“It’s like ChatGPT and Wikipedia have a baby”, said Aravind Srinivas, CEO and cofounder of Perplexity, said of his search engine at Imagination In Action’s ‘Forging the Future of Business with AI’ Summit in April 2024. On this panel, Srinivas talks about building a better search, the future of AI for enterprises, and why getting a PHD and starting a business are similar.
Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1
Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:
https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript
Stay Connected
Forbes newsletters: https://newsletters.editorial.forbes.com
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com
Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.
Subscribe to FORBES: https://www.youtube.com/user/Forbes?sub_confirmation=1
Fuel your success with Forbes. Gain unlimited access to premium journalism, including breaking news, groundbreaking in-depth reported stories, daily digests and more. Plus, members get a front-row seat at members-only events with leading thinkers and doers, access to premium video that can help you get ahead, an ad-light experience, early access to select products including NFT drops and more:
https://account.forbes.com/membership/?utm_source=youtube&utm_medium=display&utm_campaign=growth_non-sub_paid_subscribe_ytdescript
Stay Connected
Forbes newsletters: https://newsletters.editorial.forbes.com
Forbes on Facebook: http://fb.com/forbes
Forbes Video on Twitter: http://www.twitter.com/forbes
Forbes Video on Instagram: http://instagram.com/forbes
More From Forbes: http://forbes.com
Forbes covers the intersection of entrepreneurship, wealth, technology, business and lifestyle with a focus on people and success.
Category
🤖
TechTranscript
00:00So, perplexity, how many people use perplexity?
00:04Is it on your phone right now?
00:06That's gotta make you feel good.
00:07Thank you.
00:07Yeah.
00:09So, you don't every day meet somebody
00:11that's unicorn status in less than two years.
00:14And that is definitely a sign of the times in AI,
00:17so congratulations on that.
00:19Would love to hear your version of the founder journey,
00:21because it's been a very, very short timeline,
00:23and your stepping stones in there include Google,
00:26and OpenAI, and now your startup,
00:28and then before that, Berkeley.
00:30Just compress that down to your journey
00:32to get to where you are today.
00:34Can everyone hear me?
00:35Okay.
00:37Firstly, you said it right, AI,
00:40the whole AI is basically running on exponential time,
00:44so the regular time feels logarithmic,
00:46relative to AI, the speed at which we're going.
00:51He's asking us to sit down.
00:52All right, stay awake.
00:55So, yeah, it feels amazing.
00:58Exhilarating, stressful, all at the same time.
01:02There's this phrase, uncomfortably exciting,
01:05where you feel the adrenaline all the time,
01:08and you feel like you'll run out of business any day,
01:11but at the same time, you feel like
01:12your business is growing really fast.
01:13So that's what it feels like.
01:16Inflection points every single day.
01:18Wow.
01:19So, tons going on in the tech stack.
01:22Do you want to?
01:22Yeah, so I'll start with a quick story
01:25of my journey with perplexity.
01:28I had a crazy surgery just a few months ago,
01:31and I was sitting there in the hospital
01:33just before and after the surgery,
01:35and I could barely see my screen,
01:38and the only thing I could use was perplexity,
01:41because first of all, I'm one of those people
01:43who's always worried about,
01:44doctors are telling me, is this true?
01:46And if these are my blood reports, what does it mean?
01:48And so I was using perplexity all the time.
01:51Nice, large fonts, giving me a five-sentence summary
01:54of what I need to know.
01:55So I encourage everybody to look at perplexity.
01:57So magical aspect about perplexity,
02:00for those of you who have tried,
02:01is the summarization, less hallucination
02:04with links out there.
02:06So if you get a little bit more technical here,
02:09you're using underlying large language models right now,
02:12but what do you see going forward?
02:13I mean, are we going to see Transformer
02:16continue to win out over the next two years?
02:18Yeah, Lacoon is going to talk about joint embeddings
02:21with JEPA, of course, the progress in RAG.
02:23I know you have certain type of RAGs
02:25and fine-tuning that happens on the fly.
02:27Tell us a little bit about what's under the hood
02:29and what allows you to keep this mode
02:32ahead of everybody else.
02:34Right, so for perplexity, we're not bound
02:38to the Transformer or any model.
02:41It's more a software system
02:44that takes an existing language model,
02:45the state-of-the-art language model,
02:48state-of-the-art RAG system,
02:49and orchestrates the search and the LLM together
02:54into an answer engine, right?
