Latest / The Adults In The AI Room / Moritz Müller on Structuring Content for Enhanced Language Model Capabilities - The Earley AI Podcast with Seth Earley - Episode #047

Transcript

Moritz Müller on Structuring Content for Enhanced Language Model Capabilities - The Earley AI Podcast with Seth Earley - Episode #047 · The Adults In The AI Room

• 0:01
  SPEAKER_00: This is the Early AI Podcast, hosted by Seth Early and Chris Featherstone.
• 0:06
  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.
• 0:15
  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 CMS Wire.
• 0:25
  SPEAKER_00: Now, enjoy the show.
• 0:32
  SPEAKER_01: Welcome to the Early AI Podcast.
• 0:34
  SPEAKER_01: I'm Seth Early, and I'm really excited to introduce our guest today to discuss how emerging technologies like retrieval augmented generation will significantly impact organizations by improving information access.
• 0:50
  SPEAKER_01: We're also going to talk about risks around the hype, around the risks that hype large language models and what that poses for an organization if you don't adequately manage and differentiate leveraging your proprietary data, your intellectual property.
• 1:07
  SPEAKER_01: We're going to be talking about the balancing of the use of large language models with oversight for utilizing those proprietary data sources and knowledge sources.
• 1:18
  SPEAKER_01: And then how utilizing metadata for context is crucial to personalized experiences and improving information access.
• 1:25
  SPEAKER_01: And we'll talk about the evolving role of AI in repetitive task automation that will reshape career paths and managing autonomous agents.
• 1:34
  SPEAKER_01: So our guest today has been working on digital transformation programs for many years and heads product management for search and generative AI technologies.
• 1:43
  SPEAKER_01: He has been working with corporate clients and financial services manufacturing and the public sector in APAC.
• 1:50
  SPEAKER_01: He has a PhD in geophysics and has experience in oil and gas exploration and multiple years of experience with IT project implementation.
• 1:59
  SPEAKER_01: He brings a wealth of experience in enterprise search and product leadership at Squirrel, his current company, his and his technical expertise drives the understanding of technology at a very fundamental level.
• 2:11
  SPEAKER_01: Moritz Mueller, please welcome to the show.
• 2:15
  SPEAKER_02: Thanks a lot.
• 2:16
  SPEAKER_02: It's a pleasure to be here.
• 2:17
  SPEAKER_02: And I'm looking forward to speak with you about what we do at Squirrel and how we leverage, especially lately, the retrieval augmented technology to deliver state-of-the-art enterprises.
• 2:28
  SPEAKER_01: And we don't want to be a commercial about Squirrel, but we'll talk about Squirrel.
• 2:31
  SPEAKER_01: So we'll talk about concepts.
• 2:33
  SPEAKER_01: And then, yes, we do want to know about Squirrel because Squirrel seems to be a really amazing technology.
• 2:38
  SPEAKER_01: So I'm very impressed by that.
• 2:39
  SPEAKER_01: But I just want to let people know we're we're not doing a commercial here.
• 2:42
  SPEAKER_01: It's really about the knowledge and the functional.
• 2:44
  SPEAKER_01: But first, I'd like to start with what are the typical fallacies, the typical misunderstandings and myths that are out there that you run into on a day-to-day basis with executives, with colleagues, with customers, and so on.
• 2:59
  SPEAKER_02: That's a very good question.
• 3:01
  SPEAKER_02: So the standard question that I get when we talk about Gen AI, uh combining LLMs with retrieval augmented generation, then is how do I train the model on my data?
• 3:12
  SPEAKER_02: That's probably this the first question that we got when ChatGPT came up last year.
• 3:16
  SPEAKER_02: It's a question that we still get.
• 3:18
  SPEAKER_02: And I understand that people ask that question, but I do not think it really reflects the challenges of using Gen AI and especially the retrieval augmented generation technology for your enterprise and organization.
• 3:34
  SPEAKER_02: The challenge is much, much more on how you can combine the search and your own data with the respective capabilities, Gen AI capabilities of a large language model.
• 3:44
  SPEAKER_02: And that's one of them.
• 3:46
  SPEAKER_02: The other one of them, I think people often misunderstand the limitations of LLMs.
• 3:54
  SPEAKER_02: Um, so for instance, a lot of people always talk about is my data secure?
• 3:58
  SPEAKER_02: They're very worried, is my data going out to an API?
• 4:01
  SPEAKER_02: Someone is doing something with my data.
• 4:04
  SPEAKER_02: If you look at this, and specifically if you look at the example of ChatGPT, security-wise, I really think this is 100% safe.
• 4:12
  SPEAKER_02: There is absolutely no question about this.
• 4:13
  SPEAKER_02: We are so used to using uh the online resources of Microsoft like Office 365, SharePoint Online, the ChatGPT models that we use in the private cloud is exactly on the same kind of systems.
• 4:26
  SPEAKER_02: So there is absolutely no risk.
• 4:28
  SPEAKER_02: The bigger challenge is do you actually have a large language model available and not only ChatGPT, but potentially something else to use that in your environment?
• 4:38
  SPEAKER_02: That is the much bigger challenge.
• 4:40
  SPEAKER_01: Right, right.
• 4:41
  SPEAKER_01: And so just for folks who may be less familiar with the terminology, when you say training the LLM, uh define the different ways of training and then just talk about retrieval argumentary generation for a moment.
• 4:54
  SPEAKER_01: Like, what's how is that paradigm different than simply asking a question of an LLM?
• 5:00
  SPEAKER_01: So first question is training.
• 5:02
  SPEAKER_01: What does training mean?
• 5:03
  SPEAKER_01: And what are the different ways of thinking about training?
• 5:05
  SPEAKER_01: Because some people think that means you have to train the LLM itself, right?
• 5:11
  SPEAKER_01: Whereas others you think about tuning, and tuning is based on the data.
• 5:15
  SPEAKER_01: So it's a similar concept, but it's very confusing.
• 5:18
  SPEAKER_01: So can you clarify that for folks?
• 5:21
  SPEAKER_01: Yeah, that's that's a very good question.
• 5:23
  SPEAKER_02: And I start often with that.
• 5:25
  SPEAKER_02: So for me, the whole talk about training comes from our historic experience.
• 5:31
  SPEAKER_02: Whenever we talked about machine learning models or NLP models, natural language processing models, these models typically needed to be trained or they were trained for your specific data.
• 5:40
  SPEAKER_02: So for for doing data classification or for doing the task that they were supposed to be doing with your data.
• 5:47
  SPEAKER_02: Um and for that, you trained the model specifically, also a chatbot, for instance, a machine learning model for chatbot that was used.
• 5:52
  SPEAKER_02: You trained the model to give good answers with for that specific use case.
