Latest / The Adults In The AI Room / [Earley AI Podcast] Episode 72: AI in the Enterprise: What Really Matters

Transcript

[Earley AI Podcast] Episode 72: AI in the Enterprise: What Really Matters · The Adults In The AI Room

• 0:01
  SPEAKER_00: This is the Early AI Podcast hosted by Seth Early.
• 0:04
  SPEAKER_00: Join us as we delve deep into the passions, expertise, and experiences of thought leaders and practitioners to talk about what's possible with artificial intelligence.
• 0:14
  SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science, your digital transformation journey with design and deployment of innovative technology solutions, as well as Vector.
• 0:26
  SPEAKER_00: Now enjoy the show.
• 0:28
  SPEAKER_02: So welcome everyone to today's podcast.
• 0:31
  SPEAKER_02: My name is Seth Early.
• 0:33
  SPEAKER_02: I'm your host.
• 0:34
  SPEAKER_02: And today we're going to be exploring a lot of different topics in the field of AI, everything from AI content detection to the role of data strategy in this world that's being reshaped by generative AI.
• 0:46
  SPEAKER_02: We're going to talk about search integration.
• 0:48
  SPEAKER_02: We'll talk about agent technology.
• 0:51
  SPEAKER_02: There's a lot of hype.
• 0:52
  SPEAKER_02: There's no shortage of hype.
• 0:54
  SPEAKER_02: So we're going to try to cut through some of the hype and get to some of the real deal of what really matters to organizations in the AI space and to get AI to work appropriately.
• 1:04
  SPEAKER_02: We'll talk about that MIT study that just came out.
• 1:07
  SPEAKER_02: I really want to hear your thoughts on that, Chris.
• 1:09
  SPEAKER_02: And our guest today is Christopher Penn.
• 1:11
  SPEAKER_02: He is co-founder and chief data strategist at Trust Insights.
• 1:15
  SPEAKER_02: Chris is a widely recognized authority on analytics, data science, and AI machine learning.
• 1:20
  SPEAKER_02: He's the author of multiple books on marketing and analytics.
• 1:23
  SPEAKER_02: He's a keynote speaker.
• 1:24
  SPEAKER_02: He's a trusted advisor for organizations navigating digital transformation.
• 1:29
  SPEAKER_02: At Trust Insights, Chris helps companies make sense of complex data challenges and put AI into practice in ways that drive measurable outcomes.
• 1:38
  SPEAKER_02: Chris, welcome to the show.
• 1:40
  SPEAKER_01: Thank you for having me.
• 1:46
  SPEAKER_02: So you're pretty busy speaking all over the world.
• 1:49
  SPEAKER_02: So we're really honored to have you today.
• 1:51
  SPEAKER_02: So tell me what's been going on.
• 1:53
  SPEAKER_02: Like we said that a lot has changed.
• 1:55
  SPEAKER_02: One of the things we talked about way back when I had read one of your posts where people were talking about AI detection, and you had uh posted that the Declaration of Independence, you ran it through one of those AI detectors, and it said, this is totally AI created.
• 2:11
  SPEAKER_02: So apparently the founding fathers had some early version of horse-powered uh Chat GPT and were able to do it with artificial intelligence, I guess.
• 2:21
  SPEAKER_01: Yeah, uh Lynn Manuel Miranda was uh was having a good time.
• 2:24
  SPEAKER_01: AI detectors are trying to detect something that's undetectable.
• 2:28
  SPEAKER_01: They are designed to try and say, okay, this is AI generated text.
• 2:32
  SPEAKER_01: And that's a really difficult thing to do because think about this.
• 2:36
  SPEAKER_01: What if I wrote a detector that said this is going to detect a man's writing, right?
• 2:39
  SPEAKER_01: It's something like, but what does that even mean?
• 2:42
  SPEAKER_01: Like you and I, we're we're both uh signed male at birth, and we have different writing styles.
• 2:47
  SPEAKER_01: Like you and I do not write the same, we're both men, but so you know, you could not synthesize and say, oh, well, this was written by a man.
• 2:54
  SPEAKER_01: Like, no, you can't do that.
• 2:56
  SPEAKER_01: And that's kind of what these detectors are trying to do is say this is written by AI.
• 3:00
  SPEAKER_01: Now, there is a way to detect AI writing, however, it is exceptionally compute intensive.
• 3:08
  SPEAKER_01: And what you have to do is you have to build detectors that detect individual models writing, because you can detect if something was written by Deep Seek or Kimmy K2 or OpenAI's GPT-5, but it's unique to that model in the same way that you and I have unique writing styles.
• 3:25
  SPEAKER_02: Right, right.
• 3:26
  SPEAKER_02: So, so tell me, you know, it seems to me that it, you know, people get distracted by the wrong things, right?
• 3:33
  SPEAKER_02: Because it's not about whether something is AI written, it's really about the authorship and the voice.
• 3:39
  SPEAKER_02: And AI is a tool.
• 3:40
  SPEAKER_02: And I, when we talked last time, I talked about the fact that I certainly use AI, but I use it to take content that I've already written or already spoken about, and I also give it samples of my actual human AI, human non-AI writing.
• 3:56
  SPEAKER_02: So it knows my tone and my voice and my style and my intent.
• 4:00
  SPEAKER_02: And then it produces something that really does sound like me, but it's also based on my voice.
• 4:05
  SPEAKER_02: It's also based on my content.
• 4:07
  SPEAKER_02: And so I think what we're what we're realizing is it's not about using AI, it's about using AI effectively, and it's about not handing over authorship to the AI, because that doesn't do us any good anyway.
• 4:21
  SPEAKER_02: You can't see different patterns in how it's handling things, you know, and and uh, but but I think the point here is, you know, how are our organizations using it effectively?
• 4:31
  SPEAKER_02: How do they use it effectively?
• 4:33
  SPEAKER_02: And if you're turning it over to do something generic, if you're asking the question of the large language model's understanding of the world, that's not really giving you anything that's uh a competitive advantage, right?
• 4:44
  SPEAKER_02: That's giving you some efficiency, but it's what everybody else does, if it's what everybody else is doing, it's not giving you competitive advantage.
• 4:50
  SPEAKER_02: Competitive advantage comes from differentiation, differentiation comes from your voice and your knowledge and your expertise and your data and so on.
• 4:59
  SPEAKER_02: So, what is your sense of you know, this idea of it almost seems like a moot point of trying to detect whether AI is used, it's it's it's really about what are you using it for and what's the value.
• 5:10
  SPEAKER_02: So, do you want to kind of give us uh your your summation of that space in terms of the fruitlessness of trying to detect if AI is used and the real value of how to use it?
• 5:22
  SPEAKER_01: Here's the thing when you talk, when you go out into the real world and you talk to real people who are not inside the AI LinkedIn thought leadership bubble, right?
• 5:30
  SPEAKER_01: And you talk to real people, nobody cares who wrote it.
• 5:33
  SPEAKER_01: They don't care if a human wrote something or a machine wrote something.
• 5:36
  SPEAKER_01: They literally could not care less.
• 5:38
  SPEAKER_01: What they do care about is does it solve does whatever it is, a piece of content, an infographic, a YouTube video, whatever does solve their problem.
• 5:46
  SPEAKER_01: There are so many YouTube channels now that are purely synthetic.
• 5:50
  SPEAKER_01: I though I was uh uh talking about one this morning.
• 5:52
  SPEAKER_01: It's a YouTube channel that's made entirely synthetic, synthetic voice, it's Google's TTS, it's a it's a script written probably by ChatGPT, but it is tips and tricks for Adobe Premiere.
• 6:02
  SPEAKER_01: And this channel gets thousands of views.
• 6:04
  SPEAKER_01: I use the channel all the time as an end user.
• 6:06
  SPEAKER_01: Why?
