[00:00] Chad Nantais: Here's one example.
[00:02] Chad Nantais: I asked Gemini,
[00:03] Chad Nantais: give me three quotes about whatever topic it was.
[00:07] Chad Nantais: Let's say about human suffering from the Indian philosopher,
[00:10] Chad Nantais: Jiddu Krishnamurti.
[00:12] Chad Nantais: It gave me three quotes in quotations.
[00:15] Chad Nantais: I looked at those quotes and there was something fishy about them.
[00:18] Chad Nantais: And then I asked Gemini,
[00:19] Chad Nantais: are those actual quotes from talks that Krishnamurti did?
[00:23] Chad Nantais: And Gemini said,
[00:25] Chad Nantais: oh,
[00:25] Chad Nantais: I'm glad you asked.
[00:26] Chad Nantais: Actually,
[00:26] Chad Nantais: they're not actual verbatim quotes.
[00:29] Chad Nantais: they're quotes that he could have said based on you know his philosophical principles
[00:34] Paul Salazar: chad there you are let's get started welcome to episode one
[00:41] Chad Nantais: it's been a busy last couple of weeks in terms of uh existential crises related to ai there's the former employee
[00:50] Chad Nantais: of Anthropic who resigned uh Jacob Coxon he's been doing the rounds on talk shows about why he left
[00:58] Chad Nantais: over concerns about the dangerous approach that the frontier AI labs are taking towards developing AI.
[01:05] Chad Nantais: I saw a clip of him on Jon Stewart where Jon Stewart was asking him like,
[01:08] Chad Nantais: okay,
[01:09] Chad Nantais: where is this all headed?
[01:11] Chad Nantais: What is this tech utopia that the what they work what people like Jon Stewart insinuate are tech oligarchs?
[01:22] Chad Nantais: Where are they promising us?
[01:23] Chad Nantais: What does this utopia look like?
[01:25] Chad Nantais: And Jacob Coxon said
[01:28] Chad Nantais: there won't be any more work there won't be any more jobs it'll all be taken over by ai so it'll free up humans and then Jon Stewart said free up humans to do what humans derive a lot of their reason for living from their work you take the work away from them what is their reason to be alive what is their purpose and
[01:48] Chad Nantais: there was no good answer to that and to be honest i haven't heard a a good answer from any of the tech bosses either.
[01:55] Chad Nantais: And
[01:57] Chad Nantais: I was in.
[01:58] Chad Nantais: Venice and in Ghent in Belgium last month at two different conferences.
[02:04] Chad Nantais: One was the Cyber humanities Conference in Venice and the other one was SEMANTiCS in Ghent which dealt with building knowledge graphs and semantic open-linked web issues and ontologies maps of meaning that AI can sort of feed off of to become smarter and
[02:26] Paul Salazar: How different is that?
[02:27] Paul Salazar: How different is that?
[02:28] Paul Salazar: that you have just returned from,
[02:32] Paul Salazar: you know,
[02:33] Paul Salazar: we can make the argument that Europe seems to lead a lot of these efforts on humanizing technology in a broad sense.
[02:45] Paul Salazar: But I also think it's really interesting that it's in here on late night talk shows.
[02:50] Paul Salazar: It is,
[02:51] Paul Salazar: as we've talked about many times,
[02:53] Paul Salazar: the fever pitch is running high.
[02:56] Paul Salazar: And yet
[02:57] Paul Salazar: across the pond,
[02:59] Paul Salazar: it is broadly,
[03:01] Paul Salazar: I shouldn't say broadly,
[03:02] Paul Salazar: but it feels like it's more technical,
[03:04] Paul Salazar: it's more practical,
[03:06] Paul Salazar: it's more about real problem solving.
[03:08] Paul Salazar: And what a difference in just what's leading the thought patterns today.
[03:17] Paul Salazar: One very fatalistic and one very practical.
[03:19] Paul Salazar: And isn't that interesting that you have both moving in tandem at the same time,
[03:25] Paul Salazar: probably with the same
[03:27] Paul Salazar: uh rigor but one just gets a lot more attention
[03:32] Chad Nantais: absolutely um but again where did hollywood come from you know so it's the united states has an incredibly wealthy powerful tech industry and media and when you put the two together you get this feedback loop that just
[03:50] Chad Nantais: explodes things in into issues that no one can look away from all around the world right whether
[03:56] Chad Nantais: The scale of that noise is a true reflection of what's actually technically happening.
[04:02] Chad Nantais: We don't know.
[04:03] Chad Nantais: But what you're saying about Europe is correct.
[04:08] Chad Nantais: There is a different approach to how they think about AI.
[04:10] Chad Nantais: There are some people in Europe that are banging the drum that we have to keep up with the United States in the AI race.
[04:18] Chad Nantais: And there are other people in Europe who are saying that is a race we don't want to even get into because that race is purely brute force.
[04:26] Chad Nantais: who can increase the compute power at whatever cost at whatever cost to energy water consumption etc that's not a game that europe wants to play so that's what the other half are saying and of course at the humanities cyber humanities conference there
[04:44] Chad Nantais: was a call for and this is from computer science professors not just humanities people there for a more humane approach to ai one where
[04:56] Chad Nantais: computer engineers in university are being taught.
[04:59] Chad Nantais: a different perspective on AI development,
[05:02] Chad Nantais: one that isn't tech for tech's sake,
[05:04] Chad Nantais: where the ultimate goal isn't just more power,
[05:07] Chad Nantais: watch the number go up.
[05:08] Chad Nantais: It's not that.
[05:09] Chad Nantais: It's how do we now put the human back in the center of engineering,
[05:14] Chad Nantais: where we're collaborating with humans who are actual end users,
[05:18] Chad Nantais: and how are we making their lives better?