• 5:58
  SPEAKER_02: With the emergence of large language models, and I don't want to dive into the details of how complicated it is to actually train a large language model.
• 6:07
  SPEAKER_02: As I think the ChatGPT, you can break it down the training of ChatGPT, for instance, in five different stages, but it's super complex, and there is so much data being used for the training, then the limited amount of training that you would do on your data would not fundamentally change the output of the LLM.
• 6:26
  SPEAKER_02: So that's why when we talk about training in the past, very valid question.
• 6:29
  SPEAKER_02: The moment we look at these pure dimensions of these large language models, the whole conversation actually shifted on how can we tune it, as you said, how can we configure it so that it gives us the best results, specifically for the use case that we are addressing.
• 6:45
  SPEAKER_02: And the way to do this is twofold.
• 6:47
  SPEAKER_02: Um, one is as you mentioned, you can look at how we can make the whole retrieval piece better, how can we retrieve the information from an underlying search system, for instance, to provide the right contextual information.
• 7:00
  SPEAKER_02: That's one challenge.
• 7:01
  SPEAKER_02: And the other challenge is, and you probably heard about this as well, is the whole prompting challenge.
• 7:06
  SPEAKER_02: How can we make sure that we um configure the prompting in such a way?
• 7:10
  SPEAKER_02: So we basically, when we talk about prompting, it is all about giving the right instructions to the large language model to tell the model what is it that we want to achieve.
• 7:20
  SPEAKER_01: Yes, that's very interesting.
• 7:22
  SPEAKER_01: And you know, our next webinar actually is on the alignment of prompt engineering and knowledge engineering.
• 7:28
  SPEAKER_01: So we're going to be covering that coming up soon.
• 7:30
  SPEAKER_01: And I'd love to pick your brain more on that, uh, whether we do it today or we do it a little bit later.
• 7:35
  SPEAKER_01: But you know, I know it gets very complicated, but can you give a high-level understanding or overview of those five stages of training that you mentioned?
• 7:46
  SPEAKER_01: Because I have done some research and reading about large language models and the feedback mechanisms and just how unbelievably complex and sophisticated it is.
• 7:57
  SPEAKER_01: I mean, it it really is amazing.
• 7:59
  SPEAKER_01: The the just the self-referential continual feedback, you know, adjusting weights, adjusting the model as it's processing is just absolutely phenomenal.
• 8:10
  SPEAKER_01: But you want to just give a high-level understanding of of what that what those elements are?
• 8:14
  SPEAKER_01: It's okay if we dive a little bit into the weeds.
• 8:17
  SPEAKER_01: People who don't want to hear that can just fast forward.
• 8:21
  SPEAKER_02: Um that's fair, but I also need to be very honest with you.
• 8:24
  SPEAKER_02: So I myself am not an expert of training LLM on.
• 8:28
  SPEAKER_02: So I think I can break it down, um, give you a high-level understanding.
• 8:33
  SPEAKER_02: But for all the details, you really need to speak to an expert.
• 8:35
  SPEAKER_02: It wouldn't be fair if I if I if I were speaking for that.
• 8:38
  SPEAKER_02: So the way um, if you look at uh Chat GPT, the way it works, um the different stages, and again, five is not necessarily the final number, but you have different stages on how you train the model.
• 8:50
  SPEAKER_02: You train an initial model where you throw in a lot of content and you train a so-called base model um in the first version, and that can be maybe on a high level, just the amount is much, much larger compared with the original training of machine learning model.
• 9:06
  SPEAKER_02: Um, so you train the model on all the tax that you can get hold of.
• 9:10
  SPEAKER_02: You need a lot of GPU power to do this.
• 9:13
  SPEAKER_02: Uh, someone told me that um when uh I think Facebook Facebook uh trained their their first LLL, their first proper LLM, they spend 60 million dollars or so just on buying the GPU units to train at initial model.
• 9:26
  SPEAKER_02: So it's a it's it's a crazy investment that tells you already it's it's utterly impossible for an organization to just do it.
• 9:33
  SPEAKER_02: So that's the first step.
• 9:35
  SPEAKER_02: Um in the second step, you you and the first step is I think they call it pre-training, and then you do a so-called um supervised fine-tuning where you uh try to make sure that the general the general understanding is there.
• 9:50
  SPEAKER_02: So you you use certain prompts to um to teach the model certain outcomes, and that gets then refined more in the terms of a rewarding model, so that you basically tell the model you ask them some questions, and that's a manual manual process, but you can also use some technologies to do this, um, where you ask some questions, and it's the feedback learning where you say this was right or this was not right.
• 10:11
  SPEAKER_02: Um, and the more of this training you can do, the better, of course, it gets.
• 10:15
  SPEAKER_02: And that is also the stage where you later on can do the fine-tuning.
• 10:19
  SPEAKER_02: Um, you basically have a reward model to make sure as the last layer on top of your of your trained LLM, uh, you do more of an uh enforcement learning for specific use cases or kind of questions.
• 10:34
  SPEAKER_02: So that is the additional layer that you can set on top of this.
• 10:37
  SPEAKER_02: But overall, this whole process process is super complex.
• 10:40
  SPEAKER_02: Yes.
• 10:40
  SPEAKER_02: And again, not me not being an LLM expert, but what I heard from ChatGPT Chat GPT, I think they were also the surprise themselves that their ChatGPT model worked so well.
• 10:51
  SPEAKER_02: And it took them quite a while to understand which of these steps and how the combination of the stacks that's actually led to to being so successful and giving so good results.
• 11:01
  SPEAKER_01: Yeah, no, that's a great that's a great overview.
• 11:03
  SPEAKER_01: And one of the things people forget is that there is a lot of manual training, there's a lot of human-in-the-loop feedback, there are a lot of people making judgments about these armies and armies of people who are doing that.
• 11:16
  SPEAKER_01: And then when they start red teaming to try to break the model or get it to do something bad, then that's even more human in the loop.
• 11:24
  SPEAKER_01: So so that's so interesting about training.
• 11:27
  SPEAKER_01: Now, when you think about fine tuning, and this is where you know we talk about training, there's a lot of different ways to think about training, but I think I think more about fine tuning, and fine tuning is really how you're going to access your data with the LLM, and that comes into retrieval augmented generation.
• 11:45
  SPEAKER_01: Do you want to just give a quick definition of RAG and uh and and then what has to be done to the data to make it readily available for the LLM using um RAG?
• 12:01
  SPEAKER_02: Yeah, that's a very good question.
• 12:03
  SPEAKER_02: So if you look at LN's as such, it is very impressive if you go to uh ChatGPT.
• 12:10
  SPEAKER_02: I think we have all tried it ourselves, and we asked the ChatGPT a generic question or general question.