• 6:07
  SPEAKER_01: I don't not because it's got special authors, but because it solves my problem.
• 6:11
  SPEAKER_01: How do I use Chroma key in Premiere?
• 6:13
  SPEAKER_01: This channel tells me.
• 6:15
  SPEAKER_01: Go on Instagram.
• 6:16
  SPEAKER_01: There are so many synthetic influencers, people who are not real, they don't exist, they're clearly machine generated.
• 6:22
  SPEAKER_01: Nobody cares.
• 6:23
  SPEAKER_01: Now, people are fully aware this is a machine, they don't care.
• 6:27
  SPEAKER_01: They they're like, this solves my my issue.
• 6:29
  SPEAKER_01: We always apply the three e test to any piece of content.
• 6:33
  SPEAKER_01: Is it educational?
• 6:34
  SPEAKER_01: Is it engaging, or is it entertaining?
• 6:36
  SPEAKER_01: If it's gotta be at least one of those three, and if you know, two out of three ain't bad, like Meatloaf sang before uh in the back of the 70s, and three out of three is is perfect.
• 6:45
  SPEAKER_01: Again, we're so wrapped up in the the human side of things from the oh uh, is this authentic or not that we are not paying attention to the customer side of things, which is does this solve a problem I have?
• 6:56
  SPEAKER_01: And that's where that's where this big schism is.
• 6:59
  SPEAKER_01: Now, there are consequences.
• 7:01
  SPEAKER_01: So I was reading yesterday uh and today on the employment numbers and looking at some of the studies coming out about AI's impact.
• 7:09
  SPEAKER_01: It is having a discernible impact on employment, particularly for entry-level jobs.
• 7:13
  SPEAKER_01: Entry-level unemployment for particularly for new college graduates is more than double the unemployment rate for the rest of the population.
• 7:20
  SPEAKER_01: And in a study, I can't remember if it was MIT or Harvard that did it, but it came out a few uh a few days ago showing that the impact of AI is most dramatic on the least senior people and is having hiring impact.
• 7:33
  SPEAKER_01: So there are aspects of it that are having a material impact on people's lives, and that part does matter.
• 7:39
  SPEAKER_02: And it's interesting because you know it it has the ability to upskill people and to put people with less experience into a job where they have an AI assistant, but all of that programming, the AI assistance requires foundational information architecture, knowledge bases, and so on, right?
• 7:58
  SPEAKER_02: So organizations.
• 8:00
  SPEAKER_01: So I used to work at a PR agency, and our lowest ranking person was called an account coordinator.
• 8:05
  SPEAKER_01: This person had four jobs.
• 8:07
  SPEAKER_01: Job number one, take meetings during client notes, take notes during client meetings, and and send the notes to the client afterwards, two, write draft first draft press releases from templates.
• 8:18
  SPEAKER_01: Three, copy and paste Google search results into a spreadsheet, and four, fetch coffee for their manager.
• 8:25
  SPEAKER_01: The only job AI can't do of those is fetching coffee, and managers can get their own damn coffee.
• 8:29
  SPEAKER_01: And these are this is the order in very large like WPP and publicists, but these are big, big firms.
• 8:37
  SPEAKER_01: And what's happening in that's a job where yeah, you don't you don't need a human in that role at all, right?
• 8:44
  SPEAKER_02: Right.
• 8:45
  SPEAKER_02: And and and the the technologies can handle that without any kind of curated content.
• 8:50
  SPEAKER_02: It's zero shot.
• 8:52
  SPEAKER_02: That's right.
• 8:52
  SPEAKER_02: Just saying summarize this meeting, create the action items, and that's wonderful stuff, right?
• 8:57
  SPEAKER_02: We use that all the time, right, for everything you're doing.
• 9:00
  SPEAKER_02: And if you're not doing that, then you're not efficient, right?
• 9:02
  SPEAKER_02: I I guess what I'm also thinking about are some of the tech support jobs that do require knowledge, that do require training, that are that are increasingly being automated, that are are more uh challenging to uh enable more advanced tech people, right?
• 9:19
  SPEAKER_02: In other words, uh you do have to kind of go through a learning curve to get to a certain level of proficiency.
• 9:24
  SPEAKER_02: But for some of our clients, we're offloading that level one support to either self-service or to a minimally trained individual who can use the scripts, who can use the you know, the whispered coaching, who can use those things.
• 9:37
  SPEAKER_02: And that requires a little bit of data preparation and content and so on.
• 9:41
  SPEAKER_02: Uh, but that's fine.
• 9:42
  SPEAKER_02: One of the things that we talked about last time was that you know, how are people looking at generative AI?
• 9:50
  SPEAKER_02: So we have this MIT study that just came out that uh which no one read.
• 9:56
  SPEAKER_02: That no one actually read the study because there are some nuances in it, but they took the headline and everybody took the headline and the headline and the headline and the headline and headline, right?
• 10:04
  SPEAKER_02: And said, oh, it's five percent, it's five percent.
• 10:07
  SPEAKER_02: Unpack that a little bit and tell us what your thoughts are about that study.
• 10:11
  SPEAKER_01: So the headline that people are saying is 95% of AI projects fail, right?
• 10:15
  SPEAKER_01: That was the what everyone took away.
• 10:16
  SPEAKER_01: If you actually read the damn thing, you would have read 95% of AI pilot projects in an enterprise in a six-month study period did not make it out of pilot phase.
• 10:27
  SPEAKER_01: Now, you and I have worked in enterprises.
• 10:29
  SPEAKER_01: You can't get a coffee machine approved in six months, much less get a pilot project moved into production.
• 10:36
  SPEAKER_01: Zero projects in an enterprise make it out of pilot in six months, unless it's like a new stapler, right?
• 10:42
  SPEAKER_01: So that was the biggest flaw in that that that research project.
• 10:46
  SPEAKER_01: It was a good general analysis of how AI is being thought about in the enterprise, but that six-month period, nothing's gonna happen.
• 10:54
  SPEAKER_01: We have one we had one client project, for example, there's an enterprise healthcare company doing an AI deployment, two years just to get started.
• 11:04
  SPEAKER_01: I mean, not even the pilot, just to get the pilot started was two years.
• 11:09
  SPEAKER_01: People do not understand this the length of a time it takes for enterprises to move.
• 11:14
  SPEAKER_02: Yeah, yeah.
• 11:15
  SPEAKER_02: It's the clock speed, it's all the organizational complexity, it's the politics, it's the process.
• 11:21
  SPEAKER_02: Politics.
• 11:22
  SPEAKER_02: It's all across, yes, it's all across uh, you know, these cross-enterprise, cross-applications, all the fiefdoms, all the ownership, all the data ownership issues, the data security issues, all of those things.
• 11:35
  SPEAKER_01: Regulations, legislation.
• 11:38
  SPEAKER_02: Yeah, and and you know, it's funny is I said once before, you know, having the answer doesn't help, right?
• 11:42
  SPEAKER_02: Like if you have the answer, the perfect answer, and you say, here it is, an organization is still gonna be like, hmm, really?
• 11:48
  SPEAKER_02: No, I don't know, maybe you know, even if you show it with their data, it's still very difficult.
• 11:54
  SPEAKER_02: How do you overcome that?
• 11:55
  SPEAKER_02: I mean, what's is that just organizational inertia at its core?
• 11:59
  SPEAKER_02: How do you get people because we're in situations right now where the obvious thing to do is to take the POV proof of value, which is built toward a deployment, and execute on that and go back and fix the rest of the data and scale this thing, yet there's still resistance.
• 12:18
  SPEAKER_02: What is your sense of that?
• 12:20
  SPEAKER_02: I know there's a myriad of reasons why those things happen, but why do you think even with this day and age of needing to act with urgency, we're not seeing as much urgency.
• 12:30
  SPEAKER_02: What are your thoughts about that?
