# AI Engineering with Chip Huyen

An episode of [The Pragmatic Engineer](https://invisiblenetwork.ai/podcast/1769051199), 5 Feb 2025, 1 h 15 min.

Audio: https://api.substack.com/feed/podcast/156423504/32526cd5e29dcf014cbe945452876810.mp3

Supported by Our Partners • Swarmia — The engineering intelligence platform for modern software organizations. • Graphite — The AI developer productivity platform. • Vanta — Automate compliance and simplify security with Vanta. — On today’s episode of The Pragmatic Engineer, I’m joined by Chip Huyen, a computer scientist, author of the freshly published O’Reilly book AI Engineering, and an expert in applied machine learning. Chip has worked as a researcher at Netflix, was a core developer at NVIDIA (building NeMo, NVIDIA’s GenAI framework), and co-founded Claypot AI. She also taught Machine Learning at Stanford University. In this conversation, we dive into the evolving field of AI Engineering and explore key insights from Chip’s book, including: • How AI Engineering differs from Machine Learning Engineering • Why fine-tuning is usually not a tactic you’ll want (or need) to use • The spectrum of solutions to customer support problems – some not even involving AI! • The challenges of LLM evals (evaluations) • Why project-based learning is valuable—but even better when paired with structured learning • Exciting potential use cases for AI in education and entertainment • And more! — Timestamps (00:00) Intro (01:31) A quick overview of AI Engineering (05:00) How Chip ensured her book stays current amidst the rapid advancements in AI (09:50) A definition of AI Engineering and how it differs from Machine Learning Engineering (16:30) Simple first steps in building AI applications (22:53) An explanation of BM25 (retrieval system) (23:43) The problems associated with fine-tuning (27:55) Simple customer support solutions for rolling out AI thoughtfully (33:44) Chip’s thoughts on staying focused on the problem (35:19) The challenge in evaluating AI systems (38:18) Use cases in evaluating AI (41:24) The importance of prioritizing users’ needs and experience (46:24) Common mistakes made with Gen AI (52:12) A case for systematic problem solving (53:13) Project-based learning vs. structured learning (58:32) Why AI is not the end of engineering (1:03:11) How AI is helping education and the future use cases we might see (1:07:13) Rapid fire round — The Pragmatic Engineer deepdives relevant for this episode: • Applied AI Software Engineering: RAG https://newsletter.pragmaticengineer.com/p/rag • How do AI software engineering agents work? https://newsletter.pragmaticengineer.com/p/ai-coding-agents • AI Tooling for Software Engineers in 2024: Reality Check https://newsletter.pragmaticengineer.com/p/ai-tooling-2024 • IDEs with GenAI features that Software Engineers love https://newsletter.pragmaticengineer.com/p/ide-that-software-engineers-love — See the transcript and other references from the episode at https://newsletter.pragmaticengineer.com/podcast — Production and marketing by https://penname.co/. For inquiries about sponsoring the podcast, email podcast@pragmaticengineer.com. Get full access to The Pragmatic Engineer at newsletter.pragmaticengineer.com/subscribe

## Transcript

[0:00] SPEAKER_0: How would you define AI engineer or AI engineering?
[0:02] Yeah,
[0:03] so before when you wanted to build machine applications,
[0:06] you need to build your own models.
[0:08] So that means that you need your own data and you need expertise in how to train
[0:12] a babysitter model.
[0:13] However,
[0:13] nowadays if you want to build your application leveraging machine or AI,
[0:16] you can just send a direct API call and access is this wonderful capability.
[0:21] SPEAKER_1: So that's really, really lowers the entry barrier to people.
[0:24] SPEAKER_1: You don't need data anymore, you don't need a fancy AI degree anymore.
[0:28] It's a shift of focus from less machine learning and more engineering and
[0:33] more product.
[0:34] Chip Huyen is a computer scientist and writer and author of the book AI Engineering.
[0:39] SPEAKER_0: This book is currently the most-read title on the O'Reilly platform.
[0:42] Previously,
[0:42] Chip was a researcher at Netflix,
[0:44] a core developer of NeMo,
[0:46] NVIDIA's gen AI framework,
[0:47] ML Engineering's Snorkel AI,
[0:49] and founded and sold an AI startup called Claybutt AI.
