WEBVTT - Perplexity CEO Bloo Talks AI Boom

0:00:02.520 --> 0:00:07.040
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio news.

0:00:07.760 --> 0:00:10.879
<v Speaker 2>It's not just chip makers pitching the furniture of computing.

0:00:10.880 --> 0:00:14.800
<v Speaker 2>At computex Perplexity took the stage with Intel to unveil

0:00:14.840 --> 0:00:18.599
<v Speaker 2>what it calls the world's first hybrid local server AGENTIC

0:00:18.720 --> 0:00:22.880
<v Speaker 2>inference orchestrator, a phrase that sounds like it was generated

0:00:22.920 --> 0:00:26.920
<v Speaker 2>by AI itself. Here with more Perplexity, CEO arab In

0:00:26.960 --> 0:00:29.480
<v Speaker 2>shrinaves Aravin is great to have you back on the show.

0:00:29.600 --> 0:00:32.560
<v Speaker 2>I spent all morning thinking about how do I explain this,

0:00:33.120 --> 0:00:36.879
<v Speaker 2>and basically, the orchestrator is there. It's a piece of

0:00:36.920 --> 0:00:40.600
<v Speaker 2>software to decide whether a or a part of an

0:00:40.640 --> 0:00:44.000
<v Speaker 2>AI workload is best done locally on device at the edge,

0:00:44.600 --> 0:00:48.160
<v Speaker 2>or if it needs the superior computing of cloud server.

0:00:48.880 --> 0:00:50.839
<v Speaker 2>Is that right? Have I kind of nailed what you're

0:00:50.880 --> 0:00:52.400
<v Speaker 2>what you're trying to solve for here?

0:00:53.479 --> 0:00:56.200
<v Speaker 3>That's correct? So yeah, thank you for having me again.

0:00:56.400 --> 0:01:01.279
<v Speaker 3>And that is exactly correct. You don't want all your

0:01:01.400 --> 0:01:07.600
<v Speaker 3>compute centralized in gigantic servers and everything running through the

0:01:07.720 --> 0:01:12.440
<v Speaker 3>largest frontier models. You're already reading reports of how people

0:01:12.440 --> 0:01:15.000
<v Speaker 3>are freaking out about their token costs. Some people are

0:01:15.040 --> 0:01:20.040
<v Speaker 3>spending half a billion dollars per month per engineer. What

0:01:20.160 --> 0:01:24.640
<v Speaker 3>you actually want is efficient token value per WoT per user,

0:01:25.640 --> 0:01:32.640
<v Speaker 3>and that requires orchestrating privacy, accuracy, intelligence and costs all

0:01:32.760 --> 0:01:38.080
<v Speaker 3>together in one single unified system. And that orchestration capability

0:01:38.680 --> 0:01:44.280
<v Speaker 3>requires hybrid model between server side and the local and

0:01:44.280 --> 0:01:47.000
<v Speaker 3>that's what we demoed today with Intel. And we are

0:01:47.040 --> 0:01:50.960
<v Speaker 3>actually chip agnostic, So our solution works with Intel, it

0:01:51.000 --> 0:01:54.840
<v Speaker 3>works with NVIDRTX. So just like how we've been model agnostic,

0:01:54.920 --> 0:01:56.559
<v Speaker 3>we've planned to be chip agnostic here.

0:01:57.440 --> 0:01:59.760
<v Speaker 2>That's interesting. So my next question was going to be

0:02:00.040 --> 0:02:02.920
<v Speaker 2>why Intel? You know, what is it about Intel's role

0:02:03.040 --> 0:02:06.000
<v Speaker 2>in the AIPC market and on the service side that

0:02:06.040 --> 0:02:09.000
<v Speaker 2>makes it work. But if you're agnostic, what's the breakthrough

0:02:09.040 --> 0:02:11.800
<v Speaker 2>that you've cracked, Like, if you've written the software, what

0:02:11.880 --> 0:02:14.080
<v Speaker 2>is it you've managed to achieve in how those streatloads

0:02:14.080 --> 0:02:16.560
<v Speaker 2>are diverted? Okay, go go a bit further.

0:02:18.680 --> 0:02:22.640
<v Speaker 3>So, like I said, you want one single system to

0:02:22.800 --> 0:02:28.519
<v Speaker 3>route across models, files, tools, chips, servers and decide when

0:02:28.560 --> 0:02:31.560
<v Speaker 3>to use which model or when to use your local

0:02:31.600 --> 0:02:35.400
<v Speaker 3>file system, your local subasion model, your local LM, or

0:02:35.440 --> 0:02:37.560
<v Speaker 3>when to use a frontier model for depending on the

0:02:37.639 --> 0:02:40.720
<v Speaker 3>task and the prompt or depending on the confidentiality and

0:02:40.760 --> 0:02:44.720
<v Speaker 3>sensitivity if your files are apps that requires you to

0:02:44.840 --> 0:02:50.040
<v Speaker 3>make clever orchestration decisions, balance trade offs between accuracy and costs.