• 12:31
  SPEAKER_01: Uh, the easiest way to explain it is it the the number one acronym in enterprise is CYA, cover your ass, right?
• 12:39
  SPEAKER_01: And most stakeholders care about that the most because that's the way enterprises function.
• 12:44
  SPEAKER_01: We worked on an enterprise project for a med tech company auditing their CRM infrastructure.
• 12:49
  SPEAKER_01: This is pre-generative AI.
• 12:51
  SPEAKER_01: And we did our due diligence, we interviewed stakeholders, we went around to the institution stuff, and we found they were paying not for one, but three different versions of Salesforce that didn't talk to each other, weren't integrated, and sales pipeline was getting dropped from one version to the next.
• 13:05
  SPEAKER_01: What we found was a billion and a half dollars of pipeline lost.
• 13:11
  SPEAKER_01: A billion and a half dollars.
• 13:13
  SPEAKER_01: We surfaced that, we highlighted it, we said, this is the problem.
• 13:16
  SPEAKER_01: To your point, here's the answer.
• 13:18
  SPEAKER_01: This is the problem.
• 13:18
  SPEAKER_01: Guess what happened?
• 13:19
  SPEAKER_01: We were fired as an agency because we made the VP of sales look bad.
• 13:23
  SPEAKER_02: Well, yeah.
• 13:24
  SPEAKER_01: You took the bullet because they could not admit that that was a they could not admit that they had lost a billion and a half dollars of pipeline by not having any kind of orchestration or coordination in their sales CRM.
• 13:35
  SPEAKER_01: Now, this was 20.
• 13:36
  SPEAKER_01: Chris, that was clearly your fault.
• 13:38
  SPEAKER_01: Absolutely.
• 13:39
  SPEAKER_01: That's clearly our fault.
• 13:40
  SPEAKER_01: This is 2016.
• 13:41
  SPEAKER_01: This was before Trust Insights, but this was an example of in the enterprise, it is just it is actually more important to navigate the people and the politics than it is the technology.
• 13:50
  SPEAKER_01: The technology is easy, it is the humans that are the hard part.
• 13:54
  SPEAKER_02: Yeah, I totally agree.
• 13:55
  SPEAKER_02: And and you know, these days, I think there's also the challenge of people not wanting to be caught without knowing the answer, right?
• 14:04
  SPEAKER_02: And in these, in this situation, there's a lot of unanswerable questions or questions that are very difficult to answer, and there's a lot of stuff that people don't know or that's still emerging, it's still being developed.
• 14:15
  SPEAKER_02: And so when you have senior leaders who are supposed to know the answer, but don't necessarily know the answer, don't understand the nuance, then the education is even more challenging, right?
• 14:27
  SPEAKER_02: It's more, it's trickier, there's more mind fields because you can't make someone look bad, right?
• 14:33
  SPEAKER_02: You can't ask, they had you have to be able to educate them through the side door by assuming that they know stuff, but level setting anyway, right?
• 14:41
  SPEAKER_02: But really they don't know the answer, right?
• 14:43
  SPEAKER_02: What do you think?
• 14:44
  SPEAKER_02: No, here's the thing.
• 14:45
  SPEAKER_01: I sat in this, uh I did a gen AI presentation for a stakeholder at an enterprise uh about a year and a half ago.
• 14:52
  SPEAKER_01: And his thing was, and it's always this, it's always a guy that has this perspective.
• 14:56
  SPEAKER_01: His big thing was failure is not an option.
• 14:58
  SPEAKER_01: You know, generative AI is the future, failure is not an option.
• 15:01
  SPEAKER_01: I'm like, dude, you know nothing about AI because failure is like you know, 95% of the time you're gonna have all sorts of crazy stuff happen on the way to production.
• 15:11
  SPEAKER_01: And the culture at this company was one of zero tolerance for mistakes, right?
• 15:16
  SPEAKER_01: That's what you so that that was embodying that that thing, as opposed to people who actually know how to learn.
• 15:21
  SPEAKER_01: Failure is not only an option, failure is mandatory.
• 15:24
  SPEAKER_01: You have to fall down to understand what to do.
• 15:27
  SPEAKER_01: And so if a company's culture does not have a culture that tolerates or even encourages failure, you ain't going anywhere with AI.
• 15:35
  SPEAKER_02: Yeah, you're screwed.
• 15:37
  SPEAKER_02: Because number one, any failures, which are really lessons to be learned in disguise, will not be acknowledged, will not be communicated, and therefore will not be embodied, will not be, you know, will not become part of the institutional learning.
• 15:53
  SPEAKER_02: And then the you know, the the uh the messengers usually get uh the bullet in their head and the organization doesn't move forward.
• 16:00
  SPEAKER_02: But what's also interesting is you get a lot of this uh lip service and a lot of the PR saying how great their digital transformation is and how much uh they're taking advantage of AI and how successful they are.
• 16:12
  SPEAKER_02: I saw this with Verizon years ago when I was not doing a project for Verizon.
• 16:17
  SPEAKER_02: They had threatened to do business with me multiple times.
• 16:20
  SPEAKER_02: My goodness, the the the politics were unbelievable.
• 16:23
  SPEAKER_02: But this is as a customer, I, as a customer, had uh, you know, an astoundingly bad experience.
• 16:29
  SPEAKER_02: And I wrote about it in an article in Customer Think, and I wrote about it first as a rant, right?
• 16:35
  SPEAKER_02: Like I started off with this company is awful.
• 16:38
  SPEAKER_02: Here are these terrible things that happened.
• 16:39
  SPEAKER_02: And as they went through iterations, I calmed it down, I calmed it down, and then I was neutral, and then I called it six fails and six fixes for the digital transformation.
• 16:48
  SPEAKER_02: But the big observation was everything they were putting out in their press, and everything they were putting out in their board reports, and everything they were putting out to the public touted how successful that digital transformation was and how that customer experience was was was phenomenal.
• 17:05
  SPEAKER_02: And when you talk to people on the ground, it's like nowhere near that.
• 17:09
  SPEAKER_02: And when you talk as a customer, it's a disaster.
• 17:12
  SPEAKER_02: Yet somehow that was communicated up the chain, and then whoever that digital transformation person was went on to their next gig, right?
• 17:22
  SPEAKER_02: It's just long enough not to get caught, it seems, right?
• 17:26
  SPEAKER_01: So here's a real simple example, a non-AI example.
• 17:28
  SPEAKER_01: Frontier Airlines is trying something out in Denver.
• 17:31
  SPEAKER_01: They are do undergoing this transformation to make themselves a more efficient airline.
• 17:36
  SPEAKER_01: And they're doing this by stop using jet bridges.
• 17:38
  SPEAKER_01: Everyone has to use the stairways now, which is faster.
• 17:41
  SPEAKER_01: They can turn around a plane in about half the time.
• 17:43
  SPEAKER_01: But if you are handicapped, elderly, or anotherwise disabled, you're kind of screwed.
• 17:47
  SPEAKER_01: And when you read the executives' uh words about this transformation, they're talking about this is going to improve profitability.
• 17:55
  SPEAKER_01: This is going to speed up efficiency.
• 17:57
  SPEAKER_01: This is going to do all these things.
• 17:58
  SPEAKER_01: And what's not in there is this is going to make the customer experience miserable because who the hell wants to walk up a flight of stairs over across the tarmac in Phoenix in July, right?
• 18:08
  SPEAKER_01: Nobody wants that.
• 18:09
  SPEAKER_01: And the problem with the way people are implementing AI in the enterprise is twofold.
• 18:14
  SPEAKER_01: One, they're doing it with a focus on the technology.
• 18:17
  SPEAKER_01: You never start with technology.
• 18:18
  SPEAKER_01: My CEO, Katie Robert, talks about this all the time.