[05:20] Chad Nantais: How are we using AI as a way to augment human intelligence?
[05:23] Paul Salazar: You know what?
[05:25] Paul Salazar: That comment
[05:27] Paul Salazar: continues to surface in conversations that we have but that we hear that we have that we have heard that we have felt um that human experience that human component to ai and i think that shift to augmented
[05:49] Paul Salazar: intelligence versus artificial intelligence really is
[05:54] Paul Salazar: probably at the core of Syntopic.
[05:56] Paul Salazar: So I think it's a good moment.
[05:59] Paul Salazar: I'm really curious and
[06:01] Paul Salazar: I want to help everyone understand where this came from and where this idea and where the gestation of Syntopic started.
[06:09] Paul Salazar: Tell me about that in that same vein of augmented intelligence versus just pure artificial intelligence.
[06:16] Paul Salazar: Because I think it's a good starting point when someone thinks like,
[06:19] Paul Salazar: Like,
[06:19] Paul Salazar: what is Syntopic?
[06:23] Chad Nantais: Okay,
[06:23] Chad Nantais: so...
[06:24] Chad Nantais: Necessity is the mother of invention.
[06:27] Chad Nantais: And I found myself a few years ago in becoming the steward of someone's archive of thousands of hours of intellectual dialogues.
[06:37] Chad Nantais: And so this is in late 2023.
[06:41] Chad Nantais: ChatGPT had just been released on a broad scale earlier that year.
[06:46] Paul Salazar: And this archive is,
[06:47] Paul Salazar: they are podcasts,
[06:48] Paul Salazar: is that correct?
[06:49] Chad Nantais: They weren't really intended to be podcasts.
[06:51] Chad Nantais: This is an intellectual.
[06:53] Chad Nantais: who has made a habit of recording every conversation.
[06:57] Paul Salazar: Wow,
[06:57] Paul Salazar: so lectures,
[06:59] Paul Salazar: all of it?
[07:00] Chad Nantais: It's not so much lectures,
[07:01] Chad Nantais: a lot of phone conversations with other intellectuals.
[07:05] Paul Salazar: Oh,
[07:06] Paul Salazar: interesting.
[07:06] Chad Nantais: There's also radio appearances.
[07:09] Chad Nantais: There are a few lectures,
[07:10] Chad Nantais: seminars,
[07:11] Chad Nantais: things like that,
[07:12] Chad Nantais: but it's primarily like conference calls,
[07:16] Chad Nantais: one-on-one phone calls with other intellectuals and thousands of hours dating back to 1984.
[07:23] Chad Nantais: So
[07:24] Chad Nantais: I found myself in charge,
[07:26] Chad Nantais: put in charge of this to find a way to organize all this because
[07:30] Chad Nantais: I always wonder,
[07:31] Chad Nantais: like,
[07:31] Chad Nantais: OK,
[07:31] Chad Nantais: everybody's recording.
[07:32] Chad Nantais: Podcasts have been growing like this since 2005,
[07:36] Chad Nantais: since the iPod came out,
[07:37] Chad Nantais: basically.
[07:38] Chad Nantais: And they're not the growth isn't slowing down.
[07:42] Chad Nantais: But who's going to listen to all this stuff?
[07:45] Chad Nantais: Right.
[07:45] Chad Nantais: So it's going out into the feed.
[07:47] Chad Nantais: It's somewhere.
[07:48] Chad Nantais: So everybody,
[07:49] Chad Nantais: the big talk 10,
[07:50] Chad Nantais: 15 years ago was about big data.
[07:52] Chad Nantais: So look at all that.
[07:53] Chad Nantais: data that's being created by the big tech companies and it's this giant blob of data and if we could only find a way in to be able to organize it and use it to harness some kind of super intelligence so i started working with this archive and trying to find a way to best organize it having the same question what if you could tap into this big cloud of spoken thought because speech according
[08:20] Chad Nantais: to thinkers like Marshall McLuhan speech was
[08:23] Chad Nantais: the original technology it was the first extension of humans it was the extension of thought into the exterior social space so humans have been on this rampage now for the last 20 years of recording all this spoken thought and getting it out in into the cloud there's this giant cloud and i've always been a bit frustrated trying to search that cloud of audio it's been a little it's hard when you're searching on spotify you're
[08:48] Chad Nantais: basically just searching
[08:50] Chad Nantais: titles of episodes you're not deep searching into the conversations so again it's just today that does not in a in a practical omnipresent way for just the general public that really doesn't exist for us today not on a big scale no not on a large scale where you can search the whole corpus of
[09:12] Chad Nantais: podcasts on apple podcasts let's say that doesn't exist so i think
[09:17] Chad Nantais: YouTube.
[09:18] Chad Nantais: because a lot more podcasters have been putting their podcast onto YouTube that has served as sort of a way to do that.
[09:26] Chad Nantais: Like you can search YouTube,
[09:27] Chad Nantais: but it's not a really narrow search.
[09:30] Chad Nantais: You know,
[09:31] Chad Nantais: it's a very broad search for terms.
[09:33] Chad Nantais: You're going to get how-to videos.
[09:35] Chad Nantais: You're going to get all kinds of stuff.
[09:36] Chad Nantais: But just in that world of the spoken word itself,
[09:38] Chad Nantais: and YouTube isn't made for spoken word.
[09:40] Chad Nantais: YouTube is made for visual video stuff,
[09:43] Chad Nantais: right?
[09:44] Chad Nantais: So there's really no home for querying.
[09:48] Chad Nantais: this big cloud of knowledge,
[09:51] Chad Nantais: spoken knowledge.
[09:52] Paul Salazar: I'm a podcaster.
[09:53] Paul Salazar: I'm a new scaling podcaster with...
[09:58] Paul Salazar: with success.
[09:59] Paul Salazar: It's fueling me to do more.