02:55LLMs, large language models, whether it's a Transformer
02:58or state-space model or any other architecture,
03:01these are models that are like amazing reasoning engines.
03:05You can prompt them, you can instruct them,
03:08you can basically program them to do amazing things.
03:11Basically, you want to harness the reasoning power
03:14in these engines with facts that exist
03:18in the form of regular search indexes
03:20that have been built over the last two decades
03:22and combine these two powers
03:25and create a new entity called the answer engine.
03:29It's like ChatGPT and Wikipedia having a baby together,
03:34and that is perplexity.
03:35Why?
03:36Because a regular ChatGPT allows you to chat
03:38with these really general reasoning model.
03:42It feels amazing, but it doesn't always get things right.
03:45And Wikipedia has almost all the accurate facts
03:48in the world, extremely well peer-reviewed,
03:51a lot of citations, references you will trust
03:54when you read it, and you can get to know
03:57about anything, about any topic,
03:58except it's not personalized to you.
04:01The same Wikipedia page is written for everybody.
04:03So you might be reading a Wikipedia page
04:05about the Oppenheimer movie, and you might only be interested
04:07in certain things there, like about the actor.
04:10Someone else might be interested about how it was filmed.
04:13You don't have to read through the whole thing.
04:14It's just not the best experience.
04:16How would it look like if we combined the two together
04:19and bring the world's knowledge
04:20and give it to you in a personalized,
04:22conversational manner at your fingertips everywhere?
04:25That's what perplexity is.
04:27It's amazing.
04:27Maybe a little bit more about what's under the hood then.
04:30Clearly you're still using somebody else's APIs.
04:33That's right.
04:34So the user doesn't care whether you use your own models
04:39or use OpenAI's or Anthropics or Mistral's or Meta's.
04:43Like really, they don't care.
04:45When they come and use the product,
04:46they just want good answers,
04:48good personalized, fast speed, accurate answers.
04:52So our model is not our moat.
04:55Our moat is how we orchestrate many different models
04:58in one single product, lower the costs
05:02without compromising on accuracy,
05:04and make sure that the answers are accurate as possible.
05:09So lower the hallucinations,
05:11make sure that the latency is really good,
05:13and make sure the answers are presented
05:15in a readable manner.
05:17We can still do a lot of things here.
05:19We are not there yet.
05:20Nobody wants to read a big chunk of paragraphs
05:24written for any question you ask.
05:26So you have to go above and beyond
05:28to understand the user intent
05:29and present the answers differently
05:31according to the type of query.
05:32If you can nail all these three things together,
05:34which cannot be done with just the one giant model
05:37to do everything, but rather really understand
05:41how to work for different type of queries
05:43and really understand what the answer type should be
05:46and really make sure that you can use
05:48a family, an orchestra of models.
05:51One model to understand what type of query it is.
05:55One model to reformulate and expand a query
05:57into multiple queries.
06:00Go and collect pages relevant to each of them.
06:02One model to aggregate and summarize
06:04all the content you picked.
06:06One model to just chunk all the documents
06:08into different embeddings.
06:09And another model to suggest next questions to ask you.
06:13Now all of these are working in parallel
06:16at the same time when you ask a question on perplexity.
06:18And you don't care.
06:19That's the thing, you don't care what happens
06:21under the hood.
06:21When you're using the product, it should just work.
06:23It should be fast and it should feel good.
06:26And that's what, there's the same thing with,
06:29there's a whole famous Jeff Bezos video, right?
06:31Like, internet or smart internet doesn't matter.
06:33What matters is the customer experience.
06:35Like, same thing here.
06:37GPT or Claude or Lama or Mistral doesn't matter.
06:42What matters is the answer is accurate and readable.
06:45So I think that's great advice,
06:46especially for all the people doing startups.
06:47And there are a lot of them around here, as you would know.
06:51One specific, just follow up on that.
06:52If somebody was starting a company today,
06:54do you see a year or two from today perplexity
06:57or these other startups operating
06:59on open source foundation models
07:01or calling APIs like GPT-4?
07:04I would definitely, even if,
07:06it's not really a dichotomy.
07:09If you start a company and you wanna build out a product
07:12and get it in the hands of people,
07:14do the easiest possible thing.
07:15Like, take someone else's API,
07:18whether it's GPT-3.5 or Haiku
07:20or the latest Mistral model that offers an API
07:24by any of these API providers, doesn't matter.
