a lot, lot easier.
0:33:10 So I think like it’s gonna be very exciting
0:33:12 or in the near future we’re walking around
0:33:14 you know, driving around our cars
0:33:17 and walking around with our like glasses or whatever.
0:33:19 And we’re sort of updating this real time map of the world.
0:33:22 I think we’re very much on that trajectory
0:33:24 and we’re closer now than we’ve ever been.
0:33:26 – What do you think are like the business applications
0:33:27 to this tech?
0:33:28 – I mean, it’s like all the applications
0:33:30 that value stuff in the real world,
0:33:32 there’s utility and there’s delight, right?
0:33:35 I think like being able to not just like,
0:33:37 I mean, if you look at what NVIDIA is doing with Earth too,
0:33:39 right, we’re talking about the physical structure
0:33:39 of the world.
0:33:42 You can think of like the Earth having very facets, right?
0:33:44 Like there’s like the sort of the terrain,
0:33:48 like the natural like physical features of the Earth.
0:33:49 Then it’s all the like, you know,
0:33:51 human built things on top of that,
0:33:52 the structure that we built.
0:33:55 And then you can layer in like human activity
0:33:56 on top of that, right?
0:33:58 Like us moving around in the world,
0:34:00 our sensors, our cars, et cetera.
0:34:03 And then there’s other phenomena like weather, right?
0:34:04 Like tides and things like that
0:34:06 that need to be incorporated.
0:34:08 So Earth too is this really interesting initiative
0:34:11 by NVIDIA to focus on the weather like systems
0:34:13 that govern like, you know,
0:34:16 basically a day-to-day weather in the real world.
0:34:18 And so if you’ve got that understanding of like,
0:34:20 like the structure and geometry of a place
0:34:21 where the sun is going to be,
0:34:23 you can predict already things like,
0:34:25 hey, can I install solar panels here?
0:34:27 Like actually how much like sunlight would I get
0:34:30 if I installed this configuration of panels?
0:34:32 Then when you layer weather on top of that,
0:34:33 things get even more interesting.
0:34:35 So to answer your question,
0:34:37 I think there’s a bunch of applications
0:34:39 across utility and delight.
0:34:42 Like media and entertainment’s obviously in gaming is like,
0:34:45 I think the next GTA is absolutely going to be built
0:34:46 in like a twin of the real world.
0:34:49 Maybe this is the last GTA that we beat
0:34:51 will be built by humans manually
0:34:52 to emulate the real world.
0:34:55 I think that’s certainly exciting.
0:34:57 That said, a bunch of games have already used
0:34:58 reality capture, right?
0:35:00 Like from call of duty to battlefront, et cetera.
0:35:03 But I think the utilitarian aspects
0:35:05 are far, far more interesting.
0:35:07 Whether you’re like anything you’re trying to do
0:35:10 in the world of bits, like from like building stuff
0:35:12 to like disaster planning,
0:35:16 like the range of applications is just immense.
0:35:17 – Well, even just, you know,
0:35:21 one of the things that Jensen showed off at GTC this year
0:35:25 was to create these sort of virtual worlds
0:35:28 and then actually put virtual versions of like
0:35:31 humanoid robots in these worlds
0:35:34 and to sort of train them on this virtual twin
0:35:35 of the real world.
0:35:37 So they know how to navigate the real world
0:35:39 and then once they get that training data
0:35:41 then they can sort of inject that training data
0:35:43 into the real robots.
0:35:46 So like this concept of creating this digital twin
0:35:50 of the earth will allow us to train a lot of these robots
0:35:54 and machinery to operate within that digital twin
0:35:56 before actually deploying it in the real world.
0:35:59 To me, there’s a lot of huge implications there.
0:36:02 – 110%, I mean like these,
0:36:05 it’s like a way of creating all the training data
0:36:07 that these machines and perception models need
0:36:08 to be able to navigate the world, right?
0:36:10 And what better way it’s like,
0:36:12 you can create that like, you know,
0:36:14 you can 3D scan a city block
0:36:17 and then create all these different scenarios
0:36:19 of human activity on top of that
0:36:21 and feed that to, you know, like,
0:36:23 and train like self-driving cars,
0:36:25 like, you know, self-driving AI off of it.
0:36:28 I think like the fact that we’ve got a place
0:36:29 where we can basically like,
0:36:32 we can teleport reality into the digital world
0:36:34 and then also manifest the digital in the real world.
0:36:37 It’s like that bridge I think is just very powerful
0:36:38 for a bunch of different applications.
0:36:41 – Well, I want, let’s talk super quickly about TED.
0:36:45 So I’m, first of all, congrats on even having a TED talk.
0:36:47 Like that’s such an amazing accomplishment.
0:36:50 Like, you know, some people say they’ve had a TED talk.
0:36:51 They’re really talking about a TEDx talk.
0:36:54 And come on, come on, come on.
0:36:57 You’ve actually given a real TED talk, a legit TED talk.
0:36:58 And not only that,
0:37:01 but they asked you to host the TED AI podcast.
0:37:03 So tell us a little bit about that
0:37:04 and what’s going on there.
0:37:06 I mean, maybe share a little bit of your experience
0:37:09 with TED and then tell us about the TED AI podcast.
0:37:10 – Yeah, sure.
0:37:14 I mean, certainly the TED talk was a fun experience last year.
0:37:16 And I would say this year was even more fun.
0:37:18 It’s a side of the opportunity.
0:37:19 It’s a co-session too,
0:37:22 which was all of that AI with Chris Anderson.
0:37:24 And we had some amazing speakers,
0:37:28 like Vinod Kosla, Feifei, you know,
0:37:31 the CEO of GitHub, Helen Toner,
0:37:33 export member, OpenAI.
0:37:36 And even like, I don’t know if you’ve,
0:37:38 if you’ve checked her work out, but nice aunties,
0:37:42 like absolute trip down basically like,
0:37:46 to me what intergalactic social media looks like.
0:37:48 There was a super, super fun experience.
0:37:49 Yeah, I mean, like, look,
0:37:51 the T and TED is all about technology.
0:37:54 And I think right now what’s exciting is to put the,
0:37:58 like over time, TED grew to encompass, you know,
0:38:00 not just technology, entertainment and design,
0:38:03 but a plurality of topics, right?
0:38:05 And I think with AI sort of in technology,
0:38:06 being sort of this horizontal,
0:38:08 like tech is a horizontal,
0:38:11 but it’s impacting so many different verticals
0:38:13 in our daily lives, right?
0:38:15 Like we can talk about all the applications,
0:38:16 whether you’re a creator,
0:38:18 whether you’re a knowledge worker, you know,
0:38:20 whether you’re a musician, you know,
0:38:22 whether you’re thinking about like national security
0:38:25 and defense, whether you’re thinking about relationships.
0:38:30 And often in all of these sort of topics, you know,
0:38:34 there’s like a dichotomy that we as like builders
0:38:36 and consumers have to contend with.
0:38:40 And so the idea of the TED AI show really is to outline
0:38:42 those dichotomies and, you know,
0:38:45 not necessarily take an opinion one way or the other,
0:38:49 but sort of elaborate on the entire gamut of like,
0:38:51 like the good, bad and the ugly
0:38:53 and sort of let people decide for themselves
0:38:56 and do that by talking to people from all walks of life.
0:38:58 Like people whose titles haven’t even been invented yet,
0:39:01 but obviously technologists, journalists, researchers,
0:39:03 artists, you know, the list goes on
0:39:05 and, you know, I’m just super grateful
0:39:08 for the opportunity to be able to, you know,
0:39:10 just bring my excitement into the space.
0:39:12 Like obviously like I want to build,
0:39:14 bring the lens of a creative,
0:39:16 like that’s like built a following
0:39:18 for over a million folks using these tools,
0:39:20 but also as a product builder who shipped a bunch
0:39:22 of this stuff and then just like,
0:39:26 I would say like cautiously optimistic AI enthusiasts.
0:39:28 So I’m going into a bunch of these topics
0:39:30 with those three lenses in mind
0:39:32 and it’s just been a lot of fun.
0:39:35 We’ve got some really cool episodes lined up for y’all
0:39:38 and I can’t wait for y’all to check it out.
0:39:41 – You have any idea of like, you know, launch schedule?
0:39:44 Like is that, are there dates planned out for it yet?
0:39:45 – Yeah, yeah, totally.
0:39:47 So May 21st, first episode drops
0:39:48 and then it’s going to be weekly.
0:39:50 There’s going to be a little bit of a summer break there,
0:39:52 but yeah, 25 episodes in the season.
0:39:55 And let me tell you,
0:39:57 I think there’s something for everyone.
0:39:59 – Yeah, I do wonder like what was the general vibe at Ted?
0:40:03 Like are people optimistic or are they like really fearful
0:40:04 of AI?
0:40:05 – And if you look at Ted also, you know,
0:40:08 you had people like, like the laval.
0:40:10 I know there were some other speakers there.
0:40:11 I think maybe Moustafa Suleiman was there.
0:40:13 Maybe that was a more recent one.
0:40:15 But then you also had guys like Gary Marcus
0:40:20 and I’m going to totally butcher his name, Yadkowski.
0:40:22 – Oh yeah, it is Yadkowski.
0:40:24 – Which are both more on the like,
0:40:26 hey, let’s chill out on the AI side.
0:40:29 So it seems like it’s from a speaker’s front.
0:40:31 It seemed like they had speakers on both sides
0:40:33 of the arguments.
0:40:33 – Definitely.
0:40:35 I mean, the theme for this year was like the brave
0:40:36 and the brilliant and like covering
0:40:38 that gamut of opinions.
0:40:40 I would say overall the vibe is positive.
0:40:42 So I like, I’ll give you a sample size.
0:40:44 Like I taught this discovery session,
0:40:48 which was about the dichotomies of AI is about 50 people.
0:40:49 And sort of what we did is we looked at a bunch
0:40:52 of these verticals in the AI space.
0:40:54 And like, you know, essentially came up with like,
0:40:56 what happens if this goes really well
0:40:59 and what happens if this goes really poorly?
0:41:01 And like, let’s use chat GPT actually
0:41:02 to come up with like a headline,
0:41:06 like a piffy visual depiction of that desirable
0:41:08 and undesirable future.
0:41:10 And honestly, like most folks in the room
0:41:12 are optimistic about it, right?
0:41:14 Like, but they’re not blind to the downsides.
0:41:17 I think like the problem with anything is extremes, right?
0:41:19 And so like, you know, where like Nathan,
0:41:22 I hear your concern of like, if we’re like, you know,
0:41:23 oh, like we can’t have like,
0:41:27 demerism and like, you know, I was like, it’s infectious.
0:41:30 I think the same thing applies to the opposite narrative too,
0:41:32 which is like, well, we just obviously,
0:41:34 we got to keep like accelerating.
0:41:35 We got to keep shipping.
0:41:36 I think it depends, right?
0:41:39 Like, it’s sort of the boring answer to these things.
0:41:42 And I think you can’t understand the nuances
0:41:45 unless you go dissect like the full gamut
0:41:47 of like considerations.
0:41:50 And so the goal is like, really like the editorial perspective
0:41:51 I’m trying to bring.
0:41:55 And of course, Ted has a huge say in this too is like,
0:41:58 look, I’m a 60% optimistic, 70% optimistic,
0:41:59 which is not too dissimilar.
0:42:01 I think Matt, well, you and I have talked about
0:42:03 on most things, but I’m not going to be blind
0:42:06 to all like the downsides of this stuff too, right?
0:42:09 Like, and I think that’s okay to say.
0:42:12 And what I believe is just like the speaker selection,
0:42:15 you know, we’re trying really hard
0:42:17 to have a balanced perspective on the guests too.