• 12:16
  SPEAKER_02: I I even used it myself uh to uh, for instance, write some small routines in Python or get inspiration again um to yeah, because my Python days uh they are they're passed by a few years, but I'm still trying from time to time, and it's impressive what the model can do.
• 12:34
  SPEAKER_02: The moment it becomes specific, however, it is very challenging to source the right information.
• 12:42
  SPEAKER_02: And the moment we talk about enterprise rack in that sense, the task of the Gen AI technology is to give an answer that is specific to your organization or specific to your use case, and the general knowledge that the model is trained on likely does not contain that piece of information that you're looking for that is required to give you the answer.
• 13:03
  SPEAKER_02: So, in order to overcome this, when we talk about retrieval augmented generation, we pair a so-called information retrieval technology with the capabilities of Gen AI, large language model, to give us an answer on a limited amount of data.
• 13:22
  SPEAKER_02: So, what I mean with this is you can with a large language model provide a so-called prompt or context to it.
• 13:31
  SPEAKER_02: You can basically ask the large language model saying, This is my question.
• 13:35
  SPEAKER_02: I have a question, for instance.
• 13:37
  SPEAKER_02: Um, let's say in work with some organizations here, they have certain procurement guidelines.
• 13:42
  SPEAKER_02: And the question is then um, up to which sum?
• 13:45
  SPEAKER_02: Can I directly give out, for instance, um a contract or a purchase order to someone?
• 13:51
  SPEAKER_02: And up to when do I need to do a tender or something?
• 13:54
  SPEAKER_02: Now, this information is specific to the organization.
• 13:57
  SPEAKER_02: If you just ask the LLM, it will not exactly know what you're asking for.
• 14:02
  SPEAKER_02: But what you can do then is you can provide within the so-called token limit.
• 14:06
  SPEAKER_02: So this is this prompting context.
• 14:08
  SPEAKER_02: You can provide up to well, I think ChatGPT at the moment has 10 to 20 pages of text, roughly.
• 14:14
  SPEAKER_02: That's the way to look at this.
• 14:16
  SPEAKER_02: So you could provide 10 to 20 pages of text of context.
• 14:19
  SPEAKER_02: You say, This is my 10 pages with my procurement guidelines, and you can ask the model, say, This was my question, these are my guidelines.
• 14:28
  SPEAKER_02: Please answer the question based on the guidelines that I provide.
• 14:32
  SPEAKER_02: And that is what the LLM is able to do.
• 14:35
  SPEAKER_02: It is very good in doing this, it will then basically abstract the knowledge in these 10 pages and give you that specific answer form within.
• 14:44
  SPEAKER_02: Now, if you just look at this as the capability of the LLM with Jen AI to give you an answer, and this is very powerful.
• 14:53
  SPEAKER_02: Now, however, for now and in the near future, we have a limit of this token limit.
• 14:58
  SPEAKER_02: So you need to make sure you can put the right information in these 10 to 20 pages of text that you pass in this context.
• 15:06
  SPEAKER_02: Now, if you want to use the capability on top of, for instance, your knowledge base, let's say you have a lot of documents on a SharePoint or in any kind of other Google Drive system that you have here in your organization, 10 to 20 pages is probably not enough.
• 15:22
  SPEAKER_02: Right?
• 15:22
  SPEAKER_02: You are very limited in terms of that.
• 15:25
  SPEAKER_02: The way to overcome this is by doing first a search.
• 15:29
  SPEAKER_02: So that's why we use the information retrieval step.
• 15:32
  SPEAKER_02: We first conduct a search with your question and try to find the relevant documents, and not only the documents, we actually try to find the relevant pages inside of your document that contain the likely answer to your question.
• 15:49
  SPEAKER_02: And with this, we then put together basically the 10-20 pages of context.
• 15:54
  SPEAKER_02: We say we take the top 10 results of our search engine, we always take one page of text, and then we put that together at 10 to 20 pages of context, and we pass that to the large language model.
• 16:06
  SPEAKER_02: And then the LLM is basically able to give us an answer based on all the relevant pages that we have found in all of your knowledge space.
• 16:14
  SPEAKER_02: And wrapping this up, this is basically the combination of semantic search on one side where you can search through millions of documents, finding the right pages, and then you pass them on to the LLM for the generation piece to summarize your nice answer based on these relevant pages that you have in it.
• 16:33
  SPEAKER_01: That's great.
• 16:34
  SPEAKER_01: That's great.
• 16:34
  SPEAKER_01: And and when organizations uh when I've I've spoken recently with some organizations who've tried this, right?
• 16:42
  SPEAKER_01: And they've just pointed it uh to uh all of their policy documents.
• 16:48
  SPEAKER_01: And I'm not sure what repositories there were, but they they got an answer.
• 16:52
  SPEAKER_01: Uh it was about lunch, um what the reimbursement policy was for lunch and what the limit was.
• 16:59
  SPEAKER_01: And the LLM returned back$180,000.
• 17:04
  SPEAKER_01: And it was also for Singapore of all places.
• 17:08
  SPEAKER_01: And then they tried it again, and it was no matter where they they, you know, how they tried to define the area, it still came back with$180,000.
• 17:16
  SPEAKER_01: Now, clearly something went wrong.
• 17:19
  SPEAKER_01: And when you start thinking about the repository, the quality of the data, you know, why you would get an erroneous answer like that.
• 17:27
  SPEAKER_01: Can you troubleshoot that a little bit?
• 17:29
  SPEAKER_01: Can you talk a little bit about what makes good retrieval?
• 17:33
  SPEAKER_01: Right?
• 17:33
  SPEAKER_01: When you start thinking about semantic search, search, um, parametric search, metadata search, retrieval in general, what makes good retrieval?
• 17:43
  SPEAKER_01: And then what do you have to think about with the data and the content?
• 17:49
  SPEAKER_02: Yes, happy to answer that.
• 17:51
  SPEAKER_02: Um before I exactly give you an answer to that question, let me make a general statement.
• 17:58
  SPEAKER_02: I personally believe that in order to make retrieval augmented generation successful and useful, the challenge that we will face in the next years, where you will see all big organizations in the world they will use the technology, they will use it on top of their internal documents and knowledge base because it's simply so powerful and can save so much money, it's an absolute no-brainer.
• 18:21
  SPEAKER_02: But the challenge to do this is much more on making the information retrieval work well than improving and tuning the LLM.
• 18:31
  SPEAKER_02: And that is something people I'm trying to explain to people.
• 18:34
  SPEAKER_02: The LLM, if you really give it the right context in this 10 pages of text, and 10 pages of text for me is just a symbolic thing for the bigger models, this can be more, right?
• 18:43
  SPEAKER_02: But if you provide the right information to the LLM, they are really good in giving you an answer.
• 18:49
  SPEAKER_02: So the accuracies there are well above 90%, from what I've seen.