07:27Just go take it and get the product out
07:29in the hands of people,
07:30whatever costs you the minimal thing,
07:33and collect a lot of data,
07:34see where the problems of failing and not failing,
07:38and then go and collect the data set
07:40that addresses these failures.
07:42And then you have to go back to open source
07:44because that's what will let you
07:45really controllably fine-tune things.
07:48There's only so much you can do with prompt engineering
07:51and the most capable, like GPT-4,
07:53or Opus-like models are really amazing
07:55that just prompt engineering gets you the solution,
07:58but that's still pretty expensive.
08:00We are all supposed to build businesses here, right?
08:03And to do that, you have to be profitable at some point,
08:05so you cannot be spending so much
08:07on these really large frontier model providers.
08:11So you have to go back to open source
08:12and try to specialize the experience
08:15to what your product is supposed to do,
08:17and that lets you achieve better cost efficiency.
08:20That's Mark Gorenberg, the chairman of MIT,
08:22and as soon as he said that, he started taking notes,
08:24so you know what you're saying is worth remembering.
08:27Thank you.
08:28So on that note, I mean, if you,
08:30you know, the difference between open source
08:32and open weight and open APIs,
08:36and if you start thinking about not B2C, but B2B,
08:39and especially, you know, creating startups
08:41or solutions for enterprises,
08:44whether it's on-prem or, you know,
08:45kind of privacy-preserving, how do you see that?
08:47I know you right now have a B2C product,
08:50but companies like Anthropic are really targeting,
08:53you know, B2B products.
08:54Yeah, I mean, like, we will also,
08:56like, first of all, a lot of people tell us
08:58that we love using the Pro,
09:00Perplexity Pro, the subscription plan,
09:03but we are banned from using it at work.
09:07Microsoft was to ban Perplexity
09:09for all their employees today,
09:12for, I mean, because they want everyone
09:13to use their products.
09:15It's not a new thing.
09:16They've done that for Google in the past,
09:19but in general, like, there is, like,
09:21a skepticism on every employer
09:23to let people use these AI answer bots,
09:26the chat bots at work,
09:27even though they know that their employees
09:29massively benefit from it in terms of productivity,
09:32time saved, how fast they can get work done,
09:35all of these things.
09:37So any answer bot in the market today,
09:41if you want it adopted in the hands of as many people,
09:44especially when the value provided
09:45is a lot more at work time than personal time,
09:49you should figure out a way to be adopted
09:50in the enterprise,
09:51and that requires you to work on compliance,
09:55security, data retention,
09:58and making sure that you can even have
10:00self-hosted versions of your product
10:02on the, you know, the client that they have.
10:06So all these things should be worked on,
10:07and we will also work on it.
10:09And then there's the other part of, like,
10:11you know, whether the foundation model provider,
10:13if you are not the provider yourself,
10:15you have the right SLAs with them
10:17to make sure they delete the data periodically,
10:19and both Anthropic and OpenAI provide this,
10:22and if you are gonna have your own models
10:24and you're gonna let others use it,
10:25you should also provide this.
10:27So these are the current ways in which we're working.
10:29I think fine-tuning, when you're taking
10:31someone else's model and fine-tuning on your own data,
10:34that is even higher level of, like, requirements.
10:37Like, you do need to make sure
10:38that data warehouses are secure.
10:40This is where companies like Databricks or Snowflake
10:42are pretty well-positioned,
10:43because they get the full end-to-end package
10:46to the customer.
10:47We're not gonna work on that market yet.
10:50So can I go back to your founder journey
10:51for just one second here?
10:54You know, your business plan,
10:55your company's going like this,
10:56your business plan is very, very aggressive,
10:59and I don't usually associate that with PhD backgrounds.
11:02You kind of associate that with people
11:04who drop out after freshman year.
11:05So what is this about you and your journey
11:07that got you on this trajectory?
11:08Because there is a version of the world,
11:11which I think you very likely believe in,
11:13where people don't do much Google searching
11:16three years from today.
11:17They just talk to perplexity.
11:19So, I mean, first, going back to PhD,
11:22I just want to mention one thing here.
11:24I applied to MIT and Berkeley.
11:27Berkeley admitted me, MIT rejected me.
11:29And my mom, I went and told my mom,
11:35hey, I got into Berkeley.