[0:53] She taught machine learning systems design at Stanford and her current book
[0:57] is a second one on ML and AI Engineering.
[0:59] It's safe to say she's one of the most-read ML engineering and AI engineering
[1:03] experts in the world.
[1:04] In our conversation today,
[1:05] we cover what is AI engineering and why does this feel a lot more full-stacked
[1:09] than ML engineering did.
[1:11] What are typical steps to build on AI application from choosing a model
[1:15] through using RAG all the way to fine-tuning?
[1:17] What are practical ways for software engineers to get started building AI
[1:20] applications and a lot more on this very timely topic.
[1:24] If you enjoyed the show,
[1:25] please subscribe to the podcast on any platform and on YouTube.
[1:28] SPEAKER_0: Thank you.
[1:29] SPEAKER_0: This greatly helps the show get to even more listeners and viewers.
[1:32] SPEAKER_0: Chip, welcome to the podcast.
[1:34] SPEAKER_1: Hey, hi, I'm Chip.
[1:35] SPEAKER_1: I'm very excited to be here.
[1:37] So I've been following your substack for a while,
[1:40] so I was really looking forward to the chat.
[1:44] SPEAKER_0: So first of all, I really wanted to congratulate you on this book.
[1:47] SPEAKER_0: I've started to read the book, so I've not read the whole thing.
[1:49] I have started with some chapters and I went deeper in others,
[1:52] and what I found is when I looked at a table of contents,
[1:55] I was like,
[1:55] "Well,
[1:56] this looks okay in terms of,
[1:58] it'll span a broad sense because it goes from how do you understand
[2:03] foundation models,
[2:04] how do you evaluate them,
[2:06] whatispromptengineering?"
[2:08] And you go into things like fine tune RAG,
[2:11] fine-tuning dataset engineering,
[2:13] but then on each of those sections it just starts to get...
[2:16] like,
[2:16] initially there's an introduction,
[2:18] but it starts to then go deeper.
[2:19] So for example,
[2:21] for evaluation mythology,
[2:22] like here I'm just looking at a table of contents,
[2:24] but I start to read this,
[2:25] it's like,
[2:26] well,
[2:26] we know it's important to evaluate AI models.
[2:28] We know it's harder to do,
[2:30] but then you go into things like AI as a judge or ranking models with
[2:33] comparative evaluation and challenges of it.
[2:37] And then it goes,
[2:37] that's where some parts,
[2:40] in my sense is that I had to slow down,
[2:42] I had to look things up.
[2:44] So it does go really deep into a lot of these sections,
[2:47] which I found very refreshing that it's got a mix of breath,
[2:52] but it also goes deep.
[2:53] So this is definitely not a fast read for me,
[2:57] but it's one of those things where I'm just going to keep coming back to it.
[3:01] SPEAKER_1: Thank you.
[3:01] SPEAKER_1: It was not a fast write either.
[3:03] SPEAKER_1: It took quite a while and a lot of references.
[3:06] I think I also published,
[3:07] so I think in the book I cited about over a thousand references.
[3:12] SPEAKER_1: So I actually read even more papers.
[3:15] SPEAKER_1: I gone to a lot of code bases.
[3:16] I have a tracking of 1000 reposts,
[3:20] at least now 800 Stars now,
[3:23] on GitHub Stars.
[3:23] So I went to a lot of those code bases and a lot of blog posts,
[3:28] other books from the '80s,
[3:29] from the '70s,
[3:30] '90s to understand AI.
[3:32] So I also publish at least about approximately 100 links,
[3:36] reference links.
[3:37] SPEAKER_1: I felt really, really useful for me in the process of writing the book.
[3:40] So if you just want to look at those references on their own,
[3:42] it's on my GitHub.
[3:43] SPEAKER_0: And I was surprised with some of the really original research.
[3:50] It's not just like,
[3:51] oh,
[3:51] here,
[3:52] I wrote about these papers or here was what I read.
[3:55] But as you mentioned the 1000 reposts,
[3:57] you actually have a paragraph or a section about how did GitHub repositories
[4:03] change over time,
[4:04] the ones that were about infrastructure AI,
[4:07] application level,
[4:09] other things,
[4:10] andyouactuallyhave-
[4:11] SPEAKER_1: More development.
[4:11] SPEAKER_0: More than 900 reposts mapped out.
[4:13] SPEAKER_0: I never saw anything like that.