• 18:53
  SPEAKER_02: Now, the challenge is, as you mentioned, in terms of reimbursement policy, uh, if you just point the information retrieval to all of your documents, the challenge is how do you find the most relevant document with the answer?
• 19:08
  SPEAKER_02: So ideally, you point it to a most recent document.
• 19:13
  SPEAKER_02: Uh, what are we seeing in a lot of organizations is they have each year a new iteration of the document, and each document is each previous iteration is stored in their knowledge base.
• 19:22
  SPEAKER_02: So you want to have a certain time waiting factor in your information retreat.
• 19:27
  SPEAKER_02: That's one of the things.
• 19:29
  SPEAKER_02: And on top of this, we also use a lot of other metadata attributes that we, for instance, say everything that is older than one year gets a massive decay in the search functionality.
• 19:41
  SPEAKER_02: Everything that is not labeled, and you can add those labels as a knowledge-based document or that is pre-approved by someone, is not a relevant or most relevant search document.
• 19:52
  SPEAKER_02: And now imagine you have you work in a company that has been around for 50 years.
• 19:56
  SPEAKER_02: There will be tons of documents talking about reimbursements and maybe different versions, people save their own documents.
• 20:04
  SPEAKER_02: So the challenge is really: can you limit initially the amount of data that you search through to the knowledge base that is pre-approved?
• 20:13
  SPEAKER_02: And because once you do this, you will make sure that you get a much better context that you pass to the LLM to actually give you an answer.
• 20:22
  SPEAKER_02: And that is our way of making a tree block minute generation successful in any kind of organization.
• 20:29
  SPEAKER_01: That's great.
• 20:30
  SPEAKER_01: And and you know, it's funny, it it comes still comes back to the old problem.
• 20:34
  SPEAKER_01: This is where people are misunderstanding the whole uh class of technology.
• 20:40
  SPEAKER_01: It comes back to the old problem of context and and information uh hygiene and knowledge curation, right?
• 20:49
  SPEAKER_01: Because you can't give it, you can't point it to crap and get something back that's going to be useful, and you can't just point it to everything because there's no context.
• 20:58
  SPEAKER_01: So context is hugely important.
• 21:01
  SPEAKER_01: And this is where you know you're familiar with my phrase that I I coined years ago.
• 21:07
  SPEAKER_01: There's no AI without IA.
• 21:09
  SPEAKER_01: This is kind of the the exemplar of that.
• 21:11
  SPEAKER_01: Why don't you explain that concept uh in this context?
• 21:15
  SPEAKER_01: Because you really did just explain it, but it it's a good way to summarize it.
• 21:19
  SPEAKER_01: But you want to just provide your interpretation?
• 21:22
  SPEAKER_02: So a basic interpretation of this is you need to make sure in a big organization you have at least a minimum level of data hygiene and knowledge management.
• 21:34
  SPEAKER_02: Let me put it maybe like this.
• 21:36
  SPEAKER_02: Um, and it starts with every folder, as as stupid as it sounds, but every folder name that has that is called policy already or HR policy is relevant metadata information for search engine that we use.
• 21:52
  SPEAKER_02: Now, I'm not asking to go back to the Occo old days where everything was in folders.
• 21:58
  SPEAKER_02: It is much more the question.
• 22:00
  SPEAKER_02: Uh, how can we make sure this metadata is stored on the respective document?
• 22:05
  SPEAKER_02: Um and you will find big organizations nowadays, they have knowledge management departments even.
• 22:10
  SPEAKER_02: What I've seen is I've seen organizations here in Singapore, uh, the monetary authority of Singapore, they have a huge amount of people, 50 people working just on knowledge management.
• 22:19
  SPEAKER_02: Now, I'm not saying you need this to be successful, but if you can make sure that the the knowledge base, the main knowledge base of data is maintained in a certain way that you remove all the information, you archive it, or you tag documents that are pre-approved, saying this is our guide, this gets boosted in the search.
• 22:44
  SPEAKER_02: Then with this basic data hygiene, as I would call it, you make sure that not only the search is better, but also if you use retrieval documentary generation, your chatbot basically that you give to your employees or to your users finds you the best answer that you can get.
• 23:00
  SPEAKER_01: Yeah, and and and so the the whole idea of information architecture is structuring that knowledge and that content in a way that allows you to retrieve it.
• 23:11
  SPEAKER_01: And when you think about what is so incredible about LLMs is they interpret the question in a very in the query in a very sophisticated way.
• 23:22
  SPEAKER_01: And then they interpret and process the result in a very sophisticated way to make it more conversational.
• 23:28
  SPEAKER_01: And in between is that retrieval part, right?
• 23:31
  SPEAKER_01: Is that using your knowledge base, your database, your content.
• 23:35
  SPEAKER_01: And when when you just try to use a large language model by itself to answer a question, what does it tend to do?
• 23:42
  SPEAKER_01: If it doesn't have the answer, it'll make it up, right?
• 23:44
  SPEAKER_01: Unless you are very specific in your instructions to say only use the information from this data source.
• 23:50
  SPEAKER_01: And if you don't have that information, say I don't know, and then turning the temperature down to zero so it doesn't get too creative.
• 23:58
  SPEAKER_01: One is very creative, zero is not creative.
• 24:00
  SPEAKER_01: And so, you know, we've seen this work very, very well.
• 24:04
  SPEAKER_01: Uh, and it comes back to again thinking about content.
• 24:07
  SPEAKER_01: You know, in the old days, people would say, well, make it like Google, make it like Google, just make it like Google.
• 24:13
  SPEAKER_01: And I and I would say if you spent as much time and money and resources on fine-tuning your content that uh organizations do who want to get good rankings in Google, it would be like Google, right?
• 24:30
  SPEAKER_01: Because it is about that curation, it is about that architecture, it is about that terminology.
• 24:35
  SPEAKER_01: So it comes back to some of the old things.
• 24:38
  SPEAKER_01: It's this is not a magic book, but it's an amazing tool.
• 24:41
  SPEAKER_01: It's an amazing tool.
• 24:43
  SPEAKER_01: I mean, just like when you think about a chat bot and you have to, you know, interpret intents from utterances.
• 24:51
  SPEAKER_01: And utterances, you know, I forgot my password, my username doesn't work, I'm locked out of my computer, my computer's mad at me.
• 24:58
  SPEAKER_01: All right.
• 24:59
  SPEAKER_01: I just, and it, but it all comes down to the same intent, change password.
• 25:03
  SPEAKER_01: Well, large language models do that on steroids, right?
• 25:06
  SPEAKER_01: They take very sophisticated questions, and then they either they can expand those prompts or they can reconcile those prompts, but then they ask a question of the data source and using the parameters.