0:42:20 So that like, what’s what I think is going to happen
0:42:22 in the real world as well as like,
0:42:24 you’ll be able to hear both sides of that argument.
0:42:27 Like somebody who’s like super stoked and thrilled
0:42:30 on like AI art and just thinks it’s the bee’s knees.
0:42:33 And it’s totally cool to train on copyrighted material.
0:42:35 Somebody who thinks like that is the death of creativity
0:42:36 as we know it too.
0:42:37 And you have to-
0:42:38 – I think it’s good to hear both sides.
0:42:40 I mean, I agree, but I’m concerned.
0:42:41 Like I used to live in San Francisco
0:42:43 and like they’re like pushing for regulation now.
0:42:44 We’re like to-
0:42:45 – Who’s they?
0:42:46 You mean, you mean, Sam Allman?
0:42:47 – The government, yeah.
0:42:48 Well, no, the government.
0:42:49 They’re like, they’re like pushing a bill through right now.
0:42:51 They’re trying to like, they’re trying to fast track it.
0:42:53 I forgot who’s doing the bill,
0:42:55 but we’re basically,
0:42:57 if you launch a new large language model-
0:42:58 – Need approval.
0:43:00 – Yeah, you like need approval,
0:43:02 but also like it doesn’t get to basically sign something
0:43:03 and like it’s perjury if you’re lying
0:43:05 that this model can do no harm.
0:43:06 And it’s like-
0:43:07 – Yeah, like who’s gonna do that?
0:43:08 Yeah.
0:43:09 – Yeah.
0:43:12 And so I agree like nuance is important and like,
0:43:15 I do consider myself kind of part of the EAC movement,
0:43:16 but more generally just like a techno optimist.
0:43:18 – Do you have it in your bio still?
0:43:19 I don’t know.
0:43:20 – I don’t, I don’t.
0:43:21 But I, you know, I like Beth.
0:43:23 I like all the people who are part of that.
0:43:24 – That’s cool.
0:43:25 – I think in general,
0:43:27 I think in general it’s right, you know,
0:43:29 but nuance is important, I agree.
0:43:30 – Let me put it this way.
0:43:31 I think the way I see it is like,
0:43:34 we’ve got enough talented humans out there
0:43:37 that are like pushing for like acceleration
0:43:39 and there’s enough talented people out there
0:43:41 that are pushing for, you know,
0:43:43 I would say pumping the brakes
0:43:46 for the lack of a better way to put it in certain areas.
0:43:48 And I think like in totality,
0:43:52 like we’ll reach like some optimum solution
0:43:53 because of those influences.
0:43:54 And I think it’s always been like that, right?
0:43:58 Like, I mean, just like the early days of music,
0:44:00 everyone’s like, oh yeah, Napster and like peer-to-peer
0:44:01 and let’s just go crazy.
0:44:03 And then things settled down
0:44:05 and we found a business model that worked.
0:44:07 Maybe it’s not perfect, right?
0:44:09 Like people have a lot of gripes with the Apple
0:44:10 and the Spotify business model,
0:44:13 but I think we found this like globally optimum solution.
0:44:15 – I mean, it doesn’t always work out, right?
0:44:16 Though they look at like nuclear plants, right?
0:44:18 Like in the past, the US was gonna build
0:44:21 all these nuclear plants to solve energy problems
0:44:23 and we didn’t do it because of regulations
0:44:24 and because of fear.
0:44:25 And now we’re like trying to solve
0:44:26 all these global warming problems.
0:44:29 Other thing is like, well, we always kind of had nuclear there
0:44:31 that we could have been using and it worked.
0:44:32 So it doesn’t always work out.
0:44:34 Like it often does, but it doesn’t always.
0:44:37 – Totally, I mean, one of the funnest topics
0:44:42 is getting into how do politicians and regulators
0:44:46 even regulate this like sort of nebulous set of technologies.
0:44:48 It’s not just large language models, right?
0:44:50 Like there’s all the perception AI stuff
0:44:52 and like the implications there.
0:44:55 And but like a group of these technologies
0:44:57 that are like sort of permeate and go, my God,
0:45:00 like some of the stuff that I’m like doing research
0:45:04 right now on just the intersection of neuroscience and AI
0:45:07 and what we’re gonna be able to do with just like passive
0:45:10 neural like interfaces with earbuds and things like that.
0:45:14 I mean, there’s gonna be some real big ethical quandaries
0:45:15 that pop up.
0:45:18 And so yeah, yeah, trying really hard to bring like
0:45:19 still be techno-optimist,
0:45:21 but bring that balanced perspective.
0:45:23 And I think folks are gonna like it.
0:45:24 – I think at the end of the day,
0:45:26 the important thing is empathy, right?
0:45:30 Like the perspective I come from is I tend to be
0:45:31 a very empathetic person, right?
0:45:34 I want to hear both sides of the story.
0:45:35 I want to hear both perspectives
0:45:38 and I want to be empathetic to both sides too.
0:45:40 Like if somebody is genuinely worried
0:45:43 that this technology is taking their job,
0:45:44 I want to understand why.
0:45:48 I want to understand what we can do to sort of mitigate
0:45:50 the damages that could be done from this.
0:45:53 Like I’m always gonna come from that place of empathy,
0:45:55 which is why I’ve never sort of identified
0:45:58 with the EAC movement is like,
0:46:00 I don’t necessarily think we should always be pushing
0:46:03 everything forward as fast as possible.
0:46:05 I think we should be listening to the fears.
0:46:06 We should be listening to the concerns.
0:46:09 We should be figuring out sort of middle grounds.
0:46:13 Like you mentioned, there’s always people on both sides
0:46:15 which kind of creates a decent checks and balance
0:46:17 to make sure one side doesn’t go too far
0:46:20 and AI nukes the world,
0:46:22 but also the other side doesn’t go too far
0:46:25 and technology stops advancing completely.
0:46:29 Those checks and balances I think are a net positive overall.
0:46:30 I think those need to be there.
0:46:33 – I agree overall, but I mean, I think the big thing
0:46:36 that was like, the argument with EAC would be making is
0:46:39 that compounding is one of the most important powers
0:46:40 out there, right?
0:46:41 Like an idea that if we build faster
0:46:43 than the technology in the future will be better
0:46:44 and better and better and we’ll start solving
0:46:46 real world problems like cancer
0:46:48 and all these other things that maybe we could have been doing
0:46:52 if we weren’t so like quick to just regulate everything.
0:46:54 And so I think with AI, it’s the same thing.
0:46:56 Like sure, like some regulation in the future
0:46:58 might make sense, but if we just start throwing it
0:47:00 out there right now, we’re gonna slow down the compounding
0:47:03 and like the exponential, we’ll stop the exponential
0:47:05 from happening with our regulations.
0:47:08 And we, you know, yeah, maybe some jobs would be lost
0:47:10 in the short term, but in the long term,
0:47:13 we could have cured cancer, we could have solved, you know
0:47:15 global warming issues, all kinds of other problems
0:47:18 that we could have solved if we would just like weighted
0:47:19 and see what happens with the technology.
0:47:20 And yeah, now there’s a big problem.
0:47:23 Okay, maybe make a regulation, but just don’t do it
0:47:23 like right at the beginning,
0:47:24 like they’re trying to do right now.
0:47:27 – I mean, India totally flipped their decision, right?
0:47:29 Initially they were like, oh, you have to get every model
0:47:31 approved and they’re like, actually we’re gonna retract
0:47:33 this part, which was really interesting.
0:47:36 And I mean, like it’s like a regulation also,
0:47:38 like it could end up in a place where it just only benefits
0:47:41 the incumbent, like the largest AI labs too, right?
0:47:44 Like the regulatory capture like point of view,
0:47:47 it’s like it could end up in a place where like any new
0:47:50 innovation like from a startup can actually happen.
0:47:52 And they’re the ones that get these like owners compliance
0:47:55 requirements and they can’t afford a team of lawyers
0:47:57 unless they’re like super VC backed.
0:47:59 And then like, what an inefficient use of VC capital
0:48:02 instead of innovating, you’re sort of like navigating
0:48:05 like the legal landscape of like a heavily regulated
0:48:05 industry.
0:48:07 So I think like your point about nuclear is well taken too.
0:48:10 It’s like, I mean, like a lot of folks, I mean,
0:48:12 including Gary brings up the example of like, you know,
0:48:16 like basically air travel and airplanes are still well
0:48:19 regulated, but yet we’ve got this whole Boeing fiasco
0:48:20 happening, right?
0:48:23 It’s like where you’ve got one really big incumbent
0:48:25 and there’s probably a revolving door between regulatory
0:48:29 agencies and Boeing and, you know, so it’s-
0:48:30 – And when the Wright brothers got started, I mean,
0:48:31 they weren’t being heavily regulated as they were
0:48:33 like inventing the plane.
0:48:34 – Totally.
0:48:35 – They were out there just like out in Ohio,
0:48:37 just, you know, trying shit out.
0:48:37 So I mean-
0:48:39 – I think one thing it’s like a lot of people talk about
0:48:42 is like just like technology, sunsets, like regulations
0:48:45 need to sunset rather than us adding up more and more
0:48:46 regulations.
0:48:48 So I think this is where things get geopolitical too.
0:48:51 It’s like, I think China is so much savvier about AI
0:48:53 regulation than the US is right now.
0:48:57 And I feel for the politicians, I think they’re like asking
0:48:59 for this type of engagement.
0:49:01 And I think it’d be good if we engage with them on this
0:49:05 and like bring those perspectives to bear rather than,
0:49:07 I don’t know, just being like a regulation, bad,
0:49:10 innovation, good, let’s keep innovating.
0:49:13 And so it’s like, it’s nuanced, but then again, like,
0:49:18 look, I’m a, I’ve always been a Libra and like, you know,
0:49:21 kind of trying to build bridges between two worlds.
0:49:25 – Yeah, no, I mean, I think this being a super nuanced
0:49:28 conversation is like the understatement of the episode here.
0:49:31 I think there’s just so many different like rabbit holes
0:49:33 that we could potentially go down when it comes to the
0:49:34 regulation thing.
0:49:37 I think you’re going to have to be one of our sort of
0:49:38 recurring guests.
0:49:41 Maybe every few months jump on and nerd out about this stuff.
0:49:44 But, you know, I do want to give you the opportunity to
0:49:46 tell us what else you’re working on.
0:49:48 If there’s a place you think people should go check you
0:49:51 out, your Twitter, your YouTube, obviously the TED AI
0:49:53 podcast coming out later in May.
0:49:56 – Yeah, so just please follow me on Twitter @belavosadoo.
0:50:01 You can also follow me on YouTube and TikTok @billyfx.
0:50:03 If you’re interested in some more long form expositions
0:50:05 that I do, check out the Creative Tech Digest.
0:50:08 It’s both a newsletter as well as a YouTube channel.
0:50:10 And yeah, of course, check out the TED AI show.
0:50:14 Maybe the one last thing I’ll say is like if you’re a
0:50:16 founder and a builder in this space building with any of
0:50:19 the technologies that we talked about and you’re looking
0:50:21 for early stage investment, I’m also a scout for
0:50:23 A16Z Games.
0:50:26 So just hit me up on Twitter or you can email me.
0:50:28 We’ll put the email in the show notes as well.
0:50:30 And I really appreciate you guys having me on.
0:50:32 I’m wishing you all the success for your podcast.
0:50:36 And Matt, I will see you at I/O.
0:50:39 And Nathan, I hope to see you in 3D sometime soon.
0:50:42 – Oh, come out the Kyoto, come on.
0:50:43 – I gotta make it happen.