[4:14] And clearly that was you kind of slicing and dicing and doing your own kind of research.
[4:19] SPEAKER_1: Yeah, I feel like I have this thing, I do a lot of manual labor.
[4:24] I do think I get a lot of value out of doing things in non-optimally,
[4:28] I feel like a lot of people focus on what is the quickest way to do it,
[4:30] what is the funniest way to do it?
[4:32] But sometime if you are willing to put into effort into things that a lot of
[4:36] people are not willing to,
[4:37] I feel like you can get some kind insight that other people don't get.
[4:43] So one thing that is a little interesting about this book,
[4:44] it is about AI engineering and this field moves so quickly.
[4:47] Like,
[4:47] just in a week we've now had a new model come out,
[4:49] for example,
[4:49] DeepSeek,
[4:50] that people are talking about in a few weeks.
[4:52] SPEAKER_0: How did you write this book?
[4:54] How were you able to write a book about such a fast moving industry so that by
[4:57] the time it's released,
[4:57] which was clearly a few months after you finish it,
[4:59] it will still be relevant?
[5:06] SPEAKER_1: That is a great question.
[5:07] SPEAKER_1: So when I started writing the book, I was thinking the same.
[5:09] I was like,
[5:09] "Is now the right time to write it,
[5:10] becausethere'ssomanythingsarestillchanging?"
[5:12] But when ChatGPT really came out,
[5:13] a lot of people,
[5:14] I had this existential crisis.
[5:15] I was in a group chat and I was just like,
[5:18] "Oh no,
[5:18] whatdoesitmeanforusengineer?"
[5:20] SPEAKER_1: I feel like there are two things that I usually identify with.
[5:22] SPEAKER_1: One is being an engineer and the other is being a writer.
[5:24] SPEAKER_1: And guess what?
[5:24] Two use cases that AI are really good at,
[5:26] writing code and writing,
[5:27] right?
[5:37] SPEAKER_0: Yeah.
[5:37] SPEAKER_1: So it was like, oh shoot, well what does it mean?
[5:38] SPEAKER_1: So I started interview a lot of people.
[5:39] I start like reading so much and I talk to a ton of people and I started making a
[5:42] lot of note and as a process what I realized is that while a lot of those things
[5:47] seem new,
[5:47] a lot of the fundamentals have been there for a while.
[5:50] SPEAKER_1: So for example, language modeling is not a new task.
[5:52] Claude Shannon introduced that back in the 1950s as a time level,
[5:56] we'll talk about RAG,
[5:57] right?
[5:58] SPEAKER_1: RAG is actually not new, is based on retrieval augmented generation.
[6:15] SPEAKER_0: RAG, as in retrieval augmented generation, right?
[6:17] SPEAKER_1: And retrieval is a very old technology.
[6:18] It is powering,
[6:20] it's already powering a lot of use cases on the internet,
[6:24] like search or recommender systems and better databases have been around
[6:27] for a while.
[6:28] SPEAKER_1: In fact, the search is very, have so many cool algorithms already.
[6:31] So I thought,
[6:31] "Okay,
[6:31] first I see a lot of things that are new," and the second is that I try to focus on
[6:36] like asking the question.
[6:37] Okay,
[6:37] so sometime it was like,
[6:37] oh there is a problem with this,
[6:39] a solution to this problem.
[6:41] And I asked the question of whether this is due to fundamental limitations of
[6:44] AI or is this just due to the temporary capabilities of AI.
[6:46] And I try to see,
[6:47] okay,
[6:47] if it's due to something of the recent current capabilities,
[6:51] how fast is that capability changing?
[7:07] So first of all,
[7:08] in the early days a lot of people have shared a lot of prompt tips,
[7:11] like for example you try to prize the models,
[7:13] like,
[7:13] "Hey,
[7:13] say,
[7:14] if you answer this correctly,
[7:15] I'mgoingtogiveyou$200."
[7:16] Or we talk about prompt robustness,
[7:17] how robust is a model to prompt perturbations.
[7:20] And then I was reading about it and I found is actually models is getting more
[7:22] and more robust to prompts.
[7:23] First of all we saw from the GPT,
[7:25] to GPT 3.5 compared to GPT-3,
[7:26] it's already so much more robust.
[7:28] That means that the small changes to prompt actually reduce to a lot less
[7:31] variations in model performance.