• 25:18
  SPEAKER_01: This is why prompt engineering, I think, is so important, because prompt engineering has to contain elements of that architecture, right?
• 25:26
  SPEAKER_01: You have to train people how to ask the right questions and use the right concepts.
• 25:31
  SPEAKER_01: And then you're retrieving either and you're either you do, oh and we'll talk about this, you're retrieving from that data source using a semantic search or using some kind of parametric or metadata search or other signals, and then you're taking that result and you're making it more uh human understandable, more conversational, right?
• 25:51
  SPEAKER_01: That's what the beauty of this is, that's what the power of this is.
• 25:54
  SPEAKER_01: It's the in-between part that we still have to get right.
• 25:57
  SPEAKER_01: But the beauty of this thing is you don't have to, you know, formulate every response.
• 26:02
  SPEAKER_01: What I did like about the early days of chatbots is we looked at use cases and the specific content you needed for those use cases.
• 26:12
  SPEAKER_01: So when I have an example and I just found it on my phone where I tried to find out how to activate a credit card, right?
• 26:20
  SPEAKER_01: And I lost the little sticker on the front.
• 26:22
  SPEAKER_01: And I am going with their chatbot, and it's giving me pages and pages and pages of information.
• 26:30
  SPEAKER_01: No matter how many ways I just asked a question, it would give me policies about this.
• 26:34
  SPEAKER_01: And policy, all I wanted was the phone number to call.
• 26:37
  SPEAKER_01: Now, if somebody had created that use case, they could have made that content very specific.
• 26:42
  SPEAKER_01: So, so this is where, again, we have to focus on those use cases.
• 26:45
  SPEAKER_01: We have to focus on personas, we have to focus on the additional signals, right, that tell us what an individual wants.
• 26:53
  SPEAKER_01: You know, it's funny.
• 26:54
  SPEAKER_01: On um, I listened to Hard Fork.
• 26:57
  SPEAKER_01: Uh, have you ever listened to that podcast, Hard Fork?
• 26:59
  SPEAKER_01: And they're talking about the fact that Bard came out and was being very woke and very politically correct, and it wasn't giving pictures of white people and all this stuff, right?
• 27:08
  SPEAKER_01: It tried to overcorrect to not be biased about you know everything being a white person, right?
• 27:13
  SPEAKER_01: A white male or whatever.
• 27:15
  SPEAKER_01: And and essentially, um, you know, what what they said was, what persona do you want barred to answer the question is, right?
• 27:26
  SPEAKER_01: Is it was it barred or is no, it's Gemini.
• 27:28
  SPEAKER_01: I'm sorry, it's Gemini, Gemini.
• 27:31
  SPEAKER_01: What what persona, what perspective do you want it to be, right?
• 27:34
  SPEAKER_01: Like what is the user's perspective?
• 27:36
  SPEAKER_01: What is the user's orientation or mindset or cap or or uh preference?
• 27:43
  SPEAKER_01: And that comes from additional signaling, additional metadata about the user.
• 27:47
  SPEAKER_01: Do you want to talk a little bit about that whole idea and the idea of really doing personalization based on the user?
• 27:54
  SPEAKER_01: Like we all walk around with a metadata cloud around us, right?
• 27:58
  SPEAKER_01: And that metadata cloud tells us everything, you know, tells can tell the world what we want and who we are, and we change our roles.
• 28:06
  SPEAKER_01: And so, you know, we talk about high fidelity journey models that kind of represent that user intent in metadata terms.
• 28:12
  SPEAKER_01: Do you want to talk about that and how that really leads to true uh personalization?
• 28:17
  SPEAKER_01: How does that fit in with all of this?
• 28:20
  SPEAKER_02: Absolutely, absolutely.
• 28:21
  SPEAKER_02: So the first part of personalization, and I think that is important to mention in the enterprise and organization context, is you have in any organizations, you have so-called ACLs, access control layers.
• 28:36
  SPEAKER_02: That means that you're probably all familiar with this.
• 28:39
  SPEAKER_02: Uh, you have on a on a SharePoint, Google Drive, or whatever, you have only access to certain folders.
• 28:44
  SPEAKER_02: There's certain information, let's say at HR department, they will have information about salaries or contracts that is not for everyone's eyes.
• 28:51
  SPEAKER_02: Now we need to make sure when we use retrieval augmented generation that we only provide an answer based on the documents that you are allowed to see.
• 29:04
  SPEAKER_02: So, how do we do this?
• 29:06
  SPEAKER_02: The way to do this is really we honor all the so-called entitlements that are exists.
• 29:12
  SPEAKER_02: That means we know who you are when you log in.
• 29:15
  SPEAKER_02: Um we know you are working in this department, you have access to those files.
• 29:19
  SPEAKER_02: And then whenever we do the search for information, we apply these so-called ACLs and make sure you only get to see documents that you are entitled to see.
• 29:28
  SPEAKER_02: That that is for me one crucial part of personalization in the enterprise context.
• 29:34
  SPEAKER_02: Now, adding to this, of course, you also want to have an answer that is more personalized.
• 29:42
  SPEAKER_02: Now, a second layer of personalization can be introduced when we introduce so-called knowledge graphs or knowledge systems.
• 29:50
  SPEAKER_02: So you can in big organizations have goes along with the whole knowledge management, as I mentioned, the taxonomy handling.
• 29:58
  SPEAKER_02: You understand who is this person, who what is his role.
• 30:02
  SPEAKER_02: Uh, we have had an example where, for instance, uh, in a bank, an ESG investor, um, a CIO, and a normal banker, they will have different kinds of results that they get back depending on the profile that they are that they have.
• 30:18
  SPEAKER_02: Um, someone searches for for certain information, and then we use the profile information that we have, so we understand who is this person, which department is in, and we have a similar kind of tagging because of the knowledge graph, we have a similar kind of tagging of the data on as metadata on the data, and then we boost the results differently depending on in which role you're in.
• 30:41
  SPEAKER_02: So that's one of the key things that we do, and that's one of the key aspects of personalization.
• 30:45
  SPEAKER_02: And other pieces, and I believe in the future there is a lot of development in this, is that you start LLMs to ask return questions because I mean, very much if you ask an expert about something, you ask him a question, he will probably ask you a question back, right?
• 31:00
  SPEAKER_02: To focus the content much more exactly in a in a way in a conversation to focus as much more on the specific domain that you are interested in.
• 31:09
  SPEAKER_02: Yep.
• 31:10
  SPEAKER_02: Let's say you want to buy a house, you ask someone, oh, I want to buy a house, where should I buy it?
• 31:14
  SPEAKER_02: Person will ask you, Oh, uh, what are your preferences?
• 31:16
  SPEAKER_02: Uh, you look for big or small house.
• 31:19
  SPEAKER_01: Um, are you more interested in two-age children?