11:37She's like, what is Berkeley?
11:40Did you get into MIT?
11:41I was like, no, I didn't get into MIT.
11:43And she was so disappointed.
11:45And I tell her, hey, Berkeley's also a good college.
11:47And I don't even know what Berkeley is.
11:49I only know MIT.
11:50So that's the extent to which MIT's popular.
11:53And I'm very happy to be at MIT for the first time.
11:55And as a professor, I always tell the students here,
11:59although there are many examples of people dropping out
12:00and starting Microsoft and Facebook and so on,
12:04be cool, stay in school, and follow.
12:08Because if you see the current AI trend,
12:09most of the leaders have finished their PhDs.
12:12That's true, that's true.
12:13So, yeah, going back, what does PhD prepare you for?
12:16I think it prepares you for taking risks.
12:20That's a common feature between doing a PhD
12:23and doing a startup.
12:25When you're a new student,
12:26there's lots of low-hanging fruit ideas
12:29that your senior students in your own lab
12:31have ready-made for you
12:33because they don't have time to work on it
12:34and they want somebody else to help them.
12:36And you can get four or five papers published
12:39in one year if you're just literally following
12:42the instructions of your professor
12:44or your senior grad student.
12:46And generally, all the advisors want their students
12:49to do that, to feel productive,
12:51to feel like they got some wins,
12:53and get settled into the new place,
12:54and then take the risk.
12:56That's just the established wisdom in grad school.
12:59I also did that.
13:00The first paper I wrote was just someone else's idea,
13:02not mine.
13:03But I didn't like it, I was not enjoying it,
13:05and felt like, okay, this is the time to take risks.
13:08Like, why am I even here?
13:09I would rather go and work in Google Brain
13:11or some other research org if I'm not taking risks.
13:15And the part of taking risks in PhD
13:18is your advisor doesn't advise you anymore.
13:21He's like, yeah, yeah, go do your thing,
13:22and if it works out, we can consolidate
13:26and figure out a paper.
13:28So for a period of eight months,
13:30when I was an intern at OpenAI in 2018,
13:34my PhD started in RL.
13:37And I saw this guy, Alec Radford write GPT-1,
13:41the first version of GPT.
13:43It's amazing to think this was just six years ago.
13:47But at that time, everybody ridiculed it,
13:49nobody even understood it.
13:52It's like, oh, it's just like yet another autocomplete.
13:55But some of us knew that this was real.
13:57And so I had to go back to my advisor and say,
14:00we have to like, you know this whole thing,
14:02if you're in crypto, pivot to AI.
14:04So that was a moment for us.
14:06It's like, if you're in RL, pivot to generative models.
14:09It was called generative models, not generative AI.
14:11So, and I spent like eight months just not having any paper.
14:16But it was truly worth it.
14:17It helped me shape fundamentals in deep learning,
14:21generative models, unsupervised learning.
14:22And that's what helps me to do all this today.
14:25So I think PhD generally teaches you
14:30to take a lot of risks, be aggressive.
14:33There's no, like the whole shoot for the fences,
14:37it's not a truncated outcome distribution.
14:40You can change your life.
14:41So that open AI experience is a perfect segue
14:44into my next question.
14:45So you're working with GPT-1.
14:47One of our partners actually wrote a book with GPT-2
14:49and it was terrible, but it was obvious to her
14:53that the fact that it could write a book at all
14:55was indicative of something that a year later
14:57would be massive.
14:59So I think your experience with GPT-1
15:01must have given you early insight
15:03into where things are going.
15:04Now, the next iteration is multimodal.
15:07And that's only a year away
15:08and we're gonna be shocked yet again.
15:09What's perplexity gonna look like
15:11on those foundation models?
15:12Yeah, so, I mean, the first idea of perplexity,
15:15I pitched to our first investor, Elad Gil,
15:19who's an MIT alumnus, was, you know, you wear your glass.
15:24The best way to disrupt Google
15:25is ask questions through a glass.
15:27And then it has a microphone,
15:29it understands what he's saying,
15:31and then gives you an answer.
15:33And this way people don't have to type into a search bar
15:35and they get answers read out to them on their AirPods.
15:38I think this will happen, right?
15:40Whether we do it or somebody else does it, doesn't matter.
15:43It's inevitable that search will be in the form of like,
15:47on the go everywhere,
15:49hence device-free as much as possible, voice-based,
15:53and using vision inputs, not just text.