[7:34] SPEAKER_1: So this type thing, I felt like it's probably not going to stick around.
[7:38] SPEAKER_1: So those tips are not going to be very relevant to share.
[7:59] So already as people were still like,
[8:00] this was at the height of people are saying there might be a job as a prompt engineer.
[8:04] SPEAKER_0: So you already saw this is likely trending down.
[8:05] SPEAKER_0: Do I understand it correctly?
[8:07] So yeah,
[8:07] I do think just try...
[8:07] so writing I think it's a score is making a bet.
[8:10] Like when you write a topic,
[8:11] you're trying to bet on whether it's going to stay relevant in the future.
[8:15] So I think it's the people have been trying to looking at progress and trying
[8:20] to see what it's going to be like in one or two years from now on.
[8:29] SPEAKER_1: So first another example is context length.
[8:31] So at that point it was like okay,
[8:31] 8,000 context length to 128 in a few months.
[8:33] we want long context length,
[8:35] but then I kept seeing these people going really,
[8:37] really fast from what,
[8:39] SPEAKER_1: It's super, super fast.
[8:39] So it was like okay,
[8:40] maybe the question is less about context length but context efficiency
[8:47] because it's like can a model use a context efficiency really,
[8:52] really well.
[8:53] So there are certain,
[8:53] so those are the best and I do think that there are certain changes during the
[8:59] process of writing the books that make me feel more confident in the best I made,
[9:02] or like another things like multi-modality,
[9:06] I do think this.
[9:07] When I wrote about multi-modal back in 2023,
[9:09] people told me,
[9:10] "You're too early.
[9:10] SPEAKER_1: Everyone was still working on language now.
[9:12] SPEAKER_1: We are not there yet."
[9:13] SPEAKER_1: But it was like, yeah, it's just inevitable.
[9:15] I think we learn how to work with language,
[9:17] but I do want to do a lot of stuff with more than just language and I think it's
[9:23] just like,
[9:23] and nowadays just everywhere,
[9:24] almost all models nowadays are like multi-modal.
[9:38] So the title of the book is AI Engineering and we also now have this term of AI engineer,
[9:43] it's spreading like wildfire.
[9:45] How would you define AI engineer or AI engineering because I feel it's a
[9:48] little bit of a loaded term these days?
[9:54] SPEAKER_1: It is.
[9:54] SPEAKER_1: I feel like a lot of terms nowadays are loaded.
[9:57] You're not allowed to use agent anymore,
[10:00] you're not allowed to use a lot of things anymore.
[10:04] So when I was agonizing over the title for the book because people was like...
[10:08] first I know that we need a different term for machining engineering,
[10:12] and the reason is that with foundation models there are a lot of fundamentals
[10:17] like systematic approaches that still the same at machining engineering,
[10:21] there are a lot of new things.
[10:22] So for example,
[10:24] one thing is that before when you wanted to build a machining applications,
[10:28] you had to build the old models.
[10:36] SPEAKER_0: And so just be clear, you're comparing it with machine learning, right?
[10:36] SPEAKER_1: Yes.
[10:38] So what machine learning engineers did versus what AI engineers are doing.
[10:43] SPEAKER_1: Yeah.
[10:43] So yeah,
[10:43] just trying to explain why we need a different term for machining engineers
[10:47] to describe what we are doing today.
[10:49] So before when you wanted to build a machining application,
[10:53] you need to build your own models.
[10:55] So that means that you need your own data and you need expertise in how to train
[11:00] a babysitter model.
[11:01] However,
[11:01] nowadays if you want to build an application leveraging machine link or AI,
[11:06] you can just send it direct API call and access this wonderful growability.
[11:11] SPEAKER_1: So that's really, really lowers the entry barrier to people.
[11:14] SPEAKER_1: You don't need data anymore, you don't need a fancy AI degree anymore.
[11:16] A second thing is that before you need distributions because you deploy
[11:20] machining application as part of the CD applications.
[11:23] So first of all you build a recommender system,
[11:27] you need an e-commerce website so that you can have recommender,
[11:30] or some kind of website.
[11:38] So it's deployed as part CD application,
[11:41] fire detection is part of making banking applications or some kind of
[11:44] payment app.
[11:44] SPEAKER_1: But however, now you can just put it out as standalone applications.