0:50:44 – Awesome, Bill of All.
0:50:45 Well, it’s been a blast.
0:50:47 This has been one of my favorite conversations we’ve had
0:50:47 so far, so fun.
0:50:50 Excited to see you in person next week.
0:50:51 – Cool, cheers.
0:50:53 (upbeat music)
0:50:56 (upbeat music)
0:50:59 (upbeat music)
0:51:01 (upbeat music)
0:51:04 (upbeat music)
0:51:06 [MUSIC]
0:00:03 – I’m still very bullish on Google
0:00:05 because I think it’s like the tip of the iceberg
0:00:07 of what you see in tech companies
0:00:09 and the submerged part is just amazing.
0:00:10 They’ve got the most complete models
0:00:12 of the physical and digital world.
0:00:15 They’ve got ubiquitous distribution
0:00:17 and an existing ecosystem of ads and sales
0:00:19 to plug monetization in.
0:00:21 So I think it’s still a magical combination.
0:00:27 – Hey, welcome to the Next Wave podcast.
0:00:28 I’m Matt Wolf.
0:00:30 I’m here with my co-host, Nathan Lans.
0:00:33 And with this podcast, it is our goal
0:00:36 to keep you looped in with all of the latest AI news,
0:00:38 the latest AI tools,
0:00:41 and just help you keep your finger on the pulse
0:00:43 so that you are prepared for the next wave of AI
0:00:44 that’s coming.
0:00:47 And today we have an amazing guest on the show.
0:00:50 We have Bilival Sidhu on the show.
0:00:52 He is the host of the TED AI podcast.
0:00:54 He’s an ex-Googler.
0:00:56 We’re gonna talk to him about what it was like
0:00:59 working on AI and visual effects over at Google.
0:01:02 We’re gonna talk about the difference
0:01:04 between whether or not we should be accelerating AI
0:01:06 or slowing down AI.
0:01:08 We’re also gonna learn about how some
0:01:11 of these AI visual effects tools work
0:01:13 because this is the field that Bilival worked in
0:01:14 for so long.
0:01:15 It’s an amazing episode.
0:01:17 You’re gonna learn a ton
0:01:18 and can’t wait to share it with you.
0:01:20 So let’s jump on in with Bilival Sidhu.
0:01:24 Thanks so much for being on today, Bilival.
0:01:25 – Thanks for having me, gentlemen.
0:01:26 Pleasure to be here.
0:01:29 – Yeah, so I wanna just kind of dive right into it.
0:01:32 Your background is Google, right?
0:01:35 So I think when you and I first connected
0:01:38 and we first started having some chats over Twitter DM,
0:01:41 you were still actually working over at Google at the time
0:01:43 and you also were kind of doing a creator business
0:01:45 on the side with your YouTube channel
0:01:46 and everything you had going on.
0:01:48 But Bilival Sidhu, what were you doing over at Google?
0:01:49 What was your role there?
0:01:52 You know, what was your experience like over there?
0:01:53 – Gosh, yeah, it was awesome.
0:01:55 So, you know, I’ve spent a decade in tech,
0:01:57 six years at Google.
0:01:58 I’ve been able to work on projects
0:02:00 that blend the physical and digital world.
0:02:03 And I started off in the AR/VR team really
0:02:06 when spatial computing, as it’s now called,
0:02:07 was first popping off.
0:02:10 This is like right after the DK2 came out,
0:02:11 Google Glass was a thing.
0:02:15 And everyone was talking about what is the next iteration
0:02:16 of computing platforms?
0:02:19 Where are we gonna go from like this mobile revolution?
0:02:22 And so I had a chance to work on a bunch of cool stuff there,
0:02:26 YouTube VR content, livestreaming Coachella,
0:02:28 Dean Choice Awards, Elton John,
0:02:30 the camera systems that we used
0:02:33 to do stereoscopic 3D capture, AR SDKs,
0:02:35 when that became popular, augmented reality,
0:02:37 sort of hit the scene.
0:02:40 And then after that, I spent four years at Google Maps,
0:02:43 basically creating a ground up 3D model of the world,
0:02:45 sort of remapping the world, if you will,
0:02:48 and then turning the world into an AR canvas
0:02:51 with the AR Core Geospatial API.
0:02:53 It’s been a lot of fun and yeah,
0:02:55 it’s been awesome to work with some really talented folks
0:02:59 to work on these projects that have been blurring this line
0:03:01 between the world of bits and atoms.
0:03:03 – So I’m curious, working on all this stuff,
0:03:05 in my mind, I can’t even imagine
0:03:07 what a day-to-day looks like at Google.
0:03:09 I’ve been on the Google campus
0:03:13 and it looks like a giant playground for tech nerds.
0:03:14 So I’m just kind of fascinated
0:03:17 by what it’s like to work at Google.
0:03:19 Like what does a day-to-day look like over there?
0:03:20 – You know, so I was a product manager
0:03:23 and so a day-to-day for me is gonna be very different
0:03:26 than if you go talk to like an engineer or designer.
0:03:27 For me, really it was a lot of meetings.
0:03:30 Let me be perfectly honest, just like a ton of time.
0:03:31 But there’s some very cool things.
0:03:34 I think like Google and big tech companies generally
0:03:37 are sort of this like interesting microcosm
0:03:39 where it’s like, you know, I’ll send out an email
0:03:43 and the guy that wrote the book on computer vision,
0:03:46 like the computer vision book that like everyone reads,
0:03:48 like a response to it and I get a bunch of pings
0:03:51 and you’re like, oh, so-and-so respond to your thing.
0:03:53 And it’s like all these Pokemon that like, you know,
0:03:55 these companies have caught that are available
0:03:58 at your back, back in call to share ideas with,
0:04:01 pull into your own projects and really just like,
0:04:05 you know, there’s such a, it’s like the tip of the iceberg
0:04:07 of what you see in tech companies
0:04:09 and the submerged part is just amazing.
0:04:11 So it’s like, when I moved over to the maps team,
0:04:14 I was thinking of working on glasses at the time.
0:04:16 And the reason I went to maps is like,
0:04:18 I met this engineer is like,
0:04:21 oh yeah, we write CLs to move satellites around in the sky.
0:04:23 And I was like, wait, what, huh?
0:04:25 You move satellites around the sky.
0:04:27 And it’s like, yeah, like literally orchestrating a fleet.
0:04:29 Like, you know, like most people don’t know this,
0:04:31 like Google owns their own fleet of like,
0:04:33 not just street view cars, but airplanes.
0:04:34 Oh, wow.
0:04:36 And so like the ability to like task those.
0:04:38 So like, hey, we need to, you know,
0:04:41 we’ve got this like Sundar IO thing coming up
0:04:44 and we’re gonna be presenting immersive view.
0:04:47 We got to go capture this high resolution model of London.
0:04:48 And like, not, you know,
0:04:51 and suddenly things in the world of atoms
0:04:52 are moving to make that happen.
0:04:54 I think it’s like absolutely amazing.
0:04:56 I think people deeply underestimate the data mode
0:04:58 that Google has.
0:05:00 Obviously the most complete digital twin of the world
0:05:02 we’re talking about, but like search, right?
0:05:05 Like YouTube, oh my goodness.
0:05:07 All this stuff is available.
0:05:09 And so like, there’s cool things and products
0:05:11 you can build around it.
0:05:13 But along with it comes, you know,
0:05:15 which may not be a surprise to people in tech,
0:05:18 but like a ton of responsibility and sort of guardrails
0:05:20 for how you actually use this data.
0:05:21 So it’s like the size of the prize
0:05:24 and the data sets you get to play with are amazing.
0:05:25 But to be able to do stuff with it,
0:05:28 you really, really have to be exceedingly thoughtful.
0:05:30 And there’s a lot of process involved
0:05:32 in unlocking that innovation.
0:05:33 So yeah, that’s how I would describe it.
0:05:35 I think like it’s just like a,
0:05:38 it’s like, it’s like a Disneyland for nerds to be honest.
0:05:43 When all your marketing team does is put out fires,
0:05:45 they burn out fast.
0:05:46 Sifting through leads,
0:05:48 creating content for infinite channels,
0:05:51 endlessly searching for disparate performance KPIs.
0:05:52 It all takes a toll.
0:05:56 But with HubSpot, you can stop team burnout in its tracks.
0:05:58 Plus your team can achieve their best results
0:06:00 without breaking a sweat.
0:06:02 With HubSpot’s collection of AI tools,
0:06:05 Breeze, you can pinpoint the best leads possible.
0:06:08 Capture prospects attention with clickworthy content
0:06:11 and access all your company’s data in one place.
0:06:14 No sifting through tabs necessary.
0:06:16 It’s all waiting for your team in HubSpot.
0:06:17 Keep your marketers cool
0:06:20 and make your campaign results hotter than ever.
0:06:23 Visit hubspot.com/marketers to learn more.
0:06:28 – So it sounds like you’re still really bullish on Google.
0:06:30 Cause I know we were having like some playful banter
0:06:32 last year about like, you know, I was like,
0:06:34 maybe Google’s going to die.
0:06:36 And you were like, what the hell are you talking about?
0:06:38 – Yeah, yeah, we’ve actually had that conversation
0:06:40 on this podcast before where I’m like, you know what?
0:06:42 I think I give more credit to Google.
0:06:43 I think Nathan’s a little more,
0:06:46 I don’t know if they’re going to be like the top dogs in AI.
0:06:48 Like where do you stand on that?
0:06:49 Do you count Google out?
0:06:52 Do you think Google will pass the Microsoft AI,
0:06:54 you know, Avenger mega team?
0:06:56 – It’s hard to say anything with certainty,
0:06:59 but what I will say is like, you know, after I left Google,
0:07:00 I think I was like one of the few people,
0:07:03 I felt like I was in isolation saying good things
0:07:04 about Google.
0:07:06 Everyone was just like, oh yeah, they’re just too slow.
0:07:09 They foambolt, they came up with a transformer.
0:07:11 All the transformer folks are left.
0:07:15 I think it’s a situation where there was no real disruption
0:07:16 in site for the search base.
0:07:18 Yes, there were talks about like, hey, like kids
0:07:21 are like searching on TikTok and like YouTube like now,
0:07:24 but YouTube is owned by Google on TikTok.
0:07:26 It’s kind of short former people are really going to be like,
0:07:27 is that going to be a resilient thing?
0:07:30 And one might argue now social networks are places
0:07:31 where people do a lot of searching,
0:07:34 but traditional search sort of as like, you know,
0:07:37 a maps guide is to just give the maps analogy.
0:07:40 It’s like, you know, like maps is how you discover stuff
0:07:41 in the real world.
0:07:43 And Google is how you discover stuff in the digital world.
0:07:46 It’s literally your window to the worldwide web, right?
0:07:48 I don’t think anything had like sort of questioned
0:07:51 the strong position Google was in in that regard
0:07:55 until chatGPT came out when suddenly people
0:07:57 could start connecting the dots and see you like,
0:08:00 if you connect large language models with like, you know,
0:08:03 knowledge graph and search index, kind of like perplexity.
0:08:06 And you know, Microsoft co-pilots
0:08:09 and whatever the heck else OpenAI is going to announce
0:08:13 in order to kneecap any Google limelight next week.
0:08:14 Like I think people start saying,
0:08:17 hey, there’s a disruption inside.
0:08:19 And I think combined with the fact that like,
0:08:21 the search ads business model
0:08:24 was just such a money printing machine and still is.
0:08:26 And the fact that, you know, the cost per query
0:08:29 of these generative AI models is obviously going to be higher.