15:57And I think that we wanna work on that too.
16:00That's what I think true multimodal search is.
16:03And then like asking questions anywhere.
16:07Like, imagine you can talk to any device, ideally, right?
16:10Like, that doesn't have to be just your phone.
16:12And you ask questions.
16:14You go to a shopping mall
16:15and like there's just like near any row of dresses,
16:19you can just ask questions to some device there,
16:22about anything.
16:23It's almost like a concierge.
16:24A bunch of concierge everywhere.
16:26I think all these things are possible,
16:28or you can point your finger at something and ask questions.
16:30Like we have to reimagine hardware
16:33along with all these like new models.
16:35And that's what I'm excited about.
16:37Like how perplexing can play a role there.
16:40So this kind of a Tony Stark, Iron Man, also MIT guy,
16:46the Tony Stark vision of being able to interact,
16:48augment your world.
16:50This does require, I mean, you described perplexity
16:54as really an orchestration engine
16:56that brings all these models together.
16:58Once it's multimodal, what's in the physical,
17:01in the dimensional world,
17:02it's only going to get more exciting.
17:04A lot of that will be dependent on hardware.
17:07Yeah.
17:09And on one end, we're seeing Blackwell
17:11and trillions of transistors.
17:14On the other hand, if it's mobile, augmented,
17:18it's going to be edge computing.
17:21So given that, so just your thoughts
17:23on kind of the scaling laws in a trillion parameters,
17:27a hundred trillion parameters.
17:29Also share your thoughts on companies
17:30like Cerberus and Grok.
17:33What's the general landscape for compute and hardware?
17:36Yeah, I think like the NVIDIA hardware
17:40is going to be the best for training.
17:43Nobody can compete there at all.
17:47Now for inference, that's where it gets more subtle.
17:50Right now, it turns out that even though you think
17:53H100 is a bigger beast than A100s,
17:55and it's likely better for training,
17:56but for inference, you can use cheaper GPUs.
17:59It turns out not to be the case.
18:01It turns out that you can pack more density per chip
18:05because the chips are more powerful.
18:07So instead of buying like 64 GPUs,
18:09a hundred GPUs for serving more users,
18:11you can buy like 24 GPUs of H100s
18:15instead of like 128 A100s
18:17and serve the same number of users
18:18because you get higher throughput
18:21with fewer chips basically.
18:23And you spend lower too.
18:25So it helps you to scale in terms of server racks better
18:28to use more powerful chips and fewer of them
18:31if you want to do a device on the cloud.
18:33But if you want to do inference on the device itself,
18:38that's where I think NVIDIA
18:40doesn't actually necessarily play a big role.
18:42And they have like a lot of edge GPUs too,
18:45and Google has edge GPUs.
18:46Apple has probably the best hardware for on-device.
18:49I think like you have to use all,
18:51like at some point you have to build custom compilers
18:53if you're building your own hardware there.
18:56It's not clear to me
18:57what the software consolidation there is gonna be.
18:59I don't see CUDA playing a massive,
19:02having a massive mode for on-device inference.
19:05So that's where I think some people can build new companies.
19:09As for Grok, I think what they're doing is very exciting.
19:12Even if they're not replacing NVIDIA for inference,
19:18they're definitely gonna bring the prices down.
19:20Or else like margins for NVIDIA is just gonna be so high.
19:24Nobody can compete there.
19:25So I hope that it brings inference prices down
19:28when there's more competition.
19:30And that's great for everybody
19:32because we can scale our products to more users
19:34without having to pay as much.
19:36I mean, some of these hardware companies
19:38are four to five X better than NVIDIA
19:41in the price performance,
19:43training and even more on inference.
19:46So do you see this kind of one dominant player emerging,
19:50has emerged in training?
19:52Do you see dominant players emerging
19:54in inference or edge devices?
19:57I mean, you mentioned Apple, but setting that aside.
19:59Yeah, so if you decouple inference into cloud inference
20:03and on-device inference,
20:04there's not gonna be one dominant player for on-device.
20:07But for cloud inference, even today,
20:09I would say NVIDIA is the dominant player.
20:12And for other people to compete there,
20:13they need to move really, really fast on the software
20:16that helps people to deploy on their hardware.
20:19And that's what is lacking today.
20:21Like everybody still works on NVIDIA software.
20:23They use TensorRT for like inference.