[10:02] Paul Salazar: I mean,
[10:03] Paul Salazar: I'm using the traditional methods to grow user acquisition,
[10:07] Paul Salazar: et cetera,
[10:08] Paul Salazar: marketing.
[10:09] Paul Salazar: I've automated everything I possibly can.
[10:11] Paul Salazar: Why do I need this?
[10:13] Paul Salazar: What's different?
[10:14] Paul Salazar: What am I,
[10:16] Paul Salazar: how does this help me,
[10:17] Paul Salazar: the independent,
[10:19] Paul Salazar: or the someone with something meaningful to share?
[10:23] Paul Salazar: How does this help me?
[10:26] Chad Nantais: So you're putting out a podcast episode every week,
[10:29] Chad Nantais: let's say,
[10:29] Chad Nantais: or even a month.
[10:30] Chad Nantais: You're putting a lot of work into coming up with subject matter,
[10:34] Chad Nantais: finding the right guests.
[10:35] Chad Nantais: You're doing the work of being a podcaster.
[10:38] Chad Nantais: When you're finished recording your episode,
[10:40] Chad Nantais: you're going to go put it on Libsyn or you're going to put it directly on Spotify and have it distributed to the platforms.
[10:46] Chad Nantais: Put the cover on it and call it a day.
[10:49] Chad Nantais: All right.
[10:50] Chad Nantais: I just there's another one in the can.
[10:52] Chad Nantais: But then what happens?
[10:54] Chad Nantais: It goes to the top of the feed for...
[10:56] Chad Nantais: minutes,
[10:58] Chad Nantais: right?
[10:58] Chad Nantais: And it sinks quickly.
[11:00] Chad Nantais: And then it's forgotten.
[11:02] Chad Nantais: You have some core listeners who are subscribed and they're waiting for your next drop and they're listening to it and they're waiting for the next one.
[11:09] Chad Nantais: That's okay.
[11:10] Chad Nantais: But where do they get an interface where they can look back,
[11:15] Chad Nantais: they can search back into past episodes that you've done,
[11:18] Chad Nantais: or even across your episodes and see how concepts tie together.
[11:22] Chad Nantais: See how different thinkers that you interviewed you
[11:24] Chad Nantais: think differently about a specific topic or specific issue right there's no way for them to really get their hands in there and really work with it and query it so just the word Syntopic is an abbreviation of the word Syntopicon so there was an american philosopher in the middle of last century and
[11:45] Chad Nantais: um his name was Mortimer Adler and he created what's called the great books series where he was he sought to
[11:54] Chad Nantais: sort of create a compendium of all the greatest thought in western civilization and he created something called the Syntopicon which was looking at i think he had 150 different core concepts and when in western civilization and he created an index so for example there's one for angel the concept of an angel and
[12:16] Chad Nantais: across the whole canon of western thoughts which thinkers have written about
[12:21] Chad Nantais: the topic of an angel from their different perspectives and the word Syntopic Syntopicon is based on the concept of looking at something from different angles right so that's what now why is speech important again it goes back to speech is the closest that we have to thought writing
[12:39] Chad Nantais: is one step removed remember writing and the printing press came later first was oral tradition we had a very popular um movie this summer that was released the blockbuster
[12:52] Chad Nantais: about a homeric epic right yeah that's right yeah so that homer people might not realize this but homer was a blind poet and homer was a ma one of the he's the godfather of the oral tradition in western civilization right so
[13:14] Chad Nantais: when you look at the growth of podcasts and you look at the decline of literacy.
[13:20] Chad Nantais: They're having a hard time teaching kids how to read in North America now,
[13:23] Chad Nantais: to even learn how to read,
[13:25] Chad Nantais: because their eyes have been trained for this kind of motion on a screen of a video game or scrolling.
[13:31] Chad Nantais: The kids don't have the perceptual training to focus on one thing.
[13:36] Chad Nantais: So
[13:37] Chad Nantais: Marshall McLuhan predicted this,
[13:39] Chad Nantais: that the electronic technologies would move us away from literacy,
[13:43] Chad Nantais: back to the oral tradition.
[13:45] Chad Nantais: It's what
[13:46] Chad Nantais: Walter Ong calls secondary orality.
[13:49] Chad Nantais: This isn't the original orality of pre-printing press and pre-writing,
[13:55] Chad Nantais: where they sat around the fire like Homer and told tales.
[13:59] Chad Nantais: This is a different secondary,
[14:01] Chad Nantais: this is a retrieval of orality.
[14:03] Chad Nantais: So this is human thought.
[14:04] Chad Nantais: And one of the things that I learned by,
[14:06] Chad Nantais: look,
[14:07] Chad Nantais: I've been consuming podcasts for 20 years,
[14:09] Chad Nantais: but also by working with this large corpus of thousands of hours of intellectual conversations,
[14:15] Chad Nantais: what I notice is in the long form conversations,
[14:18] Chad Nantais: Like you've got some podcasters out there now like Lex Fridman who have five hour podcasts,
[14:23] Chad Nantais: right?
[14:24] Chad Nantais: So there's this long form.
[14:26] Chad Nantais: Something happens in a long form conversation that doesn't happen in a short conversation.
[14:31] Chad Nantais: New ideas are surfaced in those conversations.
[14:35] Chad Nantais: And you can hear the epiphanies happening as the speakers are discovering,
[14:39] Chad Nantais: as they're wrestling with the truth.
[14:41] Chad Nantais: You hear these ideas.
[14:43] Chad Nantais: It's almost like a laboratory.
[14:45] Chad Nantais: And this happens.
[14:47] Chad Nantais: socially and collectively as we're speaking together in a different way than it does if I go away and put my thoughts down in writing and think about them.