0:08:32 And how do you do advertising and attribution
0:08:34 and all this stuff like that it would like,
0:08:37 kind of represents a, you know, contraction
0:08:40 in the money printing machine and the pie
0:08:41 and the business that Google created
0:08:44 had all the signs of sort of innovators dilemma.
0:08:46 And I think like Google has sort of adopted
0:08:48 the playbook of the innovator solution.
0:08:51 And, you know, initially they had some of these reorgs
0:08:54 that felt more like exec reorgs.
0:08:56 And now they’re actually bringing together
0:08:58 like the brain and the deep mind teams
0:09:01 and they’re actually shipping at a really good cadence.
0:09:04 And I think they still have some of the most unique data sets
0:09:06 that other folks are talking about, you know
0:09:07 that may or may not have been scraped, you know
0:09:11 he’s case in point, the CTO of open AI
0:09:13 being asked by Joanna Stern about,
0:09:14 what does it exactly mean
0:09:16 that you train on publicly available data?
0:09:19 So all this to say, I’m still very bullish on Google
0:09:21 because I think they’ve got the most complete models
0:09:23 of the physical and digital world.
0:09:26 They’ve got ubiquitous distribution
0:09:28 and they’ve got the right infrastructure chops
0:09:31 to basically like bring that cost per query down
0:09:34 and an existing ecosystem of ads and sales
0:09:35 to plug monetization in.
0:09:39 So I think that there’s a like monetizable
0:09:41 sort of like answer engine model.
0:09:43 I think Google is one of the few companies
0:09:44 that could crack it.
0:09:46 That isn’t to say that I think like open AI
0:09:48 and Microsoft can’t take meaningful market share
0:09:50 but let’s be honest, how many of us actually use Bing?
0:09:51 Like I don’t, right?
0:09:53 Like I used it for a little bit
0:09:55 and I probably use perplexity more now.
0:09:58 Yeah, I mean, I am using chat GBT and perplexity
0:09:59 instead of Google a lot these days.
0:10:01 I agree, me too, to be honest, yeah.
0:10:03 And then, you know, I’ve been following,
0:10:04 you know, a long time ago I used to do SEO
0:10:06 like a long, long time ago
0:10:07 and I’ve been kind of following that space
0:10:09 and like in the last two months
0:10:11 and I kind of predicted this a year ago,
0:10:14 they’re like having major changes to the algorithm
0:10:16 where they’re really focusing on authority.
0:10:18 Yeah, domain authority, yeah.
0:10:20 Yeah, the reason they’re having to do that
0:10:23 is because of all the flood of AI content, right?
0:10:26 Like they just can’t deal with the flood of AI content.
0:10:28 So it’s like, okay, how do you deal with that?
0:10:31 Let’s go back to like really valuing the big brands
0:10:34 and the big names or the famous people too.
0:10:36 That’s the other thing is maybe like they’re taking
0:10:37 social signals, like you have a lot of followers
0:10:38 on social media.
0:10:40 Now that’s a signal that you’re a, you know,
0:10:42 an author they should listen to.
0:10:44 Yeah, I mean, why do you think we signed on with HubSpot?
0:10:47 We want that backlink domain authority.
0:10:48 That’s the only reason.
0:10:48 Yeah, yeah, yeah.
0:10:50 (laughing)
0:10:54 I mean, this is like, this is gonna be a meta problem though
0:10:55 for the industry, right?
0:10:57 The just the explosion of synthetic content.
0:10:59 I mean, and some social networks
0:11:01 are almost like incentivizing it.
0:11:03 Like, like on LinkedIn, it gives you,
0:11:06 it uses GPT-4 to suggest the comments.
0:11:08 And now you have just the most cringe like,
0:11:11 like regurgitations and summarizations
0:11:12 of like the original posts from like,
0:11:15 clearly a normal human being would never write this.
0:11:18 But like, if there was a meme about how to respond
0:11:20 to somebody on LinkedIn, I mean,
0:11:22 that’s like encapsulates the style I see.
0:11:22 And so.
0:11:23 But that’s happening on Quora too.
0:11:24 So like right now,
0:11:27 Quora is starting to rank at the top of Google results
0:11:29 and Quora is being dominated
0:11:31 with a chat GPT responses now.
0:11:33 100%, I mean, like books on Amazon too.
0:11:34 And I think the way they came out with a restriction
0:11:37 as well, you can only upload X number of books per day.
0:11:40 Like, I don’t know if that’s the solution.
0:11:42 But like, this is like the deep fake,
0:11:43 shallow fake problem too.
0:11:45 It’s like, everyone’s talking about detecting deep fakes.
0:11:46 And like, how do we figure this out?
0:11:50 It’s like, well, the thing that causes maximum harm today
0:11:51 are aren’t actually deep fakes.
0:11:53 They’re like super shallow fakes
0:11:55 where you take a photo from a different time or a context
0:11:58 and like, you know, kind of put it against another context.
0:11:59 This is exactly the type of stuff.
0:12:01 You’ll see like community noted on Twitter.
0:12:04 And that stuff’s like relatively easier to detect
0:12:06 ’cause you can actually find the source imagery
0:12:08 if you do like reverse image search.
0:12:10 And so like you add in the generative problem
0:12:12 on top of that, it’s like even crazier,
0:12:14 but like most platforms haven’t really even solved
0:12:16 the shallow fake problem, right?
0:12:18 It’s really like when it reaches a certain threshold
0:12:21 of distribution, there’s sort of this retroactive,
0:12:23 let’s go throttle this thing versus like,
0:12:25 how do you get ahead of this?
0:12:26 Anyway, I could talk about that forever
0:12:27 because some of the ways to avoid that is like,
0:12:30 ubiquitous surveillance, which is also not, you know.
0:12:32 Oh yeah, sounds great.
0:12:36 It’s like somehow the solutions to things that sound 1984
0:12:38 is like 1984 technology.
0:12:40 It’s like, it’s kind of weird how that works.
0:12:42 But it’s funny, somebody shared an article with me
0:12:45 just like the other day that was an article about,
0:12:47 well, not just me, but it was like an article
0:12:51 about like these seven AI influencers are, you know,
0:12:54 changing how we see AI or something like that, right?
0:12:56 And then like one of the seven like was my name
0:12:58 and I read the blurb about myself.
0:13:00 And the blurb about myself was like,
0:13:03 I grew up in Louisiana doing real estate and-
0:13:05 – You’re my neighbor.
0:13:08 – And like transitioned into computer programming
0:13:09 and then started teaching AI.
0:13:11 And I’m like, other than the fact that they’re like,
0:13:15 Matt makes content about AI, everything else about that
0:13:17 was just completely wrong.
0:13:18 – I think that’s where also like these models
0:13:21 need to be anchored in some sort of real knowledge graph,
0:13:24 you know, and like, that’s not to say that like, you know,
0:13:27 an approach like search is only gonna give you the truth,
0:13:28 right?
0:13:29 Like there’s like, what is even the truth?
0:13:31 And there’s like differing opinions on it.
0:13:33 But I think these models to be able to like,
0:13:34 just kind of fact check themselves,
0:13:36 at least with known information and come up with,
0:13:39 oh, at least the resources that are reputable are saying,
0:13:43 this is Matt’s bio would be better than, I don’t know.
0:13:46 So I’m like, this is the equivalent of like SEO dribble
0:13:50 that sort of started bleeding into Google around like 2019.
0:13:51 I think like, you know, it’s,
0:13:53 this is gonna be a huge problem.
0:13:55 And not to mention the implications of like,
0:13:57 if we’ve run out of content on the internet
0:14:01 and we are actively disincentivizing human generated content,
0:14:03 like how are we gonna train these models?
0:14:04 Like what’s gonna happen there?
0:14:05 – Yeah.
0:14:08 We’ll be right back.
0:14:11 But first I wanna tell you about another great podcast
0:14:12 you’re gonna wanna listen to.
0:14:15 It’s called “Science of Scaling” hosted by Mark Roberge.
0:14:18 And it’s brought to you by the HubSpot Podcast Network,
0:14:21 the audio destination for business professionals.
0:14:23 Each week host Mark Roberge,
0:14:26 founding chief revenue officer at HubSpot,
0:14:28 senior lecturer at Harvard Business School
0:14:30 and co-founder of Stage Two Capital,
0:14:33 sits down with the most successful sales leaders in tech
0:14:36 to learn the secrets, strategies and tactics
0:14:38 to scaling your company’s growth.
0:14:40 He recently did a great episode called
0:14:44 “How do you solve for a siloed marketing and sales?”
0:14:46 And I personally learned a lot from it.
0:14:47 You’re gonna wanna check out the podcast,
0:14:51 listen to “Science of Scaling” wherever you get your podcasts.
0:14:57 – I wanna talk real quick about search
0:14:59 because just to peek behind the curtain
0:15:02 to anybody who might be listening to this episode,
0:15:05 we’re actually recording it right before Google I/O, right?
0:15:07 Like Google I/O is next week
0:15:09 from the time we’re recording this.
0:15:12 Billavall and I are actually gonna be out at Google I/O
0:15:14 attending in person.
0:15:16 But one of the things that you mentioned
0:15:20 is that OpenAI and Microsoft have a tendency
0:15:23 that whenever Google announces something,
0:15:26 they need to jump in and like sort of one up them.
0:15:28 So by the time this episode comes out,
0:15:32 we’ll probably already know what OpenAI did with search.
0:15:35 But the rumor right now is that OpenAI
0:15:38 is creating some sort of their own search engine
0:15:41 with maybe with Microsoft involved, maybe not.
0:15:44 There’s still a lot of rumors and speculation flying around.
0:15:45 But knowing what you know about Google,
0:15:49 do you think OpenAI and chat GPT can come in and compete?
0:15:52 – I think no, like I think the search index Google has
0:15:53 is a very strong mode.
0:15:56 The fact that they can sort of almost like map the internet
0:16:00 in like almost real time is like just a hard technical
0:16:01 and infrastructure problem.
0:16:04 And they’re really well set up for it.
0:16:05 I’m curious what it is.
0:16:08 Like obviously this is complete pontification rumor mail.
0:16:11 Like what it is OpenAI is going to roll out.
0:16:14 What I think it’s gonna be is like something at parity
0:16:18 with perplexity with maybe a better like search index involved.
0:16:20 Like, and if it’s even if it’s something like that
0:16:23 where you get sort of this like multimodal summary
0:16:25 where it looks at a bunch of links, you get, you know,
0:16:27 some images, maybe some embedded videos
0:16:29 and like a summary of whatever it is you asked for
0:16:32 with citation so you can go validate sort of the quality
0:16:35 of those like the links that were summarized.
0:16:37 I think that would be a huge step up, right?
0:16:40 Like just getting, just being able to invoke like,
0:16:43 like search inside of chat GPT right now is clunky.
0:16:44 You have to be like, hey, well, and look this up,
0:16:48 research this and like explicitly prompted to do that
0:16:49 and being able to do that in a fashion
0:16:54 that is like really about leaning on like sort of a real time
0:16:58 and like sort of, you know, content that has real provenance
0:17:02 along with, you know, like sort of the distilled wisdom,
0:17:03 you know, the wisdom is debatable.
0:17:05 The distilled wisdom that is in,
0:17:07 in these large language models to summarize that,
0:17:09 I think it’s still a magical combination.
0:17:11 And I don’t know if y’all feel like this,
0:17:12 but I think the vibes have been shifting
0:17:14 with regards to just the conversation
0:17:18 and sort of Matt, I know you had this like post about like,
0:17:19 we think things are slowing down,
0:17:22 but here’s a bunch of announcement events coming up,
0:17:25 but doesn’t it feel like just like those leaps
0:17:28 that we hoped for haven’t quite come?
0:17:31 I guess Sora was kind of a leap.