• 31:22
  SPEAKER_01: Right.
• 31:22
  SPEAKER_02: There you go.
• 31:23
  SPEAKER_02: There you go.
• 31:24
  SPEAKER_02: That is the whole personalization aspect.
• 31:26
  SPEAKER_02: And I believe we get that in by applying additional filters on the data.
• 31:32
  SPEAKER_02: At the end of the day, these are just filters.
• 31:34
  SPEAKER_02: Yep.
• 31:34
  SPEAKER_02: It's a matter of questions of how we apply those filters.
• 31:38
  SPEAKER_02: Yep.
• 31:39
  SPEAKER_02: Um, if you if you take it a look at this and you abstract the whole process, is you apply a filter during your search, such that your initial document set.
• 31:48
  SPEAKER_02: So you combine a semantic search, which is a word vector, a vector-based search, with a metadata filter.
• 31:55
  SPEAKER_02: And combined, they give you a weighting of the search results that hopefully or I in the ideal work, we know it works, it gives you a better search ranking for the individual person.
• 32:05
  SPEAKER_02: But that is how we achieve uh a better degree of personalization.
• 32:09
  SPEAKER_01: Right.
• 32:09
  SPEAKER_01: And and I call that their digital body language, right?
• 32:13
  SPEAKER_01: What are all the signals that they're throwing off as you interact with the different tools and systems?
• 32:18
  SPEAKER_01: They're all storing metadata about campaign responses and interests and level of technical proficiency and equipment that they're all that they own and all of that, just like a salesperson or good service person would ask those questions, right?
• 32:33
  SPEAKER_01: I do think search is a conversation, especially the old days of metadata base search, right?
• 32:38
  SPEAKER_01: When you and when you or facet search, when you when you put in a quarry, that's ambiguous.
• 32:42
  SPEAKER_01: If you went to the Granger site and you typed in tools, imagine if you walked up to a counter at a hardware store and you said tools, they'd say, okay, what kind of tools do you want?
• 32:52
  SPEAKER_01: Right.
• 32:53
  SPEAKER_01: And that's what it does with the facets.
• 32:55
  SPEAKER_01: It says, what kind of tools do you want?
• 32:56
  SPEAKER_01: Here, power tools, NAND tools, et cetera.
• 32:58
  SPEAKER_01: And we're doing the same thing when we're asking those additional clarifying questions, right?
• 33:03
  SPEAKER_01: And that's a really important piece because I think chat-based applications are going to increasingly rely on that prompt engineering to be able to identify the the metadata and the entities and the metadata that's going to that's going to make that answer more relevant, just as you're explaining.
• 33:24
  SPEAKER_01: So this is great.
• 33:25
  SPEAKER_01: It's exactly um how I thought of this in the past.
• 33:28
  SPEAKER_01: And and this is shows you how knowledge graphs come into being uh come into play, and how you can use uh a customer identity graph or employee identity graph that includes you know things like what they're provisioned for, what they have rights to, and so on.
• 33:44
  SPEAKER_02: You can also maybe maybe Seth, or just one comment is so what we are seeing at the moment in our daily work, um, we're combining the we're building in a second layer of logic.
• 33:56
  SPEAKER_02: So we when you ask a question, we are on a first NLP model that tries to decide um which uh which information retrieval um processing line basically the command should go down.
• 34:08
  SPEAKER_02: So if you ask a text question that talks about procurement policies, it retrieves a document.
• 34:13
  SPEAKER_02: When you ask for a stock price, we point you back to a structured database that draws you a graph.
• 34:20
  SPEAKER_02: So that you do not get tax as an answer, but you actually get a graph as an answer and a visualization where you have the information in.
• 34:28
  SPEAKER_02: So the future will definitely see that these retrieval augmented generation technologies will become more and more complex and connect many different systems inside of an organization.
• 34:37
  SPEAKER_02: That's great.
• 34:38
  SPEAKER_01: So, you know, I do want to give you a chance.
• 34:40
  SPEAKER_01: I said this isn't a commercial about Squirrel, but I do want to give you a chance to talk about Squirrel because it is such an amazing tool and we're partners uh with Squirrel and starting to uh leverage it and want to introduce it to our customers.
• 34:54
  SPEAKER_01: So talk a little bit about, you know, kind of how you see Squirrel in the context of other environments and other platforms and other tools.
• 35:04
  SPEAKER_01: What are the real differentiators that you guys have?
• 35:07
  SPEAKER_01: I think one of them is that you're able to use well uh different language models.
• 35:13
  SPEAKER_01: Uh, is that correct?
• 35:14
  SPEAKER_01: And then uh tell me some of the other ones.
• 35:16
  SPEAKER_01: Give me some of the other, it's a very, very flexible, powerful tool, and you have the ability to do all this configuration, especially of those uh intents and the signals and the different types of roles and so on.
• 35:29
  SPEAKER_01: So talk a little bit more about uh Squirrel as search engine and as about uh and and about the integration with LLMs.
• 35:37
  SPEAKER_02: Yeah, I'm so happy to do this.
• 35:40
  SPEAKER_02: So if you look at us as a company, we have been around for quite a while now, uh over a decade, and we have really started off as a search company.
• 35:47
  SPEAKER_02: So Squirrel originally was considered a search engine, the platform.
• 35:51
  SPEAKER_02: In the meantime, we are in the Garner Magic Quadrant for so-called insights engine, as and also again uh an evolution of the terms.
• 35:58
  SPEAKER_02: So insights engine is the magic term.
• 36:01
  SPEAKER_02: Uh, what is in the term insights engine for me is that you use at the end of the day a lot of analytic technologies and you bring them together.
• 36:09
  SPEAKER_02: So, because you have so much data in an organization that you want to make sure your your employees, so the smart planes in your organization, they find as fast as possible the relevant information for them.
• 36:21
  SPEAKER_02: That is what this is all about.
• 36:23
  SPEAKER_02: Now, for us at SQL, as I said, we have been doing semantic search for quite a while now, so we really know how to do search.
• 36:31
  SPEAKER_02: And if someone asks me, oh, everyone talks out there about using ChatGPT, about using LLMs.
• 36:37
  SPEAKER_02: What is special about SQL?
• 36:39
  SPEAKER_02: I think the key answer that, and it potentially needs a technical person to understand that, but the key answer is that all of the solutions out there that just say they use ChatGPT, they are very much limited by this token limit at the moment.
• 36:52
  SPEAKER_02: So they can prepare 10, 20 pages of text, and they allow you to chat with these 10, 20 pages of text.
• 36:59
  SPEAKER_02: The moment it comes to your knowledge base, where you have more than just one document with 20 pages of text, or even an annual report where we have ESG questions on annual reports where you have 100 pages of text, these tools will not help you because they need first to do the so-called information retrieval.