• 18:21
  SPEAKER_01: You start with the problem and you look at the people and then you look at the process and you look at the outcome, and then you get to the technology.
• 18:27
  SPEAKER_01: And two, I saw this uh the other day on LinkedIn and it blew my mind.
• 18:31
  SPEAKER_01: This is executive who's saying, we are now a human light company.
• 18:35
  SPEAKER_01: All of our processes are human light.
• 18:37
  SPEAKER_01: We are we are more efficient and more productive.
• 18:38
  SPEAKER_01: Productivity is up 22%.
• 18:40
  SPEAKER_01: And I'm like, so you've made the customer experience worse, and you have employed fewer people.
• 18:47
  SPEAKER_01: At some point, dude, you are gonna run out of customers to sell to because no one's gonna do business with you and no one's gonna want to work with you.
• 18:53
  SPEAKER_02: Oh, that's crazy.
• 18:54
  SPEAKER_02: That's so interesting.
• 18:55
  SPEAKER_02: That's so interesting.
• 18:56
  SPEAKER_02: And you know, there's there's just this rush toward doing this, and without necessarily understanding that, you know, so much of what we do with AI, in my opinion, is to make up for our past sins in poor data and content curation.
• 19:13
  SPEAKER_01: Or worse, we just do our past sins bigger, better, faster, and cheaper.
• 19:17
  SPEAKER_02: True, true.
• 19:18
  SPEAKER_02: We go faster in the wrong direction, right?
• 19:21
  SPEAKER_02: Exactly.
• 19:22
  SPEAKER_02: Exactly.
• 19:23
  SPEAKER_02: We're automating the messes and we're continuing to fragment uh the data sources because one of the things I look at is when people are building these agent capabilities, they're building them many times with a repository or a functional or a process perspective, which actually is leading to more data fragmentation, right?
• 19:43
  SPEAKER_02: Because you end up with AI ops for this, AI ops for this particular content ops for this.
• 19:50
  SPEAKER_02: When you really need kind of a create once publish everywhere, which is something we had done at HP many, many years ago, that they're still using it, it's still in process and they save hundreds of millions of dollars.
• 20:02
  SPEAKER_02: But it's very, very few, there are very few companies that have that vision.
• 20:07
  SPEAKER_02: And what is your sense these days about kind of the connection to you know good old-fashioned data principles and data hygiene and data architecture and data quality with regard to generative AI and and retrieval augmented generation, right?
• 20:24
  SPEAKER_02: Because we know that the highest value of getting business value from AI is really about speeding up the internal information metabolism of the organization, right?
• 20:35
  SPEAKER_02: Getting people faster, better answers, getting them to data that they couldn't get to, getting them data they didn't have, getting the data that's faster, that's better quality, but it's accessing those that that information internally to make internal processes more work more effectively, which ultimately supports the customer.
• 20:54
  SPEAKER_02: I like to say you can't have acts of heroics upstream and expect a seamless customer experience, right?
• 20:59
  SPEAKER_02: Because acts of heroics don't scale.
• 21:01
  SPEAKER_02: So, what what is your sense of how they're looking at the boring, unsexy data governance, data quality, data curation, data hygiene, data architecture problems?
• 21:12
  SPEAKER_01: Well, as everyone who uses Microsoft Copilot tied to your internal SharePoint knows, Copilot looks at your internal data first.
• 21:19
  SPEAKER_01: And most people's SharePoint is like a hoarder's house that has never been cleaned out, and as a result, share uh copilot doesn't perform well.
• 21:25
  SPEAKER_01: And they're like, why doesn't copile why does copilot suck?
• 21:28
  SPEAKER_01: Because you got 20 years of crap in your SharePoint and it's referencing all of that.
• 21:31
  SPEAKER_01: But so there's there's one side which is yes, you absolutely need a solid foundation of good data for anything to work, not just AI, but literally anything to work.
• 21:41
  SPEAKER_01: But the bigger picture perspective that everyone's missing when it comes to AI is the difference between optimization and transformation.
• 21:47
  SPEAKER_01: Optimization is do what you've always done, but do it bigger, better, faster, cheaper.
• 21:50
  SPEAKER_01: That is not necessarily a good thing, going faster in the wrong direction, like you said.
• 21:55
  SPEAKER_01: Transformation or innovation is about doing something that you did not have the capable.
• 22:00
  SPEAKER_01: Capabilities to do before.
• 22:02
  SPEAKER_01: And AI is the great skill leveler in that it gives you skills.
• 22:06
  SPEAKER_01: I have zero, absolutely zero musical skills, right?
• 22:09
  SPEAKER_01: I cannot sing.
• 22:10
  SPEAKER_01: Like I can, but you don't want to hear it.
• 22:12
  SPEAKER_01: I can't play an instrument.
• 22:13
  SPEAKER_01: I can't read sheet music.
• 22:14
  SPEAKER_01: I can do none of that.
• 22:15
  SPEAKER_01: But what I can do is come up with ideas and things that I would like, and a machine can bring the skills to get me to do the thing.
• 22:23
  SPEAKER_01: Now, whether my idea is any good or not is a separate question, but at least the machine fills that gap.
• 22:28
  SPEAKER_01: So the big question for people to ask themselves is are we doing what we should be doing?
• 22:34
  SPEAKER_01: Not doing, you know, doing it bigger, bad, a fast cheap, but is this even what the customer wants?
• 22:39
  SPEAKER_01: Does the you know this is something I say all the time to content marketers because they always talk about AI, you know, oh, we can do this with Chat GPT.
• 22:45
  SPEAKER_01: I'm not saying, does anyone want another blog post?
• 22:48
  SPEAKER_01: The answer is probably no.
• 22:51
  SPEAKER_01: Go back, like this is the example I use all the time in my keynotes.
• 22:54
  SPEAKER_01: Go back to the 1930s, 1937.
• 22:57
  SPEAKER_01: There's this company in occupied Korea, because it was occupied by the Japanese during the war, called Mitsubishi.
• 23:02
  SPEAKER_01: And Mitsubishi imported and exported dried fish and seaweed.
• 23:06
  SPEAKER_01: And as technology changed and culture changed, they were able to stop using their imposed name and use the Korean pronunciation for their name and change and transform to a different kind of company.
• 23:17
  SPEAKER_01: Today, we know that company by its Korean name, Samsung, right?
• 23:21
  SPEAKER_01: And you will find no dried fish or seaweed anywhere in the Samsung store.
• 23:25
  SPEAKER_01: They have they keep the core of what they were good at, import export, right?
• 23:29
  SPEAKER_01: But now instead of fish, it's smartphones and technologies and refrigerators and all this stuff.
• 23:33
  SPEAKER_01: When we think about the enterprise and we think about digital transformation, that's almost the worst possible term you could use because it puts the technology first.
• 23:43
  SPEAKER_01: Think about transformation and then figure out where does the technology play a role?
• 23:48
  SPEAKER_01: Because, like we've said, we've talked about it's usually not the blocking obstacle.
• 23:52
  SPEAKER_01: It's human beings are usually the blocking obstacle, the cultures, the people defending their little corporate fiefdoms, like, oh, you can't take my budget, and I don't want to do this, and I don't want to talk to that department.
• 24:01
  SPEAKER_01: We had one client.
• 24:02
  SPEAKER_01: This is the most ridiculous thing in the world.
• 24:04
  SPEAKER_01: I don't want to lose my seaweed revenue.
• 24:06
  SPEAKER_01: Right.
• 24:07
  SPEAKER_01: We had we had a one client that they were so broken internally, their marketing department was not allowed to look at their sales department's performance.
• 24:16
  SPEAKER_01: Marketing and sales had separate data, they could not see each other's data.
• 24:20
  SPEAKER_01: And when we dug in and asked why, we found out because sales had this is B2B, sales had a 1.2% closing rate.