0:17:34 I think the trend from like a million to 10 million
0:17:37 to maybe infinite context is like an interesting leap too,
0:17:40 but maybe we’re just getting too used to it, you know,
0:17:43 like versus the technologies,
0:17:45 like the pace of advancements slowing down.
0:17:46 Curious, which y’all think?
0:17:47 Yeah, I don’t know.
0:17:49 I feel like it’s kind of like it’s on social media,
0:17:52 the perception is, you know, the vibe has shifted,
0:17:53 but I feel like a lot of people
0:17:56 who actually are closer to Sam Altman,
0:17:57 they don’t feel that way.
0:18:00 And so I still believe that, you know,
0:18:01 they have something amazing coming.
0:18:02 – I follow Gary Marcus.
0:18:04 I know I think you actually have Gary Marcus
0:18:06 coming on the TED AI podcast,
0:18:08 so we’ll probably get some more insights from him
0:18:10 in the near future, but, you know,
0:18:12 the sort of Gary Marcus thing
0:18:15 that he’s been sort of all over Twitter about
0:18:20 is that we’re not seeing that same leap from GPT-4 to GPT,
0:18:25 or the same leap from GPT-2 that we saw the GPT-3.
0:18:27 He’s sort of arguing that that exponential curve
0:18:28 that everybody’s talking about
0:18:31 that we’re on with AI is not true, right?
0:18:32 We’re not on this exponential curve.
0:18:36 Otherwise, like, why didn’t we get from GPT-4 to GPT-5?
0:18:40 And half the time we got from GPT-2 to GPT-3, right?
0:18:42 Like, why is it not showing that?
0:18:45 But I also, like, I think my counter argument to that
0:18:46 is just ’cause we’re not seeing it,
0:18:47 doesn’t mean it’s not there, right?
0:18:49 – Yeah, totally.
0:18:51 – I think there’s a lot of stuff happening,
0:18:53 like Nathan sort of alluding to
0:18:57 behind the scenes over at OpenAI that we’re not seeing.
0:19:00 – I think it’s probably more related to, like,
0:19:02 compute requirements to actually use
0:19:05 some of these newer, more advanced models.
0:19:06 If they were to release it right now
0:19:08 with, like, the compute that’s available,
0:19:10 it would be, like, really expensive.
0:19:12 And, you know, the 20 bucks a month
0:19:14 that people are paying to use Jet GPT
0:19:16 is probably not gonna cover the cost of inference to run,
0:19:19 you know, these newer models, same thing with Sora.
0:19:21 – I mean, that’s kind of the leading theory.
0:19:24 I know that I believe with the GPT-2,
0:19:25 like, the mysterious model that came out
0:19:27 is, like, what the hell is that?
0:19:29 Like, maybe that’s what it is, is, like,
0:19:30 they’ve actually, maybe this is something
0:19:32 they actually developed, like, a year or two ago.
0:19:35 And it’s like a, it’s a more efficient architecture
0:19:36 or something like this.
0:19:38 And possibly that’s what GPT-5’s built upon,
0:19:40 then they, you know, they have less issues
0:19:42 with the cost issues in theory, but, yeah.
0:19:45 – I think it’d be hilarious if the GPT-2 stuff
0:19:46 has actually opened AI.
0:19:49 I mean, like, what an interesting way to sort of
0:19:51 test a model in the wild versus, I don’t know,
0:19:54 running some sort of A/B where, like,
0:19:57 there’s some sort of experiments on the chat GPT website
0:19:59 where a subset of users get a certain model
0:20:01 versus another set of users.
0:20:03 Maybe, like, making it explicit that there’s
0:20:04 these two versions of the models
0:20:07 and having people respond to it separately is interesting.
0:20:09 Maybe there’s an intention to create a bit of a PR,
0:20:11 like, sort of, like, you know, kind of,
0:20:14 a seed the conversation and kind of, like,
0:20:16 grease the wheels before the reel,
0:20:17 like, sort of, I don’t know,
0:20:19 race car jump moment happens or whatever.
0:20:22 I don’t know, but it’s like, yeah, like,
0:20:24 I mean, clients agree, like, there’s the compute stuff.
0:20:26 Certainly, Sora hasn’t been rolled out widely
0:20:29 because, like, it’s just so compute-intensive, right?
0:20:31 Like, you need to come up with a completely different,
0:20:34 like, pricing model way beyond the $20 one.
0:20:37 Maybe, hence, them talking to studios and stuff like that.
0:20:39 But all these models will get optimized, too.
0:20:40 To the Gary Marcus point, it’s interesting,
0:20:43 I had a conversation with him last week and it was like,
0:20:46 it’s like, he’s been very consistent about, like,
0:20:48 this not being the right paradigm.
0:20:51 And, you know, I think people liked it,
0:20:53 like, for the lack of, let’s just put it blunt.
0:20:55 Like, people like to shit on Gary a lot.
0:20:57 But, you know, if there’s one thing Gary’s been,
0:20:59 it’s like, it’s been exceedingly consistent.
0:21:01 And so, I don’t know, like, I would like to see
0:21:06 this sort of agentic co-pilot that feels more like an employee.
0:21:08 We certainly haven’t seen it yet, right?
0:21:11 The converse of everything else is the expectations
0:21:13 around AI, like, a year ago,
0:21:15 if we go back to when GPT-4 came out,
0:21:19 we’re just so fricking high that people just thought,
0:21:20 like, whether you were a knowledge worker
0:21:23 or a visual creator, you would look at the narrative
0:21:24 and you’d be like, holy crap,
0:21:26 like, this thing’s gonna take my job.
0:21:27 And then I don’t know if y’all saw the tweet,
0:21:29 I was like, and then you go use the tech
0:21:32 and it feels less like this, like, fricking kaiju Godzilla
0:21:34 that’s gonna stomp you and more like this, like,
0:21:36 chaotic golden retriever that you can kind of coax
0:21:39 to do cool stuff with you.
0:21:41 And I think that delta between expectations
0:21:44 and reality is so stark and, you know,
0:21:46 like, there was all of these,
0:21:48 I think even in a recent Sam Altman interview where he was
0:21:50 asking one of his biggest regrets is like,
0:21:53 GPT-4 didn’t have that, like, economic impact.
0:21:55 Everyone thought it would.
0:21:56 And he’s worried that the pendulum
0:21:58 will swing sort of the other way now,
0:22:00 where if expectations were so high,
0:22:02 people were like, oh yeah, whatever.
0:22:03 And so, I don’t know,
0:22:04 I think the answer is always gonna be in the middle,
0:22:07 but I can’t help but feel like we’ve gone past
0:22:09 the, like, peak of inflated expectations
0:22:12 and we’re going into the trough of disillusionment.
0:22:12 – Well, we’ll see.
0:22:14 I mean, like, Sam Altman also said that, like,
0:22:17 he was surprised that GPT-4 was so successful
0:22:19 and that, you know, it kind of sucks.
0:22:21 – I don’t know, my feeling on that is that Sam Altman
0:22:23 is one of the, like, greatest marketers
0:22:24 of our time right now.
0:22:27 And he’s, you know, he’s really, really, really good
0:22:29 at getting that hype wheel spinning.
0:22:30 – Oh, baby.
0:22:32 – Yeah, I swear, I think Sam Altman
0:22:34 is just really, really smart.
0:22:36 And I think, you know, when you see him speak, right,
0:22:39 when he does interviews, he’s very calculated, right?
0:22:41 He’ll ask a question and he’ll sit there
0:22:43 and he’ll usually pause for a good few seconds
0:22:45 before he responds.
0:22:48 And I think that he’s got that marketer brain.
0:22:49 Like, what can I say?
0:22:52 It’s going to sort of spread the flames
0:22:53 to hype this up a little bit more.
0:22:56 And I think that’s kind of how his brain operates.
0:22:58 So I think, you know, him saying like,
0:23:01 this is going to be the dumbest model
0:23:02 you’ve ever used by a lot.
0:23:04 And we’re going to, you know, we’re going to look at back
0:23:06 at this and be embarrassed by what we put out.
0:23:08 I think that’s all marketing.
0:23:09 – Well, I’m certainly excited.
0:23:11 But, you know, if I could have one request right now,
0:23:15 it’s like just give me GPT-4 from April, last year.
0:23:17 Give me that vintage of GPT-4.
0:23:18 It was better.
0:23:19 It was better, damn it.
0:23:21 – Can you still go into like the open AI playground
0:23:23 and like select the older models?
0:23:24 – I think you can, yeah.
0:23:25 And I feel like just all of these models,
0:23:27 especially in the consumer interface,
0:23:29 follow this trajectory of like when they launched,
0:23:30 they’re really good.
0:23:32 And then over time as like, you know,
0:23:35 various efforts to make sure the output is, you know,
0:23:38 on rails and not harmful kick in,
0:23:40 you see this deterioration happen.
0:23:43 But hey, that’s why we also have open source, right?
0:23:43 – Right.
0:23:44 – Yeah.
0:23:46 – Well, let’s talk about visual effects too,
0:23:48 because that’s like, that’s really your background
0:23:49 over at Google.
0:23:53 I want to go back to like the sort of, you know,
0:23:57 3D imaging 101 for a second here.
0:23:59 Can you sort of like break down the difference
0:24:01 between things like photogrammetry,
0:24:05 LiDAR, NERFs and, you know, Gaussian splats?
0:24:06 – I would love to.
0:24:07 In fact, the thing I was talking about is like,
0:24:10 everyone talks about like generative AI a lot,
0:24:12 but I think like the part that’s getting
0:24:16 not that much attention is this like visual spatial AI space.
0:24:19 And so think about spatial intelligence as like,
0:24:21 really just like a reality capture.
0:24:23 Like the world is in fricking 3D, right?
0:24:25 And so, you know, Matt, you nailed it.
0:24:28 It’s like basically photogrammetry is the art and science
0:24:31 of taking 2D images and other sensor data like LiDAR
0:24:34 and turning it into these 3D representations
0:24:35 of the real world.
0:24:37 So photogrammetry has been around
0:24:39 since like before computers were invented even.
0:24:43 Like this is a way of basically using like math and images
0:24:45 and observations of the world really
0:24:48 to like extract 3D structure from it.
0:24:50 But you should also think of spatial intelligence
0:24:53 as the ability for machines to sort of interpret
0:24:55 the spatial data like maps, 3D models,
0:24:57 like the world as we see it, right?
0:25:00 And so like, to me like photogrammetry
0:25:03 or like reality capture, all these other techniques
0:25:05 are all about recreating reality.
0:25:07 And so photogrammetry isn’t new as I alluded to, right?
0:25:09 But I think what’s gotten a huge boost
0:25:10 in why you hear about all these things
0:25:12 is like, thanks to machine learning,
0:25:15 basically like these learned approaches
0:25:17 to modeling the complexity of reality, right?
0:25:20 Like basically like, how do I take a bunch
0:25:23 of 2D images of the world like this
0:25:25 and essentially have a model do this
0:25:26 inverse rendering problem?
0:25:28 Or it’s like, oh, here’s where these hundred photos
0:25:29 are located in 3D space.
0:25:34 Based on this, I’m literally going to like eyeball ray tracing
0:25:37 like and create a 3D representation that makes sense.
0:25:41 And since you know exactly what the model looks like
0:25:42 at the photos that you’ve taken,
0:25:44 the representation that you get eventually
0:25:46 is like good enough from all viewpoints.
0:25:49 And so like, this basically the first nerve paper
0:25:52 dropped in 2021 called neural radiance fields.
0:25:54 And then there’s just been insane progress.