• 37:18
  SPEAKER_02: And I think that puts us in quite a unique spot because we can nowadays ship out of the box a platform that does all the so-called tokenization and embedding generation that you need for vector models for the semantic search, and we give you a native integration of any LLM that you want to use.
• 37:36
  SPEAKER_02: And why is that so important?
• 37:38
  SPEAKER_02: Chat GPT is there to stay.
• 37:40
  SPEAKER_02: ChatGPT has a lot of advantages.
• 37:42
  SPEAKER_02: The big advantage of ChatGPT is it will be probably for quite a while be one of the cheapest LSM solutions out there.
• 37:48
  SPEAKER_02: That is something very, very clear.
• 37:50
  SPEAKER_02: But there will be other LLMs, and when you run uh retrieval augmented RUC framework, as we call it, you want to have a bit of flexibility.
• 37:59
  SPEAKER_02: With Squirrel, you can switch out the LLM with any other LLM that's out there.
• 38:04
  SPEAKER_02: So we are completely agnostic to the large language model that we use.
• 38:07
  SPEAKER_02: We nowadays even run it on premise.
• 38:10
  SPEAKER_02: So we use, for instance, for MISTRALL, I think the model is called MISTRAL, one small model that we can run fully on-premise.
• 38:16
  SPEAKER_02: I mean, I'm not saying this comes for free, it comes with a significant amount of costs associated with it, but we can do it fully on premise.
• 38:24
  SPEAKER_02: So for the work that we do, for instance, with our big banking clients, um, and some of these clients that are heavily regulated, they really need us to look beyond ChatGPT because the compliance is still struggling with uh approving that for some of the sensitive information that they have.
• 38:40
  SPEAKER_02: And I think um really because we have this decade of knowledge of how to do the information retrieval, how to do the semantic search.
• 38:49
  SPEAKER_02: That is the reason why we are in such a prime spot on when you want to do this on top of your own knowledge base.
• 38:54
  SPEAKER_02: And maybe just as a last step there, if I talk to clients here, everyone talks about Microsoft Copilot.
• 39:00
  SPEAKER_02: Microsoft Copilot is great, right?
• 39:02
  SPEAKER_02: Great tool.
• 39:03
  SPEAKER_02: But if you want nowadays to do retrieval augmented generation, so question answering on your SharePoint, for instance, Microsoft is not giving you an out-of-the-box product at the moment.
• 39:14
  SPEAKER_02: They're not, they will tell you here are the 20 building blocks you need to put this together, so go and try it yourself.
• 39:20
  SPEAKER_02: And that is the key advantage of Esquiro.
• 39:22
  SPEAKER_02: You can just basically buy it off the shelf.
• 39:25
  SPEAKER_02: You say, I want to do question answering on all of my documents, enterprise grade security, entitlement handling out of the box.
• 39:32
  SPEAKER_01: We we are able to provide.
• 39:34
  SPEAKER_01: That's uh that's great.
• 39:35
  SPEAKER_01: That's a huge advantage.
• 39:36
  SPEAKER_01: Now, when you think about metadata, we did an experiment with a life sciences company where we built a knowledge architecture, we componentized the content, we ingested it into an LLM, and we ingested it both without the knowledge architecture and with the knowledge architecture to compare.
• 39:55
  SPEAKER_01: And we found that when we had those enriched embeddings uh in the LLM, we were able to get much, much better results from 53% accuracy without enrichment in the metadata to 83%.
• 40:09
  SPEAKER_01: So when you think about Squirrel and how Squirrel is doing this, you're doing you are doing vector-based search, yes?
• 40:16
  SPEAKER_01: So, and then you're ingesting that content with the additional signals of whatever metadata is on those documents.
• 40:23
  SPEAKER_01: That's great.
• 40:24
  SPEAKER_02: That's that's yeah, but but maybe just to look at this, so semantic search alone will only find you the relevant paragraph inside of the document.
• 40:34
  SPEAKER_02: So let's say you ask a question about uh how to change the battery in the car, and you have all the manuals for cars loaded, uh then the then this semantic search will find you from all the manuals, likely the paragraphs that talk about how to change the battery in a car.
• 40:52
  SPEAKER_02: Now, that alone, if you have a Ford and you want to get the result for Mercedes, it will not help you.
• 41:01
  SPEAKER_02: So the way to improve this is combining, as you stated, the semantic search with the so-called metadata search.
• 41:08
  SPEAKER_02: That means when we store a document, we tag it, we say this document is related to this specific Ford model, this document is related to this specific car model, like a Mercedes BMW or whatever.
• 41:21
  SPEAKER_02: And we store that metadata on the respective items when we index them.
• 41:26
  SPEAKER_02: Okay.
• 41:26
  SPEAKER_02: Now, when you ask the question, you basically ask how to change the battery in my Mercedes.
• 41:32
  SPEAKER_02: Then we use technologies that understand oh, you want this car model.
• 41:37
  SPEAKER_02: So I first run a filter, like a faceted search with that respective metadata to filter down all the amount that you sort of data that you search in.
• 41:47
  SPEAKER_02: And in a secondary step, you then run the semantic search to find the respective paragraphs in the fitting document.
• 41:54
  SPEAKER_02: And that's how we combine the so-called metadata search with the semantic search to make sure you get the most relevant output after doing it.
• 42:01
  SPEAKER_01: That's interesting.
• 42:02
  SPEAKER_01: But you could also say that the metadata with the document would be additional signals for the LLM, but you're doing it in a stepwise fashion.
• 42:12
  SPEAKER_01: You're filtering first and then using the LLM.
• 42:14
  SPEAKER_01: But could it conceivably be done the other way, where you're saying I'm just going to ingest this and I'm going to use that, those those embeddings, I'm going to enrich those embeddings with additional metadata and then use that with a vector similarity search?
• 42:28
  SPEAKER_01: That's, I know that's a very technical question.
• 42:31
  SPEAKER_02: You can also do that, right?
• 42:33
  SPEAKER_02: Um, you could, of course, also do this.
• 42:35
  SPEAKER_02: It's just um so if you look at vector search, it becomes so we we do this, we add to, for instance, typically when we index the individual paragraphs, we decided we always want to index uh if you look at a document index, you always add the the document title as part of the embedding.
• 42:54
  SPEAKER_02: The same you can do with metadata.
• 42:56
  SPEAKER_02: Um I'm just not convinced that it solves the problem because if you, as I mentioned before, in Rug, the challenge is the so-called token limit that you have, the context.
• 43:05
  SPEAKER_02: So if you generate if you find now a lot of embeddings that talk about these kind of things, you still need to limit it to these 10 pages of text.
• 43:15
  SPEAKER_02: So if you get the right result with the semantic search approach or by combining semantic search with a metadata approach, at the end of the day, it does not matter as long as you can make sure you find the right paragraphs.