[14:58] Chad Nantais: That's the impetus behind this is if you can surface the development of these ideas that happened in spoken word conversations,
[15:08] Chad Nantais: you provide a whole new layer of learning,
[15:11] Chad Nantais: a whole new layer of insight for the listener,
[15:13] Chad Nantais: for the consumer of what's been created in these podcasts or these lecture series.
[15:18] Paul Salazar: Is it for me,
[15:19] Paul Salazar: the producer?
[15:23] Chad Nantais: It's hard to say that it's only for you,
[15:25] Chad Nantais: the producer,
[15:26] Chad Nantais: because the effect.
[15:27] Chad Nantais: that it offers the end user the consumer is very valuable as well so i think what it actually does just listening to you say this it brings the two sides closer together i think because it makes them both more deeply involved in the evolution of the ideas that are being discussed so yes there's a side effect a wonderful side effect that happens for the producers they can look back on the development of the thought that they've had with their guests or their co-hosts
[15:57] Chad Nantais: And that could shape future episodes,
[15:59] Chad Nantais: future discussions.
[16:00] Chad Nantais: So absolutely,
[16:01] Chad Nantais: that was an unintended effect.
[16:04] Chad Nantais: The original effect was simply,
[16:06] Chad Nantais: okay,
[16:06] Chad Nantais: there's all these great conversations,
[16:09] Chad Nantais: but your users,
[16:11] Chad Nantais: your listeners can't do much with them other than just keep listening to them over and over.
[16:16] Chad Nantais: So this is a good opportunity for me to explain the two main problems that had to be solved here with Syntopic.
[16:26] Chad Nantais: if you wanted to take a hundred let's say you have a hundred hours of podcast episodes
[16:31] Chad Nantais: 100 episodes times one hour each
[16:34] Chad Nantais: 100 hours if you wanted to get you want to build a website where people could really explore your content explore the concepts that are discussed across all of the episodes you'd have to have someone and again let's go back pre-ai
[16:50] Paul Salazar: the only way i could do it today is i would have to then find
[16:54] Paul Salazar: the podcast that i wanted
[16:56] Paul Salazar: through whatever search method or whatever tools I'm using.
[16:59] Paul Salazar: And then they're probably going to grab,
[17:03] Paul Salazar: a typical machine is going to grab what's high volume,
[17:07] Paul Salazar: what's popular,
[17:08] Paul Salazar: what meets the keyword search based on whatever prompts
[17:12] Paul Salazar: I shared.
[17:13] Paul Salazar: But it's not grabbing thoughts and ideas.
[17:16] Paul Salazar: That doesn't exist.
[17:17] Paul Salazar: In other words,
[17:18] Paul Salazar: I'm gonna get what's on the surface.
[17:22] Paul Salazar: I'm gonna get what's in the show notes,
[17:24] Paul Salazar: probably.
[17:25] Paul Salazar: whichever platform this producer is using but i'm not getting the ideas
[17:30] Chad Nantais: so again there's two problems that Syntopic has addressed i don't want to say completely solved but has brought them down to a manageable level where this becomes feasible where it was not feasible before so again if you go back before the quality of um
[17:48] Chad Nantais: speech to text got to where it is today so when i started in 2023
[17:54] Chad Nantais: processing this giant corpus of all these intellectual conversations.
[17:58] Chad Nantais: I started with one series and started to take it and put it through Whisper,
[18:02] Chad Nantais: which is OpenAI's speech to text.
[18:06] Chad Nantais: So you get a transcript out from there.
[18:08] Chad Nantais: And there was a certain error rate,
[18:10] Chad Nantais: right?
[18:11] Chad Nantais: There were errors produced in the process.
[18:14] Chad Nantais: Words incorrect,
[18:15] Chad Nantais: names incorrect,
[18:17] Chad Nantais: that kind of thing.
[18:18] Chad Nantais: It didn't even have speaker labels of like,
[18:21] Chad Nantais: oh,
[18:21] Chad Nantais: speaker A,
[18:22] Chad Nantais: speaker B.
[18:23] Chad Nantais: all that kind of thing didn't have that so i had to do a tremendous amount of cleanup and at that rate um it just wasn't feasible to take like a hundred hour
[18:34] Chad Nantais: 100 hours worth of dialogue and have that turned into a site that can be explored more deeply than show notes show notes are usually just some chapters of time codes where it started um what were the just an overall paraphrase of what was said in it and
[18:51] Chad Nantais: uh
[18:52] Chad Nantais: maybe the products that they're sponsoring,
[18:55] Chad Nantais: that kind of thing.
[18:55] Chad Nantais: So they're very flat,
[18:56] Chad Nantais: show notes are very flat.
[18:57] Chad Nantais: I wanted something more than that.
[18:59] Chad Nantais: But I couldn't get it back in 2023.
[19:02] Chad Nantais: The technology wasn't good enough.
[19:03] Chad Nantais: Because if you take sort of a half assed transcript,
[19:08] Chad Nantais: and you put that into another system that's trying to extract the concepts and hear where the arguments are,
[19:15] Chad Nantais: it's garbage in garbage out.
[19:17] Chad Nantais: So if you put a crappy transcript in,
[19:19] Chad Nantais: you're going to get a crappy
[19:21] Chad Nantais: website out,
[19:23] Chad Nantais: right?
[19:23] Chad Nantais: And again,
[19:24] Chad Nantais: this is more than show notes.
[19:26] Chad Nantais: This is like Wikipedia.
[19:28] Chad Nantais: You've got concepts that are connected to each other.
[19:31] Chad Nantais: So the concept of humanism connects to the concept of,
[19:34] Chad Nantais: or the person of Erasmus,
[19:36] Chad Nantais: and it connects to the city of Venice and the tradition in 1500,
[19:41] Chad Nantais: where you've got the Aldine Press printing these old,
[19:43] Chad Nantais: these Greek texts like Plato,
[19:45] Chad Nantais: and these things are connected.