0:25:56 Like we talked about from like data centers
0:25:59 to like the GPU and your fricking like, you know
0:26:03 Nvidia workstation to like the iPhone in your pocket.
0:26:04 But even this wasn’t new.
0:26:06 There was like spiritual successures
0:26:09 to these like ML based learned representations
0:26:12 to sort of encapsulate the complexity of reality.
0:26:13 Enter radiance fields, right?
0:26:17 Like the way like, think about radiance fields generally
0:26:22 sort of like imagine like a voxel grid, a cube of cubes
0:26:25 where every single cube has like a color value
0:26:28 and an alpha like transparency value.
0:26:30 And like that’s kind of what you end up getting with a nerve.
0:26:32 And then when you do volume rendering,
0:26:35 you can basically like, you know
0:26:37 end up getting these photo realistic renditions
0:26:38 of the world.
0:26:41 And so like the cool part about neural radiance field is
0:26:43 instead of photogrammetry where you get this
0:26:46 like 3D mesh model, this like with surfaces
0:26:48 with textures plastered on it.
0:26:51 Think of it like crappy GTA looking models.
0:26:54 What you get with nerfs is like a radiance field
0:26:57 this voxel grid of like all these voxels
0:26:59 and their various values they’re in
0:27:02 that change based on how the camera is looking at it.
0:27:05 And because of that, you get all these things
0:27:06 that photogrammetry couldn’t do
0:27:10 which is like modeling transparency, translucency
0:27:14 like fricking like glass, like shiny objects
0:27:17 all this stuff can be done fricking fire,
0:27:19 volumetric effects, all the stuff that photogrammetry
0:27:21 can’t do cause imagine needing to come up with like
0:27:24 a cardboard paper mache model of that thing.
0:27:25 It’s gonna look like crap.
0:27:28 How do you model hair, fire, fog, all these things.
0:27:30 And you can do all of that
0:27:32 with these implicit representations.
0:27:35 Now the problem with nerfs were the rendering speed
0:27:37 because you’ve got this voxel grid
0:27:38 and you’re doing this volume rendering
0:27:40 where you’re like first doing like
0:27:41 the training process takes forever
0:27:43 but then when you want to render an image
0:27:44 you got to do volume rendering
0:27:47 and like trace like these rays through that voxel grid
0:27:51 and add up these like values like that takes a lot of time.
0:27:54 And basically it’s like think of it like one frame per second
0:27:57 to render out some of these videos, right?
0:27:58 Along comes Gaussian splatting
0:28:01 which is like, hey, do we even need the neural part
0:28:02 of radiance fields?
0:28:04 Like do we need like ML?
0:28:08 Can we just do this with like all school statistical techniques?
0:28:10 And like, which is kind of wild, right?
0:28:13 And so instead of having this implicit black box representation
0:28:17 where like realities modeled in the weights of this like MLP
0:28:19 this like multi-layer perceptron,
0:28:22 you’ve got this explicit representation
0:28:25 of these like ellipsoidal splat looking things
0:28:26 called Gaussians.
0:28:28 Just think of them like super stretchy like fricking spheres
0:28:33 like turns out you can get like a huge jump in quality
0:28:37 while also being able to render way, way faster.
0:28:40 And so like it’s like from one FPS
0:28:42 you’re getting like a hundred frames per second.
0:28:44 And since it’s an explicit representation
0:28:46 it’s in this like most formats
0:28:48 like all these apps that I’m showing on the screen
0:28:52 it’s in this format called PLY the Stanford PLY file
0:28:53 you can basically bring it
0:28:56 into any industry standard game engine.
0:28:57 Like you can bring it into Blender
0:29:00 you can bring it into Unreal into Unity.
0:29:03 And since it’s not this like like black box
0:29:05 like this neural network that you have to deal with
0:29:07 and it’s explicit you can go and delete
0:29:09 and edit things far more easily.
0:29:13 And so it’s like super crazy to see like what’s happened there
0:29:16 but basically between nerfs and Gaussian splatting
0:29:19 think of Gaussian splatting basically as radiance fields
0:29:21 without that neural rendering part.
0:29:23 And the paper uses terms like training or whatever
0:29:27 but there’s no neural networks involved at all in 3DGS.
0:29:29 So yeah, like how crazy is that?
0:29:30 We went from like cool.
0:29:32 Yeah, you could do like cool fly through videos
0:29:34 if you remember that was what the early days
0:29:36 of like the Luma app was do the scan
0:29:38 and now you can reanimate the camera
0:29:40 and you left this thing render for 20 minutes
0:29:41 and you got back something.
0:29:43 Now you can literally take your scans
0:29:45 and drop them into these real time environments.
0:29:47 And it’s like fricking amazing.
0:29:50 Like I think on the left I’m getting like 400 FPS
0:29:52 on an NVIDIA GPU and on the right
0:29:55 I’ve got this thing in Unreal Engine.
0:29:59 And what the cool part of it is like unlike photogrammetry
0:30:02 like and very similar to, you know, neural radiance fields
0:30:05 they still model these light transport effects.
0:30:07 So like again, like imagine if this was like
0:30:08 a cardboard cut out model
0:30:10 you wouldn’t have had all these light transport effects
0:30:13 of the light going through the tree, et cetera.
0:30:17 And so the way Gaussian splatting does this
0:30:20 is like by using this OG physics concept
0:30:23 called spherical harmonics to model it.
0:30:25 And so like if you’re trying to optimize stuff
0:30:27 you can get rid of some of these view dependent effects
0:30:29 as they’re called view dependent meaning
0:30:30 as you change your view
0:30:32 that like materials look slightly different
0:30:35 but you basically get it all with Gaussian splatting.
0:30:37 So I think it’s super exciting.
0:30:40 And yeah, like you could do this stuff in the cloud
0:30:43 you can do this stuff on your fricking desktop now.
0:30:45 Like I think post shot is a tool
0:30:46 that not many people have used
0:30:48 but like if you’re working on a commercial thing
0:30:50 and you don’t wanna upload your data
0:30:52 with Luma’s terms of service
0:30:53 or Polycam’s terms of service
0:30:58 like you can train this all locally on your desktop
0:31:00 with post shot, with Nerf Studio
0:31:03 the sum of the models Nerf Studio aren’t commercial friendly
0:31:04 and then even in the phone in your pocket, right?
0:31:06 Like so if you’ve got an iPhone like a modern iPhone
0:31:09 like and you just wanna know what 3D
0:31:12 like what radiance fields and reality captures all about
0:31:15 just download the Scaniverse app and like have at it.
0:31:17 So this is maybe a dumb question
0:31:19 but like so with like Nerfs and all this new tech
0:31:22 are you able to make like a really realistic 3D model
0:31:23 of like a city like San Francisco?
0:31:25 I mean, is that what you’re showing me earlier?
0:31:27 Or like, or is it only like a certain scene?
0:31:28 Like how hard is that?
0:31:30 Yeah, I mean, there’s a bunch of new papers outright
0:31:31 the initial radiance fields.
0:31:33 Like so there’s a paper called Block Nerf
0:31:35 that tries to scale Nerfs up to city scale
0:31:37 using Waymo data sets.
0:31:39 And similarly you’re seeing in the Gaussian splatting world
0:31:42 different papers about basically having like
0:31:46 in a kind of like nested hierarchies of splats
0:31:48 that have really good transitions to model
0:31:51 an entire like city and eventually the globe.
0:31:54 So I think like that’s the path that academia
0:31:55 and industry is on.
0:31:58 And I think like already you’re seeing city scale data sets
0:32:00 that are very plausible in research.
0:32:02 And I think it’s only matter of time
0:32:05 before that stuff gets into production.
0:32:06 You think that’s like the future of Google Maps?
0:32:08 I think it’s the future of maps for sure.
0:32:11 Like, you know, in immersive view
0:32:13 there are certain indoor locations
0:32:15 where you get a pre-rendered neural radiance field
0:32:18 that you can kind of like walk around and see.
0:32:20 This is just the evolution of that.
0:32:21 I think those data sets exist
0:32:23 and there’s like a handful of companies
0:32:25 in the world that have it.
0:32:27 So I think like that is the future of geospatial
0:32:29 and like maps in general.
0:32:31 But on the other hand, I think what’s interesting is like,
0:32:33 you know, this technology like building a map
0:32:35 of the world is easy.
0:32:36 Updating it is way harder, right?
0:32:39 Like when people talk about this like
0:32:42 one to one digital twin of reality, it’s like,
0:32:45 oh yeah, well like, by the way, new stuff
0:32:47 has built all the fricking time.
0:32:48 Things change all the time.
0:32:51 Seasonality is a fricking thing, right?
0:32:54 Like so, I think with this technology we’ve got
0:32:56 and since we’ve commoditized capture
0:32:59 because sensors are cheaper, computers cheaper.
0:33:02 And now we’ve got access to the same sort of algorithms
0:33:05 and approaches to model reality.
0:33:07 I think like updating this model of the world
0:33:08 is gonna get a lot, lot easier.
0:33:10 So I think like it’s gonna be very exciting
0:33:12 or in the near future we’re walking around
0:33:14 you know, driving around our cars
0:33:17 and walking around with our like glasses or whatever.
0:33:19 And we’re sort of updating this real time map of the world.
0:33:22 I think we’re very much on that trajectory
0:33:24 and we’re closer now than we’ve ever been.
0:33:26 – What do you think are like the business applications
0:33:27 to this tech?
0:33:28 – I mean, it’s like all the applications
0:33:30 that value stuff in the real world,
0:33:32 there’s utility and there’s delight, right?
0:33:35 I think like being able to not just like,
0:33:37 I mean, if you look at what NVIDIA is doing with Earth too,
0:33:39 right, we’re talking about the physical structure
0:33:39 of the world.
0:33:42 You can think of like the Earth having very facets, right?
0:33:44 Like there’s like the sort of the terrain,
0:33:48 like the natural like physical features of the Earth.
0:33:49 Then it’s all the like, you know,
0:33:51 human built things on top of that,
0:33:52 the structure that we built.
0:33:55 And then you can layer in like human activity
0:33:56 on top of that, right?
0:33:58 Like us moving around in the world,
0:34:00 our sensors, our cars, et cetera.
0:34:03 And then there’s other phenomena like weather, right?
0:34:04 Like tides and things like that
0:34:06 that need to be incorporated.
0:34:08 So Earth too is this really interesting initiative
0:34:11 by NVIDIA to focus on the weather like systems
0:34:13 that govern like, you know,
0:34:16 basically a day-to-day weather in the real world.
0:34:18 And so if you’ve got that understanding of like,
0:34:20 like the structure and geometry of a place
0:34:21 where the sun is going to be,
0:34:23 you can predict already things like,
0:34:25 hey, can I install solar panels here?
0:34:27 Like actually how much like sunlight would I get
0:34:30 if I installed this configuration of panels?
0:34:32 Then when you layer weather on top of that,
0:34:33 things get even more interesting.
0:34:35 So to answer your question,
0:34:37 I think there’s a bunch of applications
0:34:39 across utility and delight.
0:34:42 Like media and entertainment’s obviously in gaming is like,
0:34:45 I think the next GTA is absolutely going to be built
0:34:46 in like a twin of the real world.
0:34:49 Maybe this is the last GTA that we beat
0:34:51 will be built by humans manually
0:34:52 to emulate the real world.
0:34:55 I think that’s certainly exciting.
0:34:57 That said, a bunch of games have already used
0:34:58 reality capture, right?
0:35:00 Like from call of duty to battlefront, et cetera.
0:35:03 But I think the utilitarian aspects
0:35:05 are far, far more interesting.
0:35:07 Whether you’re like anything you’re trying to do
0:35:10 in the world of bits, like from like building stuff
0:35:12 to like disaster planning,
0:35:16 like the range of applications is just immense.