• 24:28
  SPEAKER_01: So marketing was sending thousands of leads every month, and sales could close none of them.
• 24:32
  SPEAKER_01: Like they literally could have had a Furby and a dog in the sales department.
• 24:35
  SPEAKER_01: It probably would have closed at about the same rate.
• 24:37
  SPEAKER_01: That that was that was not something technology would fix.
• 24:40
  SPEAKER_01: That is, you need to fire your VP of sales and and restaff.
• 24:44
  SPEAKER_02: Wow.
• 24:44
  SPEAKER_02: Yeah, that's really interesting when you think about where those dysfunctions lie and how so much of it is based on what you said.
• 24:52
  SPEAKER_02: Cover your ass and try not to, you know, keep your head high and don't take risks, right?
• 24:58
  SPEAKER_02: Especially if it's a risk of culture or if it's a culture that that penalizes failure, right?
• 25:04
  SPEAKER_01: Because you can't exactly that's what's killing AI in in enterprises, is you can't take a risk, so you can't innovate.
• 25:10
  SPEAKER_02: Yeah, if you can't take a risk because you're gonna be punished for failure, then you're not gonna get anywhere on this.
• 25:15
  SPEAKER_02: And that's a really interesting point.
• 25:17
  SPEAKER_02: Um, where would you say, you know, uh I mean, what what are you seeing?
• 25:22
  SPEAKER_02: We self-select a little bit in terms of our clients, right?
• 25:25
  SPEAKER_02: Like we we usually will get to know people in our conversations and we'll get to know if it's a good vibe, and you know, and if they're kind of grumpy or mean or they don't respond, or they, you know, you can pick up the the signals, right?
• 25:39
  SPEAKER_02: So we kind of self-select out of people that we don't have a good cultural match with.
• 25:45
  SPEAKER_02: And so I think that our culture match is aligned with trans, you know, transparency and you know uh uh uh you know a reliance on the facts and the data, and you know, it is what it is, right?
• 25:58
  SPEAKER_02: So we don't so but but then we uh have we may have other companies that are large enough where we're dealing with one part of the group which is great, and then they have a counterpart which is not so great, right?
• 26:10
  SPEAKER_02: And they end up being, you know, the ones that kind of throw the wrench into the works because they don't have the same values.
• 26:17
  SPEAKER_02: So what are you seeing in terms of your organizations, your customers?
• 26:21
  SPEAKER_02: Are you trying to self-select out of the ones that kind of clearly demonstrate, huh, these guys are not gonna get it because they're not, you know.
• 26:29
  SPEAKER_02: Do you have any litmus test for that in terms of a cultural fit?
• 26:33
  SPEAKER_02: And then what do you do when you kind of find that that that disparity amongst uh different groups?
• 26:39
  SPEAKER_02: Is that a fair question?
• 26:40
  SPEAKER_01: So my CEO Kate, no, no, my CEO Katie and I have a very good, very simple litmus test for this because she's the CEO, she is the leader of the company, she is an expert in her own right in so many things.
• 26:51
  SPEAKER_01: What I look for, what we listen for when we're talking to somebody, new client, new prospect, new new person in a role is how likely they are to talk over her.
• 27:00
  SPEAKER_01: And that is the immediate red flag, is when someone and it's on a call with both her and I to see if someone talks over her, but not me, right?
• 27:09
  SPEAKER_01: If they give precedence to someone because of gender, right?
• 27:13
  SPEAKER_01: That's an immediate red flag that tells us this person is going to be a problem.
• 27:16
  SPEAKER_01: Wow, that's so interesting.
• 27:18
  SPEAKER_02: We never thought of that.
• 27:19
  SPEAKER_02: I'm gonna have to listen for that.
• 27:20
  SPEAKER_01: Because you if you think about it, if a person has encoded biases, and we all do, if they have encoded biases that they are not mindful of, then you know that that's going to come out to play in other areas where they don't respect certain people.
• 27:35
  SPEAKER_01: And it it this has proven true for just about everybody that has failed this test.
• 27:40
  SPEAKER_01: When I talk to their subordinates and stuff, like they're miserable, it's a bad culture for people who are because like you know, everybody has a bad day and stuff like that.
• 27:49
  SPEAKER_01: Like you give people the benefit of the doubt first.
• 27:51
  SPEAKER_01: But the other thing that's important to do, and this is super important to use AI for get as much information about a stakeholder as you can, like get their LinkedIn profile, get your emails and stuff, put it into a secure system, right?
• 28:03
  SPEAKER_01: A system that is is protected.
• 28:04
  SPEAKER_01: Build a synthetic version of that person, and then you can ask them all the questions like, hey, you seemed resistant on our last call.
• 28:11
  SPEAKER_01: What are your what are the real needs that that you're not expressing that maybe you're afraid to express?
• 28:17
  SPEAKER_02: Fascinating.
• 28:17
  SPEAKER_02: That's fascinating.
• 28:18
  SPEAKER_02: I never thought of that.
• 28:20
  SPEAKER_02: That's a wonderful, wonderful use uh of AI that people would not really readily think about.
• 28:26
  SPEAKER_02: So if you can gather information and all your recordings and all your interactions and gather their LinkedIn stuff and gather their writings and anything publicly available, you can build that really robust.
• 28:39
  SPEAKER_02: I've also thought about that when you are in just a regular meeting, who talks over whom, right?
• 28:45
  SPEAKER_02: Because that is a disrespectful signal.
• 28:48
  SPEAKER_02: So if somebody talks over me, uh that's not a good sign, right?
• 28:52
  SPEAKER_02: And and vite and and conversely, I need to be aware of talking over other people.
• 28:57
  SPEAKER_02: So that's a really interesting uh signal.
• 29:00
  SPEAKER_02: I I like that.
• 29:01
  SPEAKER_02: Wow.
• 29:02
  SPEAKER_01: If you think about it, it's not that different from I mean, this has been this has been around you know, floating around leadership things for a while, but you know, look at and look at how someone treats uh a server at a restaurant, go out to lunch with them, you know, and you see how they treat uh someone who is theoretically in a position of lesser status, right?
• 29:18
  SPEAKER_01: How does that person treat that person?
• 29:20
  SPEAKER_02: Yes, yes.
• 29:21
  SPEAKER_02: I used to uh wait tables, so I always tips.
• 29:25
  SPEAKER_02: I never leave a poor tip.
• 29:27
  SPEAKER_02: I remember when I was uh working at Ground Round and I'd have 15 people at a table and I'd be taking these massive trays and I'd get a two dollar tip and I'd be like, ooh, that's not good.
• 29:37
  SPEAKER_02: You know, that's not yep.
• 29:40
  SPEAKER_01: But uh exactly.
• 29:41
  SPEAKER_01: And these are all human behavioral things that you can predict.
• 29:44
  SPEAKER_01: One of my I I did cover this in my newsletter this past weekend.
• 29:47
  SPEAKER_01: One of the things that like job seekers should be doing is using these tools to practice interviewing.
• 29:53
  SPEAKER_01: I had it at one point say, Okay, you're gonna play the role of an adversarial hiring manager.
• 29:58
  SPEAKER_01: Your job is to weed me out, your job is to be confrontational, your job is to pick at things.
• 30:02
  SPEAKER_01: And it I gave it a job description, I gave it my LinkedIn profile, and it wrote a question like, you know, you have a background in this, you don't appear even remotely qualified for this job.
• 30:10
  SPEAKER_01: Explain why I should even talk to you.
• 30:12
  SPEAKER_01: Like, can you imagine someone saying that in an interview?
• 30:15
  SPEAKER_01: Like, whoa! But if a machine asks you that and then you're ready for it, then if a human does that, you're you're ready with an answer.
• 30:23
  SPEAKER_02: Yeah, that's great.