[19:47] Chad Nantais: It's just seeing how things connect.
[19:49] Chad Nantais: That's what the
[19:50] Chad Nantais: the studious listener of a podcast or lecture series wants.
[19:53] Chad Nantais: They don't want to just listen to one episode and get something from that.
[19:57] Chad Nantais: They want to connect across.
[20:00] Chad Nantais: I like this podcast host.
[20:01] Chad Nantais: I like their guests.
[20:02] Chad Nantais: I want to see connections.
[20:03] Chad Nantais: I want to learn.
[20:04] Chad Nantais: That's what knowledge is.
[20:05] Chad Nantais: It's connecting.
[20:06] Chad Nantais: In other words,
[20:06] Paul Salazar: this is turning your
[20:10] Paul Salazar: MP3 back into vinyl.
[20:14] Paul Salazar: In other words,
[20:15] Paul Salazar: bring fidelity back to the podcast itself.
[20:18] Paul Salazar: It's not that you were changing the content.
[20:22] Paul Salazar: We're making it deeper.
[20:25] Paul Salazar: Fidelity.
[20:27] Chad Nantais: So that's the other side of the coin.
[20:29] Paul Salazar: A record player versus the MP3 in my earbuds or something like that.
[20:37] Chad Nantais: That's a good point.
[20:38] Chad Nantais: So you're getting,
[20:39] Chad Nantais: you're taking a flat conversation.
[20:42] Chad Nantais: I mean,
[20:42] Chad Nantais: flat in terms of it's just linear from second zero up to one hour.
[20:48] Chad Nantais: It's this linear thing of a conversation.
[20:50] Chad Nantais: And that's the best we've had was some show notes.
[20:54] Chad Nantais: But originally.
[20:56] Chad Nantais: The problem was this could not be done in an economic way.
[21:00] Chad Nantais: The cost of getting someone to,
[21:02] Chad Nantais: you have to have different responsibilities,
[21:04] Chad Nantais: getting someone to transcribe.
[21:05] Chad Nantais: So they have to listen to the 100 hours.
[21:08] Chad Nantais: They have to transcribe.
[21:10] Chad Nantais: Then you take the next person,
[21:12] Chad Nantais: will extract the entities.
[21:14] Chad Nantais: They'll see that Marshall McLuhan was mentioned in episode one,
[21:18] Chad Nantais: episode 17,
[21:19] Chad Nantais: episode 47,
[21:21] Chad Nantais: and then connecting those concepts to,
[21:23] Chad Nantais: oh,
[21:24] Chad Nantais: his book.
[21:24] Chad Nantais: Understanding Media was discussed in Episode 8,
[21:26] Chad Nantais: and connecting all those is a
[21:29] Chad Nantais: Herculean task for a human being.
[21:32] Chad Nantais: It requires an incredible amount of reading comprehension and being able to connect concepts across time,
[21:37] Chad Nantais: okay?
[21:38] Chad Nantais: So first of all,
[21:39] Chad Nantais: was that even really possible?
[21:40] Chad Nantais: I'm sure with some really intense,
[21:43] Chad Nantais: studious approach of someone that can extract these concepts and connect the dots,
[21:48] Chad Nantais: okay?
[21:49] Chad Nantais: And then from there,
[21:49] Chad Nantais: you have a whole other level of verification of truth.
[21:53] Chad Nantais: right that what's being stated on the summary pages that's connecting these concepts that it is true to the words that were originally spoken now at these conferences that i went to last month there were people there who do this with
[22:10] Chad Nantais: texts whether it's ancient texts from the middle ages or you know a famous writer from switzerland from the 20th century they're putting in the text into
[22:23] Chad Nantais: They're using an LLM to read the text and build a knowledge graph,
[22:28] Chad Nantais: okay?
[22:29] Chad Nantais: But nobody at these two conferences was doing this for the spoken word.
[22:33] Chad Nantais: They were doing it for books,
[22:35] Chad Nantais: monographs,
[22:37] Chad Nantais: old manuscripts,
[22:38] Chad Nantais: things like that.
[22:39] Chad Nantais: But nobody was doing it for audio.
[22:41] Chad Nantais: Audio is a different animal because when you're writing a book,
[22:45] Chad Nantais: you're very careful to use certain proper syntax and citations and all that.
[22:50] Chad Nantais: But the way that we communicate and share our thoughts with...
[22:52] Chad Nantais: in the spoken word is a different shape and it's less linear okay so there has to be something that's built for purpose and that's what Syntopic is it's built for the oral tradition and again number one this wasn't feasible for definitely not for a podcaster and even for a large institution that needs to process
[23:14] Chad Nantais: 20 000 hours of cultural historical recordings it's still not feasible
[23:21] Chad Nantais: at that level to be paying upwards of
[23:25] Chad Nantais: $5,000 per hour to do that whole pipeline of processing and building this sort of wiki site around a spoken word series.
[23:34] Chad Nantais: So
[23:35] Chad Nantais: Syntopic uses AI,
[23:39] Chad Nantais: uses an LLM,
[23:40] Chad Nantais: to do some of the work,
[23:42] Chad Nantais: well most of the work,
[23:43] Chad Nantais: and having a harness so that the human still drives the process.