0:35:17 – Well, even just, you know,
0:35:21 one of the things that Jensen showed off at GTC this year
0:35:25 was to create these sort of virtual worlds
0:35:28 and then actually put virtual versions of like
0:35:31 humanoid robots in these worlds
0:35:34 and to sort of train them on this virtual twin
0:35:35 of the real world.
0:35:37 So they know how to navigate the real world
0:35:39 and then once they get that training data
0:35:41 then they can sort of inject that training data
0:35:43 into the real robots.
0:35:46 So like this concept of creating this digital twin
0:35:50 of the earth will allow us to train a lot of these robots
0:35:54 and machinery to operate within that digital twin
0:35:56 before actually deploying it in the real world.
0:35:59 To me, there’s a lot of huge implications there.
0:36:02 – 110%, I mean like these,
0:36:05 it’s like a way of creating all the training data
0:36:07 that these machines and perception models need
0:36:08 to be able to navigate the world, right?
0:36:10 And what better way it’s like,
0:36:12 you can create that like, you know,
0:36:14 you can 3D scan a city block
0:36:17 and then create all these different scenarios
0:36:19 of human activity on top of that
0:36:21 and feed that to, you know, like,
0:36:23 and train like self-driving cars,
0:36:25 like, you know, self-driving AI off of it.
0:36:28 I think like the fact that we’ve got a place
0:36:29 where we can basically like,
0:36:32 we can teleport reality into the digital world
0:36:34 and then also manifest the digital in the real world.
0:36:37 It’s like that bridge I think is just very powerful
0:36:38 for a bunch of different applications.
0:36:41 – Well, I want, let’s talk super quickly about TED.
0:36:45 So I’m, first of all, congrats on even having a TED talk.
0:36:47 Like that’s such an amazing accomplishment.
0:36:50 Like, you know, some people say they’ve had a TED talk.
0:36:51 They’re really talking about a TEDx talk.
0:36:54 And come on, come on, come on.
0:36:57 You’ve actually given a real TED talk, a legit TED talk.
0:36:58 And not only that,
0:37:01 but they asked you to host the TED AI podcast.
0:37:03 So tell us a little bit about that
0:37:04 and what’s going on there.
0:37:06 I mean, maybe share a little bit of your experience
0:37:09 with TED and then tell us about the TED AI podcast.
0:37:10 – Yeah, sure.
0:37:14 I mean, certainly the TED talk was a fun experience last year.
0:37:16 And I would say this year was even more fun.
0:37:18 It’s a side of the opportunity.
0:37:19 It’s a co-session too,
0:37:22 which was all of that AI with Chris Anderson.
0:37:24 And we had some amazing speakers,
0:37:28 like Vinod Kosla, Feifei, you know,
0:37:31 the CEO of GitHub, Helen Toner,
0:37:33 export member, OpenAI.
0:37:36 And even like, I don’t know if you’ve,
0:37:38 if you’ve checked her work out, but nice aunties,
0:37:42 like absolute trip down basically like,
0:37:46 to me what intergalactic social media looks like.
0:37:48 There was a super, super fun experience.
0:37:49 Yeah, I mean, like, look,
0:37:51 the T and TED is all about technology.
0:37:54 And I think right now what’s exciting is to put the,
0:37:58 like over time, TED grew to encompass, you know,
0:38:00 not just technology, entertainment and design,
0:38:03 but a plurality of topics, right?
0:38:05 And I think with AI sort of in technology,
0:38:06 being sort of this horizontal,
0:38:08 like tech is a horizontal,
0:38:11 but it’s impacting so many different verticals
0:38:13 in our daily lives, right?
0:38:15 Like we can talk about all the applications,
0:38:16 whether you’re a creator,
0:38:18 whether you’re a knowledge worker, you know,
0:38:20 whether you’re a musician, you know,
0:38:22 whether you’re thinking about like national security
0:38:25 and defense, whether you’re thinking about relationships.
0:38:30 And often in all of these sort of topics, you know,
0:38:34 there’s like a dichotomy that we as like builders
0:38:36 and consumers have to contend with.
0:38:40 And so the idea of the TED AI show really is to outline
0:38:42 those dichotomies and, you know,
0:38:45 not necessarily take an opinion one way or the other,
0:38:49 but sort of elaborate on the entire gamut of like,
0:38:51 like the good, bad and the ugly
0:38:53 and sort of let people decide for themselves
0:38:56 and do that by talking to people from all walks of life.
0:38:58 Like people whose titles haven’t even been invented yet,
0:39:01 but obviously technologists, journalists, researchers,
0:39:03 artists, you know, the list goes on
0:39:05 and, you know, I’m just super grateful
0:39:08 for the opportunity to be able to, you know,
0:39:10 just bring my excitement into the space.
0:39:12 Like obviously like I want to build,
0:39:14 bring the lens of a creative,
0:39:16 like that’s like built a following
0:39:18 for over a million folks using these tools,
0:39:20 but also as a product builder who shipped a bunch
0:39:22 of this stuff and then just like,
0:39:26 I would say like cautiously optimistic AI enthusiasts.
0:39:28 So I’m going into a bunch of these topics
0:39:30 with those three lenses in mind
0:39:32 and it’s just been a lot of fun.
0:39:35 We’ve got some really cool episodes lined up for y’all
0:39:38 and I can’t wait for y’all to check it out.
0:39:41 – You have any idea of like, you know, launch schedule?
0:39:44 Like is that, are there dates planned out for it yet?
0:39:45 – Yeah, yeah, totally.
0:39:47 So May 21st, first episode drops
0:39:48 and then it’s going to be weekly.
0:39:50 There’s going to be a little bit of a summer break there,
0:39:52 but yeah, 25 episodes in the season.
0:39:55 And let me tell you,
0:39:57 I think there’s something for everyone.
0:39:59 – Yeah, I do wonder like what was the general vibe at Ted?
0:40:03 Like are people optimistic or are they like really fearful
0:40:04 of AI?
0:40:05 – And if you look at Ted also, you know,
0:40:08 you had people like, like the laval.
0:40:10 I know there were some other speakers there.
0:40:11 I think maybe Moustafa Suleiman was there.
0:40:13 Maybe that was a more recent one.
0:40:15 But then you also had guys like Gary Marcus
0:40:20 and I’m going to totally butcher his name, Yadkowski.
0:40:22 – Oh yeah, it is Yadkowski.
0:40:24 – Which are both more on the like,
0:40:26 hey, let’s chill out on the AI side.
0:40:29 So it seems like it’s from a speaker’s front.
0:40:31 It seemed like they had speakers on both sides
0:40:33 of the arguments.
0:40:33 – Definitely.
0:40:35 I mean, the theme for this year was like the brave
0:40:36 and the brilliant and like covering
0:40:38 that gamut of opinions.
0:40:40 I would say overall the vibe is positive.
0:40:42 So I like, I’ll give you a sample size.
0:40:44 Like I taught this discovery session,
0:40:48 which was about the dichotomies of AI is about 50 people.
0:40:49 And sort of what we did is we looked at a bunch
0:40:52 of these verticals in the AI space.
0:40:54 And like, you know, essentially came up with like,
0:40:56 what happens if this goes really well
0:40:59 and what happens if this goes really poorly?
0:41:01 And like, let’s use chat GPT actually
0:41:02 to come up with like a headline,
0:41:06 like a piffy visual depiction of that desirable
0:41:08 and undesirable future.
0:41:10 And honestly, like most folks in the room
0:41:12 are optimistic about it, right?
0:41:14 Like, but they’re not blind to the downsides.
0:41:17 I think like the problem with anything is extremes, right?
0:41:19 And so like, you know, where like Nathan,
0:41:22 I hear your concern of like, if we’re like, you know,
0:41:23 oh, like we can’t have like,
0:41:27 demerism and like, you know, I was like, it’s infectious.
0:41:30 I think the same thing applies to the opposite narrative too,
0:41:32 which is like, well, we just obviously,
0:41:34 we got to keep like accelerating.
0:41:35 We got to keep shipping.
0:41:36 I think it depends, right?
0:41:39 Like, it’s sort of the boring answer to these things.
0:41:42 And I think you can’t understand the nuances
0:41:45 unless you go dissect like the full gamut
0:41:47 of like considerations.
0:41:50 And so the goal is like, really like the editorial perspective
0:41:51 I’m trying to bring.
0:41:55 And of course, Ted has a huge say in this too is like,
0:41:58 look, I’m a 60% optimistic, 70% optimistic,
0:41:59 which is not too dissimilar.
0:42:01 I think Matt, well, you and I have talked about
0:42:03 on most things, but I’m not going to be blind
0:42:06 to all like the downsides of this stuff too, right?
0:42:09 Like, and I think that’s okay to say.
0:42:12 And what I believe is just like the speaker selection,
0:42:15 you know, we’re trying really hard
0:42:17 to have a balanced perspective on the guests too.
0:42:20 So that like, what’s what I think is going to happen
0:42:22 in the real world as well as like,
0:42:24 you’ll be able to hear both sides of that argument.
0:42:27 Like somebody who’s like super stoked and thrilled
0:42:30 on like AI art and just thinks it’s the bee’s knees.
0:42:33 And it’s totally cool to train on copyrighted material.
0:42:35 Somebody who thinks like that is the death of creativity
0:42:36 as we know it too.
0:42:37 And you have to-
0:42:38 – I think it’s good to hear both sides.
0:42:40 I mean, I agree, but I’m concerned.
0:42:41 Like I used to live in San Francisco
0:42:43 and like they’re like pushing for regulation now.
0:42:44 We’re like to-
0:42:45 – Who’s they?
0:42:46 You mean, you mean, Sam Allman?
0:42:47 – The government, yeah.
0:42:48 Well, no, the government.
0:42:49 They’re like, they’re like pushing a bill through right now.
0:42:51 They’re trying to like, they’re trying to fast track it.
0:42:53 I forgot who’s doing the bill,
0:42:55 but we’re basically,
0:42:57 if you launch a new large language model-
0:42:58 – Need approval.
0:43:00 – Yeah, you like need approval,
0:43:02 but also like it doesn’t get to basically sign something
0:43:03 and like it’s perjury if you’re lying
0:43:05 that this model can do no harm.
0:43:06 And it’s like-
0:43:07 – Yeah, like who’s gonna do that?
0:43:08 Yeah.
0:43:09 – Yeah.
0:43:12 And so I agree like nuance is important and like,
0:43:15 I do consider myself kind of part of the EAC movement,
0:43:16 but more generally just like a techno optimist.
0:43:18 – Do you have it in your bio still?
0:43:19 I don’t know.
0:43:20 – I don’t, I don’t.
0:43:21 But I, you know, I like Beth.
0:43:23 I like all the people who are part of that.
0:43:24 – That’s cool.
0:43:25 – I think in general,
0:43:27 I think in general it’s right, you know,
0:43:29 but nuance is important, I agree.
0:43:30 – Let me put it this way.
0:43:31 I think the way I see it is like,
0:43:34 we’ve got enough talented humans out there
0:43:37 that are like pushing for like acceleration
0:43:39 and there’s enough talented people out there
0:43:41 that are pushing for, you know,
0:43:43 I would say pumping the brakes
0:43:46 for the lack of a better way to put it in certain areas.
0:43:48 And I think like in totality,
0:43:52 like we’ll reach like some optimum solution
0:43:53 because of those influences.
0:43:54 And I think it’s always been like that, right?
0:43:58 Like, I mean, just like the early days of music,
0:44:00 everyone’s like, oh yeah, Napster and like peer-to-peer
0:44:01 and let’s just go crazy.