0:02:50.760 --> 0:02:54.000
<v Speaker 3>And that's what we're doing in our software and that

0:02:54.240 --> 0:02:57.720
<v Speaker 3>computer essentially an operating system and balances all these different

0:02:57.720 --> 0:02:59.440
<v Speaker 3>objectives simultaneously.

0:02:59.560 --> 0:03:03.079
<v Speaker 1>I mean sit in such an interesting place as an orchestrator,

0:03:03.120 --> 0:03:06.720
<v Speaker 1>whether it be letting people use your own house models,

0:03:06.720 --> 0:03:09.160
<v Speaker 1>whether it's using a mix of third party models, and

0:03:09.200 --> 0:03:11.720
<v Speaker 1>the third party models are up to or not right now,

0:03:11.880 --> 0:03:13.359
<v Speaker 1>I don't hen I just want to get your take

0:03:13.400 --> 0:03:18.440
<v Speaker 1>on how you feel about competitive modes or competitive threats

0:03:18.720 --> 0:03:21.639
<v Speaker 1>if these big companies Anthropic, Open AI, SpaceX ll go

0:03:21.720 --> 0:03:23.200
<v Speaker 1>public in the next few months.

0:03:23.840 --> 0:03:28.560
<v Speaker 3>We actually love Andthropic, Open AI, XAI, all these frontier labs.

0:03:29.360 --> 0:03:34.160
<v Speaker 3>Every time any of their AI gets better, our unified

0:03:34.360 --> 0:03:38.320
<v Speaker 3>system also gets better because we route across all of them.

0:03:38.480 --> 0:03:42.640
<v Speaker 3>We basically think of Perplexity computer as taking the best

0:03:42.640 --> 0:03:45.040
<v Speaker 3>of all AI and putting it together in one single

0:03:45.120 --> 0:03:49.000
<v Speaker 3>unified interface and system. So all of you know how

0:03:49.080 --> 0:03:52.040
<v Speaker 3>much Anthropics models have improved since the beginning of the year.

0:03:53.200 --> 0:03:56.160
<v Speaker 3>What has it led to for us, our revenue actually

0:03:56.240 --> 0:03:58.400
<v Speaker 3>triples since the beginning of the year. It's just been

0:03:59.040 --> 0:04:01.720
<v Speaker 3>five months in the year and our revenue is already

0:04:01.720 --> 0:04:06.520
<v Speaker 3>tripled to what that's So we're actually like very happy

0:04:06.560 --> 0:04:10.240
<v Speaker 3>with all these companies's progress and they completely deserve their IPOs,

0:04:10.320 --> 0:04:11.920
<v Speaker 3>so we're very excited for them.

0:04:12.160 --> 0:04:14.800
<v Speaker 1>Can I follow up to are you able to discuss

0:04:14.920 --> 0:04:17.159
<v Speaker 1>what that revenues jumped to? There were reports in the

0:04:17.200 --> 0:04:18.880
<v Speaker 1>ft that you're up to about four hundred and fifty

0:04:18.920 --> 0:04:19.720
<v Speaker 1>million dollars just in.

0:04:19.640 --> 0:04:23.279
<v Speaker 3>The and March. Yeah, we crossed that. I think I

0:04:23.560 --> 0:04:26.640
<v Speaker 3>publicly tweeted that we cross five hundred million about some

0:04:26.640 --> 0:04:30.279
<v Speaker 3>somewhere around mid April. We are announcing new numbers yet,

0:04:30.320 --> 0:04:31.640
<v Speaker 3>but we're doing really well.

0:04:32.880 --> 0:04:36.679
<v Speaker 2>Irvin. I've been thinking a lot about where perplexity sits

0:04:37.160 --> 0:04:41.560
<v Speaker 2>in the suite of available tools and technologies. Right research

0:04:41.640 --> 0:04:45.720
<v Speaker 2>seems to be a really interesting place with perplexity, And

0:04:45.960 --> 0:04:49.719
<v Speaker 2>I'm wondering, like how you measure the engagement on the platform,

0:04:49.760 --> 0:04:52.640
<v Speaker 2>so like it's not just like one query and done,

0:04:52.680 --> 0:04:55.320
<v Speaker 2>but do you kind of track the time that an

0:04:55.360 --> 0:04:58.760
<v Speaker 2>individual desk or user would stick with one query as

0:04:58.800 --> 0:05:02.440
<v Speaker 2>sort of indication of success, you know, how the platform

0:05:02.520 --> 0:05:04.799
<v Speaker 2>is being used, the behaviors of the userbase.