• 43:26
  SPEAKER_01: Understood.
• 43:26
  SPEAKER_01: Totally, totally get it.
• 43:28
  SPEAKER_01: And we we're also thinking in the fact that you have to componentize content, you know, because you because when you bring it into an LLM, you kind of break it up into arbitrary, you know, frames and you have overlap between those frames for context, but that doesn't always give you the context.
• 43:43
  SPEAKER_01: You know, the example I use is if you have an error code for a particular product.
• 43:48
  SPEAKER_01: Well, if you just said to a service tech, I have an error code of 92, right?
• 43:53
  SPEAKER_01: Well, okay, what do you own?
• 43:55
  SPEAKER_01: What what modem?
• 43:56
  SPEAKER_01: What what's the configuration, et cetera?
• 43:57
  SPEAKER_01: And not and that would have to be on that piece of
• 44:00
  SPEAKER_02: content right that that contextual metadata right so then in that in that case yes you could do it filtering it first right doing a parametric search a retrieval and then you could use the LLM so I could see how both those ways would work so we're almost out of time here but I wanted to give a couple of questions to learn a little bit more about who you are you are located in Singapore and uh do you want to talk a little bit about how you got to what you're doing and uh you know I know you you worked in in oil and gas exploration I know you have a PhD how did all that kind of impact what you're doing now that's a story out like yeah so I mean I mean the the the main part is five to eight minutes yeah happy happy to speak about this so I worked in oil and gas after my my my bachelor and and masters and I really enjoyed this it's exciting oil and gas is technology wise one of the most advanced fields that we have as as humanity as you it's it's really amazing and if you look at it nowadays um a large part of our economy is still driven by it so really really amazing but then I decided look I want to do a PhD I went back to university I appreciated this I learned a lot about the techniques um and then I I think it was a phase where the oil price was down I decided I didn't want to go back into a cyclical business so I started and at the time in Zurich uh I started to go into consulting work and I ended up at an AI company that's how it often goes I'm sorry what company so I first worked for a consulting company um uh in in Switzerland and then I switched um uh I switched at some point over to Squirrel um AI company again very exciting for me and when we had our first project in Singapore I came here to to lead this project because it was a lighthouse customer for us a very big project and then I committed um at the very beginning of 2020 I committed myself to open up an office in Singapore and I flew in one week before the covert lockdown to Singapore and that was a hell of a ride then for one and a half years.
• 45:59
  SPEAKER_02: Wow and but I don't have any requests I really like the place it is fantastic to to see um wow to see Singapore and also how our technology here um is is evolving.
• 46:13
  SPEAKER_02: I think um there's a lot of potential in this area I've never been to Singapore I would love to visit and um you're always welcome yes yes yes yes and uh you you uh I've forgotten do you have a family out there no uh I I don't have a family here but then pretty sure they would also enjoy it otherwise um so um everyone that comes and visits me only complains that it is basically warm every single day but I but but I really like it.
• 46:40
  SPEAKER_02: Yeah yeah yeah I do miss the the Swiss winters I have to admit that a little bit of snow once in a while but yeah um you can get that during Christmas break.
• 46:49
  SPEAKER_01: Yeah in New England we're we're getting we're to the point where it's warm a lot of the time a lot more than the winter sports people want it to be but and then what else do you do outside of uh work what what's fun for you I uh love to do cycling here in Singapore road cycling is it's quite a big community um but you have to go out early to to to escape and and beat the the sun in that sense and also a bit of traffic.
• 47:14
  SPEAKER_02: Right right so what time do you go out?
• 47:18
  SPEAKER_02: Uh we normally go very early at six in the morning on a on a weekend so it's it you continue an early start to the to the weekend.
• 47:24
  SPEAKER_02: Yeah and then um yeah you end with a with a typically with a coffee and a good local breakfast uh at nine in the morning and then you have had your workout when when many other people are just about to start their day.
• 47:35
  SPEAKER_01: Yeah that's wonderful that's a great way to start your day uh I'm not a big morning workout person and I had to take some time off because of some um medical stuff that I had to deal with but I do know when I make myself work out in the morning it's it's fantastic and it's a great way to start the day.
• 47:52
  SPEAKER_01: Well listen this has been really wonderful you know I'd like to ask one final question of folks you know when you when you kind of think about your life and you think about what you've done if you went back to yourself and could talk to yourself when you got out of college is there some piece of advice you would have given yourself based on what you know now yes so one of the things that I would have done earlier is I think I would have come to Asia before.
• 48:20
  SPEAKER_02: So I first time I came to Asia I was already in my 30s um I definitely would have loved to come here early and also learn Chinese that's one of the things I would have told myself learn Chinese because in this area in this region Chinese is is thriving a lot of things even so the business language is English and on the relationship side it helps to open a lot of doors not only in China but in all the other regions around the world.
• 48:43
  SPEAKER_02: Wow um and the other thing I would tell everyone um really also myself just just follow what you want to do um just just try out technology and and you see it with that hype of chat GPT right now there is there's always going to be something new with technology.
• 48:59
  SPEAKER_02: So just the moment you're interested in follow it and and think about how you can potentially have a business or career even out of it.
• 49:09
  SPEAKER_02: That is something that I would recommend to everyone be be passionate about it right as as long as you're passionate about what you do uh you really enjoy it um and the moment I I would lose my passion I would really have to reconsider what I'm doing.
• 49:23
  SPEAKER_01: That's wonderful and we'll put uh your contact information your LinkedIn information um uh uh is there do you have a simple LinkedIn um handle that you can mention right now what is your LinkedIn um uh you will find me um uh with Moritz B Miller I think you should find me in in LinkedIn um uh that should be um uh in there but if you also you also find me with Morris Miller and Scarrow in LinkedIn so if you want to reach out just discuss the triple augmented generation how it works in enterprise setup um please feel free reach out i i really think this is a huge topic for the next it's a huge two to five years it will it will revolutionize the the whole uh um the whole way we we handle data and and you know all the consultancies and the analysts and everyone is saying that it's going to have the biggest impact of any AI initiative in large organization in any organization retrieval augmented generation and the market uh for uh for rag is predicted to be it was what 13 billion dollars in 2023 it's predicted to be 183 billion dollars in 2027 so yes we're at the beginning of that hockey stick and uh it is an exciting exciting time well moritz thank you so much for being here it's really been a pleasure i really appreciated your time thank you so and thank you to our audience and uh we will see you next time on our next podcast for early the early um uh ai podcast again thanks everyone and thank you moritz thank you for joining us on another deep dive into ai innovation tune in next time when we introduce another industry expert and discuss how to maximize AI in your world the early AI podcast is sponsored by early information science and CMSwire to learn more visit early.com that is early comes for tuning in