[11:48] You don't need an existing distribution channel,
[11:50] but having a distribution channel is really,
[11:52] really useful.
[11:59] So I think that another very big thing is that it's a shift of focus from less
[12:02] machine learning and more engineering and more product.
[12:05] So before you start from,
[12:05] if you good machine engineer,
[12:07] you start from data,
[12:08] you have to gather data,
[12:09] you maybe had human annotations and then you train a model and now once a model
[12:14] is good,
[12:14] deploy that into the audio product.
[12:16] But nowadays you actually start with a demo so you have a good idea,
[12:19] okay,
[12:19] let's just try it out and see if it works.
[12:22] So we start a product and after you say,
[12:26] "Okay,
[12:26] sounds like it works pretty well,
[12:28] Iwanttomakeitbetter."
[12:37] So they started gathering more data maybe a bit as more of problem examples or
[12:40] in very rare case I don't recommend most people do in the early days it's like fine-tuning,
[12:45] but it's very rare.
[12:47] But basically you start getting more data maybe for evaluation,
[12:52] extremely important.
[12:53] Having good evaluations become even way way more important with AI Engineering.
[12:57] So you've got data and then after that it was maybe you've been sending a lot of
[13:00] API calls to OpenAI and Anthropic or Google and okay,
[13:03] now it's too expensive,
[13:03] "NowIneedtodevelopmyownmodel."
[13:10] So you start hosting your own model using open source alternative or
[13:13] fine-tuning model.
[13:14] Yeah,
[13:15] so before machine engineer you go from a data model to product and now with
[13:20] engineering you go from product to data into model.
[13:23] So it plays a lot more focus on product and data,
[13:26] which wouldn't be a competitive advantage when everyone shares a similar
[13:30] base like AI capabilities.
[13:31] So I do think that says need a different term,
[13:34] just separate it from machine engineer and I didn't know what terms to use.
[13:36] SPEAKER_1: And then I was like, "Okay, that's just ask the people."
[13:39] So I surveyed a bunch of people,
[13:40] if I was doing,
[13:41] what does this building applications on top 5-digit models and almost
[13:45] everyone was like,
[13:45] "AI Engineering" and I was like,
[13:47] "Okay,
[13:47] if that's the case and this,
[13:48] let'sgowithAIEngineering."
[13:57] So do I understand correctly that the biggest difference is that machine
[14:02] learning engineers did a lot more kind of groundwork getting the data and
[14:07] building the model.
[14:08] Whereas AI Engineering,
[14:09] you have a lot of that at least initially you can start either as APIs or something.
[14:12] So there's more of an engineering,
[14:14] you kind of hack things together,
[14:16] you put it together,
[14:16] and then over time as things become more serious or your product is bigger,
[14:20] you do a lot more of the...
[14:21] you might build your own model or you might host it.
[14:23] You might one day even build your own model if it is there,
[14:26] but it is just a lot later.
[14:27] So a lot of ML engineering comes down the path if it's big enough,
[14:30] if it works,
[14:31] et cetera.
[14:31] Whereas with ML engineering it was the other way you had to put in all this
[14:36] effort and then see if it even works?
[14:43] So first,
[14:43] I feel like every,
[14:44] companies may have different definition of the role.
[14:47] I think it's even the same company,
[14:48] people with the same title,
[14:49] same role,
[14:49] can do very different things.
[14:51] SPEAKER_1: So it's not very, never like a clear-cut definition.
[14:58] Second things is that,
[14:59] I don't think the question is like machine or engineering.
[15:02] In the vast majority of general AI system have since there are very strong
[15:07] traditional machine learning classifier component.
[15:10] So imagine you're building a customer support chatbot,
[15:13] which is whenever I ask at a conference,
[15:15] I see a lot people raise your hand,
[15:17] very classic general applications.
[15:18] And so you get a request from a customer and maybe you have several different
[15:22] potential solution for it.
[15:24] Maybe if it's an easy query you might send it to a cheap model,
[15:27] or if it's a harder query,
[15:28] you might send it to a more expensive model.
[15:29] But if it's something very sensitive like,
[15:31] "Hey,
[15:31] why did you charge me twice for the bill last month," right?
[15:34] SPEAKER_1: Then you might want to send it to human operator.