0:05:05.880 --> 0:05:09.920
<v Speaker 3>So we're actually not trying to maximize engagement per user

0:05:10.000 --> 0:05:11.800
<v Speaker 3>in the sense we're not actually trying to keep them

0:05:11.839 --> 0:05:14.960
<v Speaker 3>longer on the platform or something. In fact, like accurate

0:05:15.160 --> 0:05:18.960
<v Speaker 3>accuracy is somewhat like towards the opposite end of that,

0:05:19.040 --> 0:05:21.359
<v Speaker 3>like if you give the user an accurate answer in

0:05:21.400 --> 0:05:23.479
<v Speaker 3>the first turn, it's likely that they're not going to

0:05:23.480 --> 0:05:26.479
<v Speaker 3>continue in the same chat. What we actually look at

0:05:26.600 --> 0:05:30.320
<v Speaker 3>is like retentive uses, like if the same user is

0:05:30.440 --> 0:05:36.240
<v Speaker 3>using Perplexity for a lot more research tasks, not just

0:05:36.240 --> 0:05:38.719
<v Speaker 3>like that one single task that came with and that's

0:05:38.839 --> 0:05:41.800
<v Speaker 3>that's always been the case. For example, we introduce a

0:05:41.839 --> 0:05:45.120
<v Speaker 3>max plan, and that's already like you know, at the

0:05:45.160 --> 0:05:48.039
<v Speaker 3>beginning of the year, in terms of subscriptions, split between

0:05:48.360 --> 0:05:50.400
<v Speaker 3>the max plan that is a two hundred dollars month

0:05:50.440 --> 0:05:53.760
<v Speaker 3>plan versus the pro plan was somethad like nine is

0:05:53.800 --> 0:05:57.800
<v Speaker 3>to ninety one. Today it's more like thirties to seventy.

0:05:58.040 --> 0:06:00.000
<v Speaker 3>So I think that that already shows that they are

0:06:00.080 --> 0:06:03.080
<v Speaker 3>these power users who are willing to pay like two

0:06:03.120 --> 0:06:05.120
<v Speaker 3>thousand dollars a year out of pocket. This is not

0:06:05.160 --> 0:06:10.200
<v Speaker 3>even enterprise because they allow these supreior research and orchestration

0:06:10.279 --> 0:06:12.360
<v Speaker 3>and accuracy that we bring in our product.

0:06:12.440 --> 0:06:16.039
<v Speaker 1>Interesting, So we're seeing the growth you're talking about your

0:06:16.920 --> 0:06:20.599
<v Speaker 1>average revenue run rate right tripling, going up to almost

0:06:20.640 --> 0:06:23.279
<v Speaker 1>five hour million dollars. I'm interested as to where the

0:06:23.360 --> 0:06:25.200
<v Speaker 1>combative nature does come in because it looks like you're

0:06:25.200 --> 0:06:26.920
<v Speaker 1>playing well with all the other players out there. But

0:06:27.200 --> 0:06:30.000
<v Speaker 1>there are some issues in the courts in particular, CNN,

0:06:30.040 --> 0:06:32.039
<v Speaker 1>for example, has just the latest to hit you with

0:06:32.080 --> 0:06:36.440
<v Speaker 1>a lawsuit alledging that you violated federal copyright laws. How

0:06:36.480 --> 0:06:39.080
<v Speaker 1>are you dealing with how people get paid and what

0:06:39.120 --> 0:06:42.000
<v Speaker 1>you train upon and what you feed and source to

0:06:42.120 --> 0:06:42.880
<v Speaker 1>us as a user.

0:06:44.240 --> 0:06:46.040
<v Speaker 3>I mean, the fact of the matter is that, like,

0:06:46.120 --> 0:06:49.680
<v Speaker 3>nobody has any copyright over truth and facts. Like I

0:06:49.720 --> 0:06:53.400
<v Speaker 3>think we've being consistent with our position. We're very confident

0:06:53.440 --> 0:06:56.600
<v Speaker 3>in our position, and we will let the legal process,

0:06:57.080 --> 0:06:59.080
<v Speaker 3>you know, decide what the right thing is in that

0:06:59.120 --> 0:07:01.440
<v Speaker 3>particular situation. I don't want to comment further on that,

0:07:02.000 --> 0:07:04.960
<v Speaker 3>but nobody has any copyright over truth in facts.

0:07:05.800 --> 0:07:09.159
<v Speaker 1>Perplexity CEO staying up late. It is like eleven thirty

0:07:09.160 --> 0:07:12.720
<v Speaker 1>pm with you. We so appreciate you coming by after

0:07:12.760 --> 0:07:15.880
<v Speaker 1>your Yeah jet lag works worldwide. Our of industry and

0:07:15.880 --> 0:07:18.960
<v Speaker 1>of ours safe flight back from Taiwan, we appreciate it.