20:26And NVIDIA is like moving really fast.
20:29That's why they are our investor too.
20:31We work with them on TensorRT.
20:32We work with them on fine tuning infrastructure
20:34like Nemotron.
20:36And they just like moving really fast.
20:37So whoever wants to compete there
20:39should also move equally fast on the software stack too.
20:42Yeah.
20:43So we only have time for one more question
20:44according to John there.
20:45And I really have to ask you this
20:46because I was looking at your LinkedIn profile.
20:48So your founder journey makes it seem like PhD,
20:51open AI, brilliant idea, perplexity.
20:54But somewhere in there, you ended up seed investing
20:57in Mistral, Cognition Labs, Grok, 11 Labs.
21:00There's gotta be more than a story there.
21:02How did that happen?
21:03I mean, someone on Twitter told,
21:04I just invest in hype companies.
21:06Actually, that's what I do.
21:07I don't have any time to like evaluate all of them.
21:11They sometimes they ask me for investing
21:13and if the products are good, it feels good to use them.
21:18I just invest like there's no time to evaluate.
21:21Oh, like what's the business model?
21:22What's your mode?
21:23Because if somebody did that to us, to me, or perplexity,
21:27it would be ridiculed very fast.
21:29Like, oh, how are you ever gonna compete
21:30with open AI or Google?
21:31Like for every startup, you can come up
21:34with like thousands of reasons of why they wouldn't work.
21:38And so like, if you only focus on that,
21:40you're never gonna be able to invest
21:42in any of these companies.
21:43So rather you would focus on, okay,
21:45like what does the success outcome look like
21:48in case it works?
21:49And if that seems pretty big,
21:51and the downside is not that high either,
21:55you should just go for it.
21:56These are asymmetric bets.
21:58So-
21:59That makes it sound like all these insanely awesome startups
22:02came to you.
22:02Did it all happen in the last 18 months
22:04that you received-
22:05Yeah, we share an investor with Levin Labs, Nat Friedman.
22:10He's one of the most popular AI investors.
22:14And so he connected us, Matty and me,
22:17and we are customers of Levin Labs.
22:21Perplexity's voice features, our podcast,
22:24Discover Daily Podcast is all powered by Levin Labs voice.
22:27And we work very closely with them on voice-to-voice.
22:31For Grok, we got connected through Chamath.
22:34He's an investor in Grok.
22:37So I think like Mistral was similar.
22:39Like, you know, I knew some of their founders before,
22:41and also like we use their models,
22:45like a lot of people here.
22:47So there's like some synergy
22:48between what we do at Perplexity.
22:49If Grok works out, we'll do inferencing on Grok.
22:52We use Mistral's models.
22:54We use Levin Labs' software.
22:55So there's a lot of synergy between like my investments
22:58and how Perplexity can make use
23:00of these other companies' services or products.
23:02Really, I just want to point out how mind-blowing.
23:05All of what you just said is all in the last 20 months.
23:08That's right.
23:09The founding of the company, the unicorn valuation,
23:11the investments, everything you just described,
23:14that is completely unprecedented in world history
23:17to be moving through life at a pace like that.
23:21Just to end, I know we have 30 seconds left.
23:25What would you share as top three sectors
23:27or even top three startup ideas
23:28that folks should be thinking about here?
23:30I think on-device would be really cool.
23:33I'm learning more about it myself.
23:36Somebody was telling me yesterday
23:38that if they were to start a company today,
23:41they would really, really work
23:43on a seven billion parameter model
23:46that's as good as Cloud Sonnet
23:48or somewhere like 3.5 to four,
23:53GPT 3.5 to four in that spectrum capability.
23:57And that model could let you control the phone,
24:02be able to control the OS.
24:04And that's the sort of company
24:06that you would expect OpenAI or Anthropic
24:10to not overly invest in
24:12because they want to go after the most general,
24:15extremely capable reasoning model
24:17that lives on the cloud and can do everything.
24:19But this is the sort of model
24:20that will help you build a lot of value for every customer.
24:24Every consumer can just use it.
24:27Some of you may have seen MKBHD's criticism
24:29of humane chip, humane pin.
24:32While it was harsh, there's one major point
24:34that the latency is very poor
24:36and that can be only addressed
24:37if the models are living on-device.
24:40Fantastic.
24:41Absolutely an honor, pleasure.
24:44Thank you.
24:45Thank you for coming.