[15:36] So you might have this one like a router or an intent classifier,
[15:38] you choose what to send it to,
[15:39] which is like a traditional classical machine learning model that you can build.
[16:00] Or after you get a response from an AI model,
[16:01] you might think it's like okay,
[16:02] this is contain like PII because you don't want to send back to users
[16:04] responses contain private information.
[16:05] SPEAKER_1: So that PII detection can be, or toxicity detections can be a classifier.
[16:08] Or in the RAG system nowadays we talk about how a lot of it using retrieval systems,
[16:11] which are also I think it's a RAM of classifier,
[16:12] classical machine learning that you can build yourself.
[16:32] So what are the most common techniques used when building AI applications,
[16:37] things that the software engineer who's going into building AI
[16:40] applications I should know about and later I can go deeper into?
[16:47] So the assumption here is that you have tried a lot of solutions and now you try
[16:52] generative AI and you think this generative AI is a solution for you and it think,
[16:56] what is the first,
[16:57] like what should I progress from there,
[16:59] right?
[16:59] SPEAKER_1: Is that correct?
[17:00] SPEAKER_0: Yeah, yeah.
[17:00] SPEAKER_0: It's just like what are some common approaches?
[17:03] I think things like RAG,
[17:04] fine-tuning,
[17:05] or other things that are just good to know about,
[17:07] I should probably learn more about those topics.
[17:11] SPEAKER_1: Yeah.
[17:11] So I think those techniques are useful and I would usually recommend some development,
[17:15] not recommend,
[17:16] but a common pattern you have seen is a certain developmental power path.
[17:20] So initially the first thing it would say is trying to understand what means a
[17:24] good response,
[17:24] what a good response is and what a bad response is.
[17:27] SPEAKER_1: So what you want is a large share rate and it's not always intuitive.
[17:38] SPEAKER_1: So for example, LinkedIn has a very great example.
[17:39] So they build like job fit assessments for candidates and they found out that
[17:43] a majority of their time spending just to understand what candidates needed
[17:47] from the model.
[17:47] So initially they focus on correct,
[17:49] but since they realize it's candidates \[inaudible 00:17:58\] found it not helpful.
[17:53] Say if a candidate asks the model,
[17:55] am I a good fit for this job and the AI respond with,
[17:57] "You're terrible fit," it was like,
[17:58] "Okay,
[17:58] wellwhatamIdoingthisinformation?"
[18:11] So you need to try to understand,
[18:14] okay here is they want more understanding what are the gaps and how they can
[18:17] fill the gaps or get suggestion for other roles that are a better fit for them
[18:21] right now.
[18:22] So have this pictures clear understanding and then build a guideline,
[18:26] like,
[18:26] okay,
[18:26] given this response,
[18:28] answer this,
[18:29] be helpful,
[18:29] show them the gaps or show them the role.
[18:31] SPEAKER_1: Have very clear guideline for the model in the prompt.
[18:34] So you try those prompts and then you see,
[18:36] look at the,
[18:37] and then maybe you can try to add more examples and then just go through,
[18:42] get really good response.
[18:42] Like try to evaluate,
[18:44] maybe have a,
[18:44] create,
[18:45] a set of queries and expert responses using both automated metrics as a judge
[18:47] and also human evaluations to measure the progress.
[18:50] So you have done prompting or we have added more examples just as a prompt and
[18:54] maybe you will start make it more complex.
[19:11] So may use,
[19:11] you can give then the model more context so they can answer better questions.
[19:13] So maybe when users ask the questions,
[19:15] you can have the model pull out all the documents or all the job listings
[19:18] related to this question or information about the company,
[19:21] get a candidate resume.
[19:29] SPEAKER_1: So you build a system, you augment the context with the documents.
[19:31] SPEAKER_1: So that's a RAG pattern.
[19:33] SPEAKER_1: So I do think it's like RAG is a very, very powerful pattern.
[19:38] SPEAKER_1: It's nothing really, really fancy.
[19:42] And something interesting about RAG is that a lot of people equate RAG with
[19:47] vector search.
[19:48] When I see a lot of people like,
[19:49] "Oh I want to use RAG and here's,
[19:52] likewhatdatabaseIshoulduse?"
[19:57] SPEAKER_0: So people jump straight to vector search.
[20:00] SPEAKER_1: Yes.
[20:00] SPEAKER_1: Yes.