[23:48] Chad Nantais: At the end of the day,
[23:49] Chad Nantais: the human can review what is being written into the articles on the website
[23:54] Paul Salazar: question that's a good question i wanted to ask that so i feel like we have this problem that is not it's not absurd it's not groundbreaking that none of us have thought about this but the idea that that there is a problem that spoken word is hard to search spoken word is hard
[24:18] Paul Salazar: to capture and i don't know about most people's memory but let's face it i don't think everyone's memory is uh maybe do you remember do you remember the last podcast you listened to how much of it do you actually remember
[24:32] Chad Nantais: good point i'm going to defer that
[24:35] Chad Nantais: question to later so that's partly i just want to throw that in there that's another part of it is i thought i i listened to all these podcasts and i realized like it's going in one ear and out the other i need some way to freeze that
[24:47] Chad Nantais: and so that i can explore it
[24:49] Paul Salazar: here's the elephant in the room let's call it out how do i know that what Syntopic produces is the right stuff we could talk ad nauseum about ai slop hallucination
[25:03] Paul Salazar: um you know all the sinister things all the hollywood type horror stories that could happen how is this technology not one of those stories
[25:14] Chad Nantais: so that goes to the second problem so the first problem was the economics right how do you have a human in
[25:20] Chad Nantais: the loop so that they can ensure that this is a very uh sound website that's been developed
[25:31] Chad Nantais: from the audio,
[25:32] Chad Nantais: right?
[25:32] Chad Nantais: You have to have the human involved to a certain degree,
[25:34] Chad Nantais: but you have to have it in a way that's economical,
[25:37] Chad Nantais: that's affordable.
[25:38] Chad Nantais: So that was the first problem is make it reviewable by a human,
[25:42] Chad Nantais: just enough reviewable,
[25:44] Chad Nantais: harness the power of the LLM,
[25:47] Chad Nantais: make it so it's economically feasible,
[25:49] Chad Nantais: even for a podcaster.
[25:50] Chad Nantais: The second problem is exactly what you're saying.
[25:53] Chad Nantais: How do you know that these notes and this article and these connections between these different concepts,
[25:58] Chad Nantais: how do you know it's true?
[26:00] Paul Salazar: so how does the product in what you've built where's the check and balance where's the human in the loop
[26:07] Chad Nantais: well before the human in the loop part see that's the thing we have to reduce we want the human to be in the loop we don't want the human to be in the loop for everything because the human's in the loop for everything
[26:17] Paul Salazar: i don't have time for that
[26:19] Chad Nantais: well
[26:19] Chad Nantais: you don't have time for that but also nobody has budget for that like it's it's ridiculous like i'm pegging it at somewhere between three and five thousand dollars per hour of recorded material
[26:30] Paul Salazar: oh gosh
[26:30] Chad Nantais: right so who even an institution like it would have to be the most like smithsonian level treasure of a country that they'd be willing to pay that much to do this for spoken word it have to be like jfk speeches or something like that but barring
[26:46] Chad Nantais: that making it affordable means harnessing the llm but in a way that the human is still in charge but the human's not involved in every little aspect so that means the human only reviews the the most
[26:59] Chad Nantais: error-prone aspects,
[27:01] Chad Nantais: the parts that the LLM makes the most mistakes on,
[27:06] Chad Nantais: and presenting those in a way with a quick,
[27:09] Chad Nantais: we built a quick review panel,
[27:11] Chad Nantais: so the human is only presented with what they need to actually check.
[27:15] Chad Nantais: They can go deeper if they want to,
[27:17] Chad Nantais: but it hits,
[27:18] Chad Nantais: you know,
[27:18] Chad Nantais: the whole 80-20 thing,
[27:19] Chad Nantais: it's hitting like the 90%
[27:22] Chad Nantais: of things that can go wrong with 10%
[27:24] Chad Nantais: of the effort.
[27:26] Chad Nantais: Brings me to how do we do that?
[27:28] Chad Nantais: How do we make it?
[27:29] Chad Nantais: How do we trust that the machine has already filtered out a set of the most common possible mistakes that could have happened?
[27:39] Chad Nantais: So what we did was there's a two level process of gates.
[27:44] Chad Nantais: And this is what I did the presentation on in Ghent at SEMANTiCS.
[27:47] Chad Nantais: It's called from paraphrase to provenance.
[27:51] Chad Nantais: Because if you people have asked me,
[27:52] Chad Nantais: why don't I just take my podcast and dump it into Google Gemini or something like that?
[27:58] Chad Nantais: And it'll just transcribe,
[27:59] Chad Nantais: it'll do all this and say,
[28:00] Chad Nantais: good luck,
[28:01] Chad Nantais: because you will get paraphrasing where you think you're getting verbatim quotes.
[28:06] Chad Nantais: How do I know?
[28:06] Chad Nantais: Here's one example.
[28:08] Chad Nantais: I asked Gemini,
[28:10] Chad Nantais: give me three quotes about whatever topic it was.
[28:13] Chad Nantais: Let's say about human suffering from the Indian philosopher,
[28:17] Chad Nantais: Jiddu Krishnamurti.
[28:18] Chad Nantais: It gave me three quotes in quotations.
[28:22] Chad Nantais: I looked at those quotes and there was something fishy about them.
[28:24] Chad Nantais: And then I asked Gemini,
[28:26] Chad Nantais: are those actual quotes?
[28:27] Chad Nantais: from talks that krishnamurti did and gemini said oh i'm glad you asked actually they're not actual verbatim quotes they're quotes that he could have said based on you know his philosophical principles so as i started to process our first paying client their corpus of
[28:47] Chad Nantais: 81 episodes a total of 160 plus hours of dialogues for the first half of that at least
[28:55] Chad Nantais: I was just rolling along quickly,
[28:57] Chad Nantais: letting the LLM build out this wiki-like website with episodes linked to concepts and books and thinkers that are being discussed and all this.
[29:08] Chad Nantais: I'm flying along.
[29:09] Chad Nantais: And then I noticed some key errors in what was being quoted as verbatim from the speakers.
[29:16] Chad Nantais: So I dug deeper and I realized,
[29:18] Chad Nantais: uh-oh,
[29:18] Chad Nantais: we've got a problem here.
[29:20] Chad Nantais: I have overtrusted the LLM to surface verbatim quotes.
[29:25] Chad Nantais: from these recordings.
[29:27] Chad Nantais: So I had to hit the brakes on that and go back and engineer a solution on how we can.
[29:34] Chad Nantais: Yeah.
[29:34] Chad Nantais: So there's two levels of gates and I won't get too technical about that,
[29:38] Chad Nantais: but that's a whole other episode.
[29:42] Chad Nantais: It's very techie.
[29:43] Chad Nantais: So one level is deterministic,
[29:45] Chad Nantais: which means we're not trusting the LLM.
[29:48] Chad Nantais: This is actual like Python code.
[29:50] Chad Nantais: This is deterministic,
[29:52] Chad Nantais: not
[29:53] Chad Nantais: LLM.
[29:54] Chad Nantais: LLMs are probabilistic.
[29:58] Chad Nantais: they're using statistics to guess what the next word should be based on the data that it has about the past.
[30:04] Chad Nantais: So we need a level in there that's not
[30:07] Chad Nantais: AI because AI is guessing.
[30:10] Chad Nantais: They're confidently wrong when they're wrong.
[30:12] Chad Nantais: They don't warn you that they guessed.
[30:14] Chad Nantais: Right.
[30:14] Chad Nantais: So we need one layer has to be deterministic,
[30:17] Chad Nantais: which is just Python,
[30:18] Chad Nantais: like arithmetic.
[30:19] Chad Nantais: Basically,
[30:20] Chad Nantais: it's the rules are the rules.
[30:23] Chad Nantais: You're not guessing.
[30:25] Chad Nantais: You're screening.
[30:26] Chad Nantais: based on this for potential errors in what was quoted from the spoken word so you got that level and then you've got the llm level on top of that that's checking the deterministic level
[30:39] Paul Salazar: so you've infused a a level or a layer of critical thinking into
[30:44] Paul Salazar: the into the digestion
[30:47] Chad Nantais: exactly but you've got actually i said two layers actually three layers right so you've got deterministic good old-fashioned code
[30:54] Chad Nantais: Then you've got the AI level,
[30:56] Chad Nantais: the LLM,
[30:57] Chad Nantais: and then you've got the human.
[30:58] Chad Nantais: So there's stacked layers of checking.
[31:01] Chad Nantais: So quotes from the corpus,
[31:04] Chad Nantais: from the recordings,
[31:05] Chad Nantais: bubble up to the top,
[31:06] Chad Nantais: and eventually they're going to hit the human.
[31:08] Chad Nantais: And the human's going to verify a claim before it goes on the web page for that episode as a notable quote.
[31:16] Chad Nantais: It's going to be checked by a human before it goes out there.
[31:20] Chad Nantais: But the human doesn't have to review all the transcripts.
[31:22] Chad Nantais: doesn't have to review 100 hours of transcripts.
[31:25] Chad Nantais: The human just has to review what's about to go on the article page.
[31:29] Chad Nantais: And again,
[31:29] Chad Nantais: we have different filters and checks.
[31:32] Chad Nantais: There's another one called the sensitivity filter that makes sure that nothing's sensitive and sensitive is defined by the client.
[31:39] Chad Nantais: One example could be,
[31:42] Chad Nantais: you know,
[31:42] Chad Nantais: making sensitive claims about a living person saying,
[31:46] Chad Nantais: for example,
[31:46] Chad Nantais: saying that,
[31:47] Chad Nantais: you know,
[31:47] Chad Nantais: that guy really went to a lot of drug parties in New York in the
[31:51] Chad Nantais: 60s.
[31:52] Chad Nantais: things like that that could be considered um libelous any of that stuff so there's different layers of filters and the sum total of this is again the human gets to be in charge the final arbiter the final judge but doesn't have to review every single letter that was transcribed so this goes back to the augmented technology concept
[32:11] Paul Salazar: that we've been talking about um in practical terms i know we've had this conversation before but how much time are we saving
[32:21] Paul Salazar: Is this exponential time savings?
[32:24] Chad Nantais: Well,
[32:24] Chad Nantais: I wouldn't say exponential.
[32:25] Chad Nantais: I would say it's a multiple.
[32:27] Chad Nantais: It's geometric.
[32:29] Chad Nantais: So
[32:29] Chad Nantais: I ran some numbers on it,
[32:31] Chad Nantais: and it looks like,
[32:33] Chad Nantais: and this can't be held up in a court of law level,
[32:37] Chad Nantais: but just based on estimates,
[32:38] Chad Nantais: it looks like
[32:39] Chad Nantais: Syntopic creates basically eight times savings in money and time.
[32:50] Paul Salazar: Not to mention the fact that the platform itself creates an opportunity to even do this in the first place.
[32:57] Chad Nantais: There you go.
[32:58] Chad Nantais: Without that
[32:59] Chad Nantais: 8x increase in efficiency,
[33:01] Chad Nantais: this isn't feasible for most.
[33:04] Paul Salazar: So really,
[33:06] Paul Salazar: really interesting.
[33:08] Chad Nantais: And that's only sorry.
[33:09] Chad Nantais: I would say that
[33:10] Chad Nantais: 8x improvement in efficiency.
[33:13] Chad Nantais: That's just off the top.
[33:15] Chad Nantais: Like we haven't even begun to like.
[33:18] Chad Nantais: It's only version one of the review panel.
[33:21] Chad Nantais: And what the review panel is,
[33:22] Chad Nantais: so this is where the human in the loop does their work,
[33:24] Chad Nantais: this is their control panel.
[33:26] Chad Nantais: As an episode gets processed,
[33:28] Chad Nantais: the reviewer is notified to come and check over not everything,
[33:33] Chad Nantais: but just the key claims,
[33:34] Chad Nantais: the key statements on the episode page,
[33:38] Chad Nantais: and on the pages for the people and books.
[33:41] Paul Salazar: Talk to me about one of these core values of Syntopic and provenance.
[33:47] Paul Salazar: How does that fit into this world where more and more we have concerns?
[33:53] Paul Salazar: Sometimes these things are demonized.
[33:54] Paul Salazar: Sometimes they're perhaps sensationalized.
[33:58] Paul Salazar: But
[33:58] Paul Salazar: I do think that there is a place for conversation about how much distortion is okay.
[34:08] Paul Salazar: And I think
[34:09] Paul Salazar: Syntopic and you and I,
[34:11] Paul Salazar: we've thought about this a lot.
[34:13] Paul Salazar: But tell me from its core.
[34:15] Paul Salazar: and kind of as a value where this provenance fits into the product itself and and what we've built
[34:23] Chad Nantais: well there's a relevant story um in terms of current events that i found out about a few months ago that Anthropic
[34:36] Chad Nantais: had something this is public knowledge now because of some court case and the publication
[34:44] Chad Nantais: of some memo that was leaked from Anthropic.
[34:48] Chad Nantais: So this isn't something that they wanted the world to know,
[34:51] Chad Nantais: apparently,
[34:52] Chad Nantais: but something called Project Panama.
[34:56] Chad Nantais: inside Anthropic,
[34:59] Chad Nantais: the goal of which was to buy all the books in the world,
[35:03] Chad Nantais: scan them all,
[35:05] Chad Nantais: and then slice them up.
[35:08] Chad Nantais: The reason for this is because of something called the
[35:13] Chad Nantais: AI epoch or the LLM epoch.
[35:19] Chad Nantais: After 2022,
[35:21] Chad Nantais: you don't know whether a book or a text was completely written by a human.
[35:27] Chad Nantais: or partly written by LLM or wholly written by an llm you only know before ChatGPT went live before that a book a text was the the product of human thinking after
[35:45] Chad Nantais: that you don't know if it's synthetic or not so there's a race across the frontier labs to get their hands on
[35:56] Chad Nantais: what is verified as human thought,
[35:58] Chad Nantais: human originated thought in the form of books,
[36:01] Chad Nantais: texts,
[36:02] Chad Nantais: you name it.
[36:04] Chad Nantais: So
[36:05] Chad Nantais: Project Panama's goal was to get all the books they could get their hands on,
[36:09] Chad Nantais: scan them into the system,
[36:10] Chad Nantais: knowing that these books contain true human thought,
[36:17] Chad Nantais: not mixed in with LLM thought.
[36:20] Chad Nantais: With
[36:21] Chad Nantais: Syntopic, we're focused on spoken word.
[36:24] Chad Nantais: So the same problem can happen with the spoken word,
[36:27] Chad Nantais: right?
[36:29] Chad Nantais: How do you know that what was paraphrased when an LLM is summarizing a conversation in a podcast,
[36:36] Chad Nantais: how do you know that it's a true reflection,
[36:38] Chad Nantais: it's the exact words that were spoken?
[36:40] Chad Nantais: How do you know that?
[36:41] Chad Nantais: So that's where this word provenance comes from.
[36:44] Chad Nantais: When you buy a bottle of fine wine,
[36:46] Chad Nantais: expensive bottle of wine,
[36:48] Chad Nantais: you want to know the provenance.
[36:49] Chad Nantais: You want to know who owned this bottle of wine as it was traded.
[36:54] Chad Nantais: over the years since 1840.
[36:56] Chad Nantais: A painting,
[36:57] Chad Nantais: you want to know the provenance.
[36:58] Chad Nantais: You want to know that it is truly a Salvador Dali painting.
[37:02] Chad Nantais: You want to trace the ownership back to it.
[37:04] Chad Nantais: They're doing this now with luxury goods.
[37:07] Chad Nantais: They've introduced provenance as a way to know that this is a genuine article.
[37:11] Chad Nantais: So we're focused on this summary of this conversation.
[37:16] Chad Nantais: Is this the genuine article?
[37:17] Chad Nantais: Is this truly what was said?
[37:19] Chad Nantais: Or is it distorted by AI hallucinations?
[37:23] Chad Nantais: or has it been distorted on the level of when it was scanned?
[37:27] Chad Nantais: You know what I mean?
[37:28] Chad Nantais: So that's why provenance matters.
[37:32] Chad Nantais: Provenance was incredibly important in the discussions at the
[37:37] Chad Nantais: IEEE Cyber humanities Conference in Venice because they're dealing with protecting the collective memory,
[37:46] Chad Nantais: let's say,
[37:46] Chad Nantais: of Europe in the form of paintings,
[37:49] Chad Nantais: texts,
[37:50] Chad Nantais: recordings,
[37:51] Chad Nantais: you name it.
[37:52] Chad Nantais: right?
[37:53] Chad Nantais: So they don't want to lose the cultural memory.
[37:55] Chad Nantais: So there's all kinds of threats.
[37:57] Chad Nantais: And we are exclusively focused on spoken word and how to protect that as it's being interpreted by LLMs.
[38:06] Chad Nantais: And the more content that is generated by LLMs,
[38:09] Chad Nantais: because again,
[38:09] Chad Nantais: we're in the post-epoch now.
[38:11] Chad Nantais: Now we don't know if something was indeed written by a human or edited by an LLM or elaborated on by an LLM.
[38:21] Chad Nantais: So you're getting...
[38:22] Chad Nantais: even if someone does use an llm to take what was said in a podcast and create pages about it on the internet we still don't know unless we have receipts whether those are the actual words that were said and sometimes you get a situation where one llm is quoting another llm and if it's wrong you get the telephone effect you know the telephone game where with you know time
[38:46] Chad Nantais: it changes