• 30:24
  SPEAKER_02: I that's another really interesting scenario and use case.
• 30:27
  SPEAKER_02: I I really love that as well.
• 30:29
  SPEAKER_02: That would that's great for for everyone.
• 30:31
  SPEAKER_02: I'm gonna share that with one of my uh I support.
• 30:34
  SPEAKER_02: Um are you familiar with the Phoenix Academy uh in Massachusetts?
• 30:37
  SPEAKER_02: I am not.
• 30:38
  SPEAKER_02: It's a it's for underrepresented population.
• 30:41
  SPEAKER_02: Um they try to prepare them for college or work life of someone, and they're then asking me about ideas around AI, and I think that's a wonderful, perfect, perfect example.
• 30:51
  SPEAKER_02: So I'm gonna share that and give you attribution.
• 30:54
  SPEAKER_02: By the way, I um I have a part of a slide that I usually explain to people where I talk about, I call it my metrics layer cake, where I talk about, you know, everybody wants to know how something is going to improve uh uh revenue, but we're working in the weeds, but we can have process, you can have uh uh data scorecards, data supports the process, we can have process scorecards, process support and outcomes.
• 31:15
  SPEAKER_02: So we have business outcomes uh scorecards, and then those align with the strategy and we need the linkage up and down the food chain.
• 31:22
  SPEAKER_02: But I always say that at that process level, that's where you either get a bonus, or as my colleague Chris Penn likes to say, you catalyze a resume generating event.
• 31:33
  SPEAKER_01: Blah, blah.
• 31:34
  SPEAKER_01: There's two the last two issues of my newsletter.
• 31:36
  SPEAKER_01: The first is how do you find jobs using AI, using the deep research tools?
• 31:40
  SPEAKER_01: And then how do you land the job through the interviews?
• 31:42
  SPEAKER_01: And this Saturday, I'm gonna be doing what's the Sunday I'm gonna be going on how do you negotiate pay?
• 31:47
  SPEAKER_01: Because that's one of the parts that people hate the most.
• 31:49
  SPEAKER_02: Can you, Carolyn?
• 31:51
  SPEAKER_02: Let's make sure Carolyn has those newsletters in the show notes.
• 31:54
  SPEAKER_02: And I know you have a ton of different platforms and vehicles.
• 31:59
  SPEAKER_02: How do you keep up with all that stuff?
• 32:01
  SPEAKER_02: I mean, I'm sure you're using some of the technology for that, but you have Substack, you have Twitter, you have LinkedIn, you have Facebook, you have what else do you have?
• 32:09
  SPEAKER_02: You have a lot of stuff, right?
• 32:11
  SPEAKER_02: So how do you how do you syndicate out to those things?
• 32:14
  SPEAKER_02: Is it the same content that you just syndicate?
• 32:16
  SPEAKER_02: Are you varying the content?
• 32:18
  SPEAKER_02: How do you how do you handle all that?
• 32:19
  SPEAKER_02: It's a massive, massive output.
• 32:22
  SPEAKER_01: It's it is somewhat massive, but a lot of it is just uh capturing thinking.
• 32:26
  SPEAKER_01: Like that's the most important thing is having your voice memos app on your phone, like hotkeyed to a button so that you can immediately take a voice memo.
• 32:32
  SPEAKER_01: Um, my newsletter goes on Substack.
• 32:35
  SPEAKER_02: What do you use for your voice memos?
• 32:36
  SPEAKER_01: Do you just the built-in voice memos app?
• 32:39
  SPEAKER_01: So Substack is where my newsletter goes because that that handles the email distribution.
• 32:43
  SPEAKER_01: Every day I I try to write something for LinkedIn, that then gets put into a markdown file on my computer, and that is automatically then uh reprocessed and queued to my blog so that that content doesn't get lost.
• 32:54
  SPEAKER_01: And then uh it also goes, those those also go into Substack of Substack notes, which are different than than newsletters.
• 33:01
  SPEAKER_01: From I also have a backlog of almost 2,000 questions from conferences and events and things, QA sessions, webinars.
• 33:09
  SPEAKER_01: And so I record those as YouTube videos, just answering the question.
• 33:13
  SPEAKER_01: Um, that goes on the YouTube channel along with the video version of the newsletter, because I just read the newsletter out loud.
• 33:18
  SPEAKER_01: Um that those videos then the ones that do well then get turned into uh bigger pieces of content.
• 33:24
  SPEAKER_01: So the the way I think about it is you try to seed content and ideas around um in small pieces at first.
• 33:32
  SPEAKER_01: See how which ones take off, see which ones you know uh trigger something, and then you can turn you can make those bigger, right?
• 33:38
  SPEAKER_01: Because it's it's the the cost of failure.
• 33:40
  SPEAKER_01: Like I publish a post on LinkedIn that's 800 words and nobody likes it.
• 33:44
  SPEAKER_01: Okay, great.
• 33:44
  SPEAKER_01: That was not a huge loss of time, as opposed to spending four, five, six hours on a big newsletter that no one's gonna like.
• 33:51
  SPEAKER_02: Yeah, yeah.
• 33:51
  SPEAKER_02: No, that's a great, a great way to think about kind of triaging and seeing what floats to the top and what ideas are resonating with people.
• 33:59
  SPEAKER_02: It always surprises me.
• 34:00
  SPEAKER_02: Sometimes I just repost something from somebody and it gets a million views, not a million views, but it gets a lot of views.
• 34:06
  SPEAKER_02: Um, but uh it is interesting what resonates with different people.
• 34:10
  SPEAKER_02: One of one of my most secret productivity tools that I love is uh a tool called SuperKey.
• 34:16
  SPEAKER_02: We won't tell anyone else, it'll just be between us.
• 34:19
  SPEAKER_01: It's a tool I built called SuperKey.
• 34:21
  SPEAKER_01: So there's all these like voiced AI voice memo tools, and that they all cost$19.99 a month, and they all you don't know where your data is going.
• 34:28
  SPEAKER_01: So I built my own uh that runs on my computer, it doesn't need the internet.
• 34:32
  SPEAKER_01: And these kinds of tools are so handy where you push a button on your keyboard, hold it down, you speak, and then when you let go, it pastes what you just said into the the current window.
• 34:41
  SPEAKER_01: I am such a faster, faster better producer because I can sit there and just think, knock out a couple sentences, think again, you know, write a paragraph or two, or I can just hold down the button, just foam at the mouth or whatever.
• 34:54
  SPEAKER_01: But the those kinds of real-time, very quick translation tools are so useful.
• 34:59
  SPEAKER_02: So tell me again how you built it.
• 35:01
  SPEAKER_02: What is it?
• 35:02
  SPEAKER_02: What does it actually do?
• 35:04
  SPEAKER_01: So, what it does is it's a little app that's always running, and there's um I just there's a hotkey on the keyboard that I press.
• 35:10
  SPEAKER_01: When I press that, it starts listening to the microphone, it records the audio as long as I'm holding now the keys.
• 35:16
  SPEAKER_01: When I let go, it passes it through a transcription model.
• 35:18
  SPEAKER_01: Uh, I use NVIDIA's parakeet, which is free and open source, and then it takes the it takes the the transcribed text and just immediately pastes it into whatever window I'm in.
• 35:28
  SPEAKER_02: Oh, okay, okay, great.
• 35:29
  SPEAKER_02: And uh and uh the difference between that and using dictation, pressing the dictation button is pretty much just easier for you, and much better accuracy.
• 35:38
  SPEAKER_01: Nvidia's parakeet model has the lowest word error rate of any of the transcription models out there.
• 35:42
  SPEAKER_02: Great, great.
• 35:43
  SPEAKER_02: Yeah, and that's wonderful because your your ideas come and they go, right?
• 35:47
  SPEAKER_02: You don't remember later.