[20:05] SPEAKER_0: So as engineers who are like, "I need a vector search database."
[20:06] SPEAKER_0: A vector search database.
[20:11] SPEAKER_1: Yeah, people love databases, very interesting.
[20:14] But I feel like there was the first solution is probably not like
[20:17] jump-stretchembeddingbaseretrievalbecauseyouhavetouse-
[20:20] SPEAKER_0: Oh, interesting.
[20:22] SPEAKER_1: Yeah, you need to build, it's like an embedding model.
[20:24] SPEAKER_1: The qualities that really highly dependence on its equality embeddings.
[20:28] SPEAKER_1: If you have embeddings numbers, you retrieve things well.
[20:30] SPEAKER_1: Also, very databases can be quite expensive to run and a latency.
[20:33] Also,
[20:34] another thing is vector databases can,
[20:35] embeddings can also obscure certain keywords.
[20:37] For example,
[20:38] if I'm searching for a specific arrow code,
[20:39] right,
[20:40] through embedding,
[20:41] you don't really get the exact error code anymore.
[20:49] SPEAKER_1: So there's certain challenging with vector databases and vector search.
[20:52] SPEAKER_1: So I think it's like the usual common approach.
[20:55] SPEAKER_1: Maybe just start with something as simple like keyword retrieval.
[20:59] If you ask,
[20:59] just extract all the keywords from the user query and find all the document
[21:03] that with the query and you say,
[21:04] okay,
[21:04] maybe the documents are too long and now it can't fit into context,
[21:08] that's when it's start with chunking.
[21:09] Okay,
[21:10] now how do I chunk this documents into context that can fit into the context
[21:14] link and now it's chunking have other problems.
[21:16] So for example,
[21:16] maybe the keyword,
[21:17] the document is about the company X,
[21:19] but people say from now on company X is referred to as company,
[21:22] so it is a rest document.
[21:23] SPEAKER_1: You don't have X appear anywhere.
[21:24] SPEAKER_1: So you don't get the chunk.
[21:26] SPEAKER_1: If you search for X, you don't get the chunk below.
[21:42] So now you might want to,
[21:45] okay,
[21:45] now I need to extract the keywords from documents and get X metadata to every
[21:48] chunk or get the title of the document or some people started adding summary
[21:52] to it.
[21:52] SPEAKER_1: Like Anthropic has a very, very good article called Contextual Retrieval.
[21:57] It's like as they asked ChatGPT to generate the key information metadata
[22:00] about each document and append it to prepare that to each chunk to have you
[22:07] retrieve the right chunk.
[22:08] So I do think this having data preparations,
[22:09] right,
[22:10] actually give a really huge performance boost.
[22:13] I have seen this giving way,
[22:14] way better performance boosts and focusing on which better databases I
[22:17] should use.
[22:25] So,
[22:25] what is a little bit of...
[22:26] I'm not saying that they're not useful,
[22:28] it's just in the beginning you probably wants to try something simple with
[22:30] the biggest performance against and then you started moving up the
[22:33] complexity level.
[22:34] Also for a lot of retrieval,
[22:35] someone told me very interesting,
[22:38] a little bit of heartache,
[22:40] but he was just like,
[22:41] "I'm not going to take any retrieval system seriously if they don't
[22:44] benchmark against BM25.
[22:52] SPEAKER_1: So BM25 is a pretty old school, like 20+ years now.
[22:55] SPEAKER_1: Retrieval is a lot simpler, like turn-based retrieval not embedding.
[22:59] And it's a really,
[22:59] really hard-to-beat retrieval system and I think a lot of time we use it,
[23:04] if you start a complex,
[23:05] you can now combine both turn-based retrieval and a simpler solutions with
[23:09] embedded databases.
[23:11] So you have both the semantic on the embedding side,
[23:14] but you have \[inaudible 00:23:18\] the exact keyword match on the turn-based.
[23:16] SPEAKER_1: So a lot of things they hybrid search is very common.
[23:25] So I guess we talk about prompt engineering,
[23:28] add more examples,
[23:29] RAG,
[23:29] right?
[23:30] I think after that maybe have any max out on a lot of those things,
[23:34] which usually take why people maybe then it might consider fine-tuning,
[23:38] but I think they usually have a lot of reservations against fine-tuning
[23:42] because fine-tuning has a whole host