0:44:03 And then things settled down
0:44:05 and we found a business model that worked.
0:44:07 Maybe it’s not perfect, right?
0:44:09 Like people have a lot of gripes with the Apple
0:44:10 and the Spotify business model,
0:44:13 but I think we found this like globally optimum solution.
0:44:15 – I mean, it doesn’t always work out, right?
0:44:16 Though they look at like nuclear plants, right?
0:44:18 Like in the past, the US was gonna build
0:44:21 all these nuclear plants to solve energy problems
0:44:23 and we didn’t do it because of regulations
0:44:24 and because of fear.
0:44:25 And now we’re like trying to solve
0:44:26 all these global warming problems.
0:44:29 Other thing is like, well, we always kind of had nuclear there
0:44:31 that we could have been using and it worked.
0:44:32 So it doesn’t always work out.
0:44:34 Like it often does, but it doesn’t always.
0:44:37 – Totally, I mean, one of the funnest topics
0:44:42 is getting into how do politicians and regulators
0:44:46 even regulate this like sort of nebulous set of technologies.
0:44:48 It’s not just large language models, right?
0:44:50 Like there’s all the perception AI stuff
0:44:52 and like the implications there.
0:44:55 And but like a group of these technologies
0:44:57 that are like sort of permeate and go, my God,
0:45:00 like some of the stuff that I’m like doing research
0:45:04 right now on just the intersection of neuroscience and AI
0:45:07 and what we’re gonna be able to do with just like passive
0:45:10 neural like interfaces with earbuds and things like that.
0:45:14 I mean, there’s gonna be some real big ethical quandaries
0:45:15 that pop up.
0:45:18 And so yeah, yeah, trying really hard to bring like
0:45:19 still be techno-optimist,
0:45:21 but bring that balanced perspective.
0:45:23 And I think folks are gonna like it.
0:45:24 – I think at the end of the day,
0:45:26 the important thing is empathy, right?
0:45:30 Like the perspective I come from is I tend to be
0:45:31 a very empathetic person, right?
0:45:34 I want to hear both sides of the story.
0:45:35 I want to hear both perspectives
0:45:38 and I want to be empathetic to both sides too.
0:45:40 Like if somebody is genuinely worried
0:45:43 that this technology is taking their job,
0:45:44 I want to understand why.
0:45:48 I want to understand what we can do to sort of mitigate
0:45:50 the damages that could be done from this.
0:45:53 Like I’m always gonna come from that place of empathy,
0:45:55 which is why I’ve never sort of identified
0:45:58 with the EAC movement is like,
0:46:00 I don’t necessarily think we should always be pushing
0:46:03 everything forward as fast as possible.
0:46:05 I think we should be listening to the fears.
0:46:06 We should be listening to the concerns.
0:46:09 We should be figuring out sort of middle grounds.
0:46:13 Like you mentioned, there’s always people on both sides
0:46:15 which kind of creates a decent checks and balance
0:46:17 to make sure one side doesn’t go too far
0:46:20 and AI nukes the world,
0:46:22 but also the other side doesn’t go too far
0:46:25 and technology stops advancing completely.
0:46:29 Those checks and balances I think are a net positive overall.
0:46:30 I think those need to be there.
0:46:33 – I agree overall, but I mean, I think the big thing
0:46:36 that was like, the argument with EAC would be making is
0:46:39 that compounding is one of the most important powers
0:46:40 out there, right?
0:46:41 Like an idea that if we build faster
0:46:43 than the technology in the future will be better
0:46:44 and better and better and we’ll start solving
0:46:46 real world problems like cancer
0:46:48 and all these other things that maybe we could have been doing
0:46:52 if we weren’t so like quick to just regulate everything.
0:46:54 And so I think with AI, it’s the same thing.
0:46:56 Like sure, like some regulation in the future
0:46:58 might make sense, but if we just start throwing it
0:47:00 out there right now, we’re gonna slow down the compounding
0:47:03 and like the exponential, we’ll stop the exponential
0:47:05 from happening with our regulations.
0:47:08 And we, you know, yeah, maybe some jobs would be lost
0:47:10 in the short term, but in the long term,
0:47:13 we could have cured cancer, we could have solved, you know
0:47:15 global warming issues, all kinds of other problems
0:47:18 that we could have solved if we would just like weighted
0:47:19 and see what happens with the technology.
0:47:20 And yeah, now there’s a big problem.
0:47:23 Okay, maybe make a regulation, but just don’t do it
0:47:23 like right at the beginning,
0:47:24 like they’re trying to do right now.
0:47:27 – I mean, India totally flipped their decision, right?
0:47:29 Initially they were like, oh, you have to get every model
0:47:31 approved and they’re like, actually we’re gonna retract
0:47:33 this part, which was really interesting.
0:47:36 And I mean, like it’s like a regulation also,
0:47:38 like it could end up in a place where it just only benefits
0:47:41 the incumbent, like the largest AI labs too, right?
0:47:44 Like the regulatory capture like point of view,
0:47:47 it’s like it could end up in a place where like any new
0:47:50 innovation like from a startup can actually happen.
0:47:52 And they’re the ones that get these like owners compliance
0:47:55 requirements and they can’t afford a team of lawyers
0:47:57 unless they’re like super VC backed.
0:47:59 And then like, what an inefficient use of VC capital
0:48:02 instead of innovating, you’re sort of like navigating
0:48:05 like the legal landscape of like a heavily regulated
0:48:05 industry.
0:48:07 So I think like your point about nuclear is well taken too.
0:48:10 It’s like, I mean, like a lot of folks, I mean,
0:48:12 including Gary brings up the example of like, you know,
0:48:16 like basically air travel and airplanes are still well
0:48:19 regulated, but yet we’ve got this whole Boeing fiasco
0:48:20 happening, right?
0:48:23 It’s like where you’ve got one really big incumbent
0:48:25 and there’s probably a revolving door between regulatory
0:48:29 agencies and Boeing and, you know, so it’s-
0:48:30 – And when the Wright brothers got started, I mean,
0:48:31 they weren’t being heavily regulated as they were
0:48:33 like inventing the plane.
0:48:34 – Totally.
0:48:35 – They were out there just like out in Ohio,
0:48:37 just, you know, trying shit out.
0:48:37 So I mean-
0:48:39 – I think one thing it’s like a lot of people talk about
0:48:42 is like just like technology, sunsets, like regulations
0:48:45 need to sunset rather than us adding up more and more
0:48:46 regulations.
0:48:48 So I think this is where things get geopolitical too.
0:48:51 It’s like, I think China is so much savvier about AI
0:48:53 regulation than the US is right now.
0:48:57 And I feel for the politicians, I think they’re like asking
0:48:59 for this type of engagement.
0:49:01 And I think it’d be good if we engage with them on this
0:49:05 and like bring those perspectives to bear rather than,
0:49:07 I don’t know, just being like a regulation, bad,
0:49:10 innovation, good, let’s keep innovating.
0:49:13 And so it’s like, it’s nuanced, but then again, like,
0:49:18 look, I’m a, I’ve always been a Libra and like, you know,
0:49:21 kind of trying to build bridges between two worlds.
0:49:25 – Yeah, no, I mean, I think this being a super nuanced
0:49:28 conversation is like the understatement of the episode here.
0:49:31 I think there’s just so many different like rabbit holes
0:49:33 that we could potentially go down when it comes to the
0:49:34 regulation thing.
0:49:37 I think you’re going to have to be one of our sort of
0:49:38 recurring guests.
0:49:41 Maybe every few months jump on and nerd out about this stuff.
0:49:44 But, you know, I do want to give you the opportunity to
0:49:46 tell us what else you’re working on.
0:49:48 If there’s a place you think people should go check you
0:49:51 out, your Twitter, your YouTube, obviously the TED AI
0:49:53 podcast coming out later in May.
0:49:56 – Yeah, so just please follow me on Twitter @belavosadoo.
0:50:01 You can also follow me on YouTube and TikTok @billyfx.
0:50:03 If you’re interested in some more long form expositions
0:50:05 that I do, check out the Creative Tech Digest.
0:50:08 It’s both a newsletter as well as a YouTube channel.
0:50:10 And yeah, of course, check out the TED AI show.
0:50:14 Maybe the one last thing I’ll say is like if you’re a
0:50:16 founder and a builder in this space building with any of
0:50:19 the technologies that we talked about and you’re looking
0:50:21 for early stage investment, I’m also a scout for
0:50:23 A16Z Games.
0:50:26 So just hit me up on Twitter or you can email me.
0:50:28 We’ll put the email in the show notes as well.
0:50:30 And I really appreciate you guys having me on.
0:50:32 I’m wishing you all the success for your podcast.
0:50:36 And Matt, I will see you at I/O.
0:50:39 And Nathan, I hope to see you in 3D sometime soon.
0:50:42 – Oh, come out the Kyoto, come on.
0:50:43 – I gotta make it happen.
0:50:44 – Awesome, Bill of All.
0:50:45 Well, it’s been a blast.
0:50:47 This has been one of my favorite conversations we’ve had
0:50:47 so far, so fun.
0:50:50 Excited to see you in person next week.
0:50:51 – Cool, cheers.
0:50:53 (upbeat music)
0:50:56 (upbeat music)
0:50:59 (upbeat music)
0:51:01 (upbeat music)
0:51:04 (upbeat music)
0:51:06 [MUSIC]
Episode 8: Is Google’s dominance in search engines at risk with the rise of generative AI models? Hosts Matt Wolfe (https://x.com/mreflow) and Nathan Lands (https://x.com/NathanLands) dive into this topic with guest Bilawal Sidhu (https://x.com/bilawalsidhu), a host of The TED AI Show and a former Google employee experienced in AR/VR projects and creating 3D maps for Google Maps.
In this episode, Bilawal explores the potential challenges facing Google’s search engine supremacy due to advancements in generative AI models and discusses the implications for the future of search engines and advertising. He dives into the impact of AI-generated content on search results and the need for a nuanced approach in navigating the evolving landscape of digital information.
Check out The Next Wave YouTube Channel if you want to see Matt and Nathan on screen: https://lnk.to/thenextwavepd
—
Show Notes:
- (00:00) Decade in tech, six years at Google.
- (03:17) Time at Google as a product manager .
- (06:27) Discussion about impact of TikTok and YouTube.
- (12:28) Models need to be anchored in real knowledge.
- (18:23) Testing GPT-2 in the wild with users.
- (19:37) AI expectations not met yet.
- (23:29) Spatial intelligence for machines and photogrammetry summary.
- (26:02) Implicit representations enable flexible, accurate rendering techniques.
- (32:08) Nvidia’s Earth 2 initiative predicts real-world weather.
- (34:41) Training AI by creating virtual reality scenarios.
- (36:33) Exciting Ted show explores impact of AI technology.
- (40:15) Editorial perspective in Ted show seeks balanced, nuanced viewpoints.
- (45:12) Delay regulations to accelerate technological advancement impactfully.
- (47:18) Regulations should sunset, engage with China’s perspective.
—
Mentions:
- Get HubSpot’s Free AI-Powered Sales Hub: enhance support, retention, and revenue all in one place https://clickhubspot.com/gvx
—
Check Out Matt’s Stuff:
• Future Tools – https://futuretools.beehiiv.com/
• Blog – https://www.mattwolfe.com/
• YouTube- https://www.youtube.com/@mreflow
—
Check Out Nathan’s Stuff:
- Newsletter: https://news.lore.com/
- Blog – https://lore.com/
The Next Wave is a HubSpot Original Podcast // Brought to you by The HubSpot Podcast Network // Production by Darren Clarke // Editing by Ezra Bakker Trupiano