• 35:49
  SPEAKER_02: Like I always think to myself, oh, that's a great idea.
• 35:51
  SPEAKER_02: I'm gonna have to write about that later.
• 35:53
  SPEAKER_02: I guess what you never do.
• 35:54
  SPEAKER_02: Late later happens and it's like it's gone.
• 35:57
  SPEAKER_02: It's like a comic who is always writing their bits down, right?
• 36:00
  SPEAKER_02: You always have to capture those ideas when they happen because that's a stream of consciousness and awareness that's gonna go away.
• 36:08
  SPEAKER_02: So I really like that idea, and I need to do more of it.
• 36:11
  SPEAKER_02: I end up writing emails to myself when I that's an old way to do it.
• 36:16
  SPEAKER_02: So you're speaking at MACOM.
• 36:18
  SPEAKER_02: Tell us about MACOM that's coming up uh in a couple of months.
• 36:21
  SPEAKER_01: Yeah, uh it's gonna be what?
• 36:22
  SPEAKER_01: October 13th to 17th, I think.
• 36:24
  SPEAKER_01: It's uh it's in Cleveland, it's the marketing AI conference, Paul Retzer's event.
• 36:28
  SPEAKER_01: I've been a speaker at that event every year.
• 36:30
  SPEAKER_01: Uh in fact, I was the first keynote speaker of the first MAKON ever, uh, which is a a nice distinction.
• 36:36
  SPEAKER_01: And my my CEO, Katie Robert and I are doing a workshop on using generative AI for analytics, which is a very, very tangled session.
• 36:46
  SPEAKER_01: But we're gonna spend a couple hours walking through how you can use these tools to actually get correct answers with your data.
• 36:54
  SPEAKER_02: Right, right.
• 36:54
  SPEAKER_02: And I imagine there's some data normalization that needs to happen with that and reference architecture and so on.
• 37:00
  SPEAKER_02: So yeah, that's a that's a really great, a great topic.
• 37:03
  SPEAKER_02: And MACOM is actually where I got many of my case studies for my book five years ago.
• 37:10
  SPEAKER_02: So you'll recognize some people, Cal, I'll do and Jeff um uh Jeff uh uh from Market News and a couple of other coyle, yeah, yeah.
• 37:19
  SPEAKER_02: Jeff Coyle, yeah.
• 37:20
  SPEAKER_02: And uh I remember at a conference and I had zero expectations because everything at that time, this is 2019, right?
• 37:28
  SPEAKER_02: Um, there was no chat GPT.
• 37:30
  SPEAKER_02: There was no chat GPT, but every time I went to a conference and people talked about AI, it was all arm wavy, high-level, completely useless, right?
• 37:40
  SPEAKER_02: It was yeah, because no one can code.
• 37:42
  SPEAKER_02: What's that?
• 37:42
  SPEAKER_02: Yeah, because nobody was a coder back then.
• 37:44
  SPEAKER_02: Nobody knew, nobody knew, nobody had any idea.
• 37:47
  SPEAKER_02: And when I went to MACOM, I was like, wow, these people really know what they're talking about.
• 37:51
  SPEAKER_02: And um, I was very impressed.
• 37:53
  SPEAKER_02: And it's been and I know the institute has now grown quite a bit.
• 37:57
  SPEAKER_02: I know they spun off the marketing organization, and the conference is great, and highly, highly recommend it to folks.
• 38:05
  SPEAKER_02: Uh, do you have a discount code that you can share?
• 38:08
  SPEAKER_01: I do, I haven't the foggest idea what it is, so we'll have to find it after the show.
• 38:12
  SPEAKER_02: All right, so let's uh let's also post that in the show notes uh for that.
• 38:16
  SPEAKER_01: Carolyn has it.
• 38:18
  SPEAKER_01: It's N200.
• 38:21
  SPEAKER_02: Thank you, Carol.
• 38:22
  SPEAKER_02: So, what else do you think is happening?
• 38:24
  SPEAKER_02: Like we're hearing a lot about things like MCP model context protocol.
• 38:29
  SPEAKER_02: Give me give me five minutes on that topic because to me it's it's APIs for agents, right?
• 38:36
  SPEAKER_02: And agent context, but it's also has a lot of complexity because you're kind of abstracting what can be what is multidimensional intelligence, right?
• 38:46
  SPEAKER_02: You're talking about a vector space, and then you're taking kind of a projection of that, which gives you some explicit metadata, but there's a lot of implicit metadata behind the scenes, right?
• 38:55
  SPEAKER_02: You can't possibly capture the richness.
• 38:58
  SPEAKER_02: Like when you ask Chat GPT what it knows about you when you have long-term memory turned on, it'll say, oh, you're this, you're that, you're this, but that's a fraction of what it really knows because that vector representation is so much more complex and cannot be easily explained in explicit terms, right?
• 39:16
  SPEAKER_02: So when you think about model context protocol, you're kind of trying to do the same thing.
• 39:20
  SPEAKER_02: You're trying to pass parameters to a vector model or large language model that has infinitely more parameters than you're able to extrapolate or extract.
• 39:32
  SPEAKER_02: So, what are your thoughts about MCP and agent to agent and and really the the whole agent space?
• 39:39
  SPEAKER_02: Because I know that's highly overhyped at this point.
• 39:42
  SPEAKER_01: Well, so the agent space is really important because agents, uh, the agent space and the way people are handling agentic AI who actually get it to work, um, are recognizing that the language model is the engine of the car, right?
• 39:55
  SPEAKER_01: Nobody drives down the road on an engine.
• 39:57
  SPEAKER_01: Like you drive down the road in a car, so you kind of need the rest of the car.
• 40:00
  SPEAKER_01: You need seats and wheels and a steering wheel, maybe a radio, seat belt, etc.
• 40:05
  SPEAKER_01: And the agentic infrastructure around it is what allows you to put the engine into the rest of the car.
• 40:12
  SPEAKER_01: So that's in a nutshell, the where agentic is going is saying, okay, we're we're we're gonna give you the rest of the car.
• 40:18
  SPEAKER_01: So your language model can talk to your Gmail and you can say, Hey, what emails haven't I read today that I need to?
• 40:24
  SPEAKER_01: MCP is I simplify it's APIs for AI.
• 40:27
  SPEAKER_01: If you know what an API is, you know what an MCP is.
• 40:29
  SPEAKER_01: However, it is the field has gotten so varied that there you could find MCP servers for pretty much anything.
• 40:37
  SPEAKER_01: There's two things there to pay attention to.
• 40:40
  SPEAKER_01: One, most of them are deeply insecure.
• 40:42
  SPEAKER_01: So you better know where your data is going because when an AI model talks to an MCP server, it's sending the context.
• 40:48
  SPEAKER_01: That means your prompt, your documents, anything else.
• 40:51
  SPEAKER_01: And if it's your MCP service is being hosted by somebody you don't know or you don't trust, your data's leaving on a jet plane.
• 40:58
  SPEAKER_01: And two, the MCP space, people overhype and overcomplicate it.
• 41:03
  SPEAKER_01: It is not that complicated.
• 41:04
  SPEAKER_01: It is just building an app that can talk any specific protocol.
• 41:08
  SPEAKER_01: I've built MCP servers using fast MCP.
• 41:11
  SPEAKER_01: I built one, it's the dumbest MCP server in the world.
• 41:14
  SPEAKER_01: All it does is has a it's a Python entry to the word count utility.
• 41:18
  SPEAKER_01: And all it does is count words.
• 41:19
  SPEAKER_01: You say, use the word count MCP and you in Claude, use the word count MCP to actually count how many words are in this document.
• 41:25
  SPEAKER_01: And it calls the MCP server, it sends the post, the post comes back and it says, Yeah, that's 500 words.
• 41:30
  SPEAKER_01: So easy, so simple, but you go, that's what it's for.
• 41:32
  SPEAKER_01: It's for the things that AI models suck at.
• 41:36
  SPEAKER_01: The most powerful one that I use, that is my personal favorite, is called Serena.
• 41:40
  SPEAKER_01: Uh, Serena MCP is an open source, locally run, so no data leaves your computer, semantic layer for mainly for coding, but you can use it for anything that can very quickly traverse a whole code base or a document base and very quickly find and return relevant chunks without having to go through the whole rag embedding dance, right?
• 42:01
  SPEAKER_01: It just uses plain old text.
• 42:03
  SPEAKER_01: It is so cool, so powerful, it saves me.
• 42:06
  SPEAKER_01: It is literally saved my my my soul from being eroded by coding bugs because it can understand the context of an environment.
• 42:14
  SPEAKER_01: So if you've ever coded a big project, you know you've got to have a dictionary.
• 42:17
  SPEAKER_01: Here's the functions are, here's what the handles are, here's where all these pieces are.
• 42:21
  SPEAKER_01: And then you got to remember oh, I didn't name that function return value.
• 42:24
  SPEAKER_01: I mean it returned this value, like and and so this this style server can just do that.
• 42:29
  SPEAKER_01: Serena is something I would, if you are doing any kind of coding, you absolutely must use this MCP.
• 42:34
  SPEAKER_01: It's free, it's local, it your data never leaves, it's totally private, and it will save your brain.
• 42:40
  SPEAKER_02: Tell me again, what is it?
• 42:41
  SPEAKER_02: What is the output?
• 42:42
  SPEAKER_02: What is it doing for you?
• 42:44
  SPEAKER_02: It's it's reviewing your code, and then what is it producing for output?
• 42:48
  SPEAKER_01: So it spits back code chunks, it's or or file references.
• 42:52
  SPEAKER_01: So you can say, like, hey, I've got this error in my code, go fix it.
• 42:55
  SPEAKER_01: And it will say, Let me look at the error.
• 42:57
  SPEAKER_01: I see what files mentioned, let me go look in the files for those functions, and then it sends back to the language model, hey, here's the function that we think is bad.
• 43:04
  SPEAKER_02: Gotcha.
• 43:04
  SPEAKER_02: Okay.
• 43:05
  SPEAKER_02: Immensely helpful, immensely valuable.
• 43:08
  SPEAKER_01: It's like a little librarian for your project.
• 43:10
  SPEAKER_02: Right, right, right.
• 43:11
  SPEAKER_02: So um, interesting stuff.
• 43:13
  SPEAKER_02: Uh, we're right, we're getting to the end of the uh the time, but um, what do you see as kind of the most important things for organizations to look at before coming into Q4?
• 43:24
  SPEAKER_02: Rest of the year uh in 2025, and then what's on tap for 2026?
• 43:28
  SPEAKER_02: What do you see as kind of the the realizations, the understanding, the kind of back to basics?
• 43:34
  SPEAKER_02: What do you see as kind of the fundamentals for what organizations are going to be doing or should be doing uh moving forward in this next uh several quarters?
• 43:44
  SPEAKER_01: So that's a big question.
• 43:45
  SPEAKER_01: Uh, in terms of where organizations are going, they need to figure out whether or not they should be using AI, right?
• 43:51
  SPEAKER_01: Um, and on a on a project-by-project basis.
• 43:54
  SPEAKER_01: Like, is this a problem that is well suited for AI?
• 43:59
  SPEAKER_01: That's a that that question is getting missed because people are like, we got to be AI first, we gotta be AI for we gotta do AI for this.
• 44:04
  SPEAKER_01: Like, that's like saying blenders, we are all in on blenders, everything is blenders.
• 44:08
  SPEAKER_01: Yeah, but you're a sushi restaurant.
• 44:10
  SPEAKER_01: You know, nobody wants blended sushi.
• 44:12
  SPEAKER_01: No one wants that.
• 44:13
  SPEAKER_01: And so it's it's it's taking the technology and making the technology take a back seat.
• 44:17
  SPEAKER_01: Again, this is something my CEO Katie Robert talks about all the time.
• 44:20
  SPEAKER_01: Put the technology in the backseat, figure out where you're going first before you figure out what car you want to drive, right?
• 44:26
  SPEAKER_01: That's super important.
• 44:27
  SPEAKER_01: The technology itself, agentic is is exploding in terms of capabilities and risks and things.
• 44:32
  SPEAKER_01: So you better know what you're doing.
• 44:34
  SPEAKER_01: I strongly recommend that people pay attention to the free open source scene when it comes to uh anything agentic because you want code and services and servers that are under your control that that uh adhere to the rest of your standards.
• 44:47
  SPEAKER_01: And then in terms of the models themselves, uh, we are hitting diminishing returns on the Transformers model.
• 44:53
  SPEAKER_01: I mean, that's that's a known uh we have already hit some serious issues with RAG, return retrieval augmented generation.
• 45:00
  SPEAKER_01: However, what is happening that is very interesting is everything that's coming out of China, the Chinese labs are coming out with some amazing innovations because, ironically, because of all the export restrictions, they've had to get really clever with math to work around the fact that they have less powerful chips.
• 45:18
  SPEAKER_01: So you see Alibaba's Quen family, phenomenal set of models.
• 45:22
  SPEAKER_01: Kimi K2 Instruct, Deep Seek, all these labs have some really good models and they are super cheap to use.
• 45:28
  SPEAKER_01: And you can use them with like US-based providers, so you don't have to worry about like sending your data off to China.
• 45:34
  SPEAKER_01: We use a company uh called Deep Infra.
• 45:36
  SPEAKER_01: And man, you can get like top state-of-the-art performance for like a tenth of state-of-the-art prices.
• 45:43
  SPEAKER_02: Wow, that's incredible.
• 45:44
  SPEAKER_02: Hey, Chris, this has been wonderful.
• 45:46
  SPEAKER_02: I will see you at MAKON.
• 45:47
  SPEAKER_02: Uh I mean, I know Chris is going to your session, but uh looking forward to seeing you there.
• 45:52
  SPEAKER_02: Highly recommend the conference, it's a great event.
• 45:55
  SPEAKER_02: I remember the uh code that you'll see in the show notes.
• 45:58
  SPEAKER_02: Take a look at Chris's newsletters.
• 46:00
  SPEAKER_01: I'll have some book signings too.
• 46:01
  SPEAKER_01: I got a new book.
• 46:02
  SPEAKER_01: A new book.
• 46:03
  SPEAKER_01: It's called Almost Timeless 48 Foundation Principles of Generative AI, which you can get at AI formarketersbook.com.
• 46:10
  SPEAKER_02: I love it.
• 46:10
  SPEAKER_02: I love it.
• 46:11
  SPEAKER_02: That looks great.
• 46:12
  SPEAKER_02: You write books faster than I can read them.
• 46:14
  SPEAKER_02: So anyway, Chris, this has been wonderful.
• 46:16
  SPEAKER_02: Thank you again uh for your time and participation today.
• 46:19
  SPEAKER_01: Thanks for having me.
• 46:20
  SPEAKER_02: And thanks uh for everyone listening.
• 46:22
  SPEAKER_02: Thank you, Carolyn, for setting this up.
• 46:24
  SPEAKER_02: And we will see you next time on the Early AI podcast.
• 46:28
  SPEAKER_02: Bye now.
• 46:30
  SPEAKER_00: Thank you for joining us on another deep dive into AI innovation.
• 46:34
  SPEAKER_00: Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world.
• 46:41
  SPEAKER_00: The Early AI Podcast is sponsored by Early Information Science and Vector.
• 46:46
  SPEAKER_00: To learn more, visit early.com.
• 46:48
  SPEAKER_00: That's E A R L E Y dot com.
• 46:51
  SPEAKER_00: Thank you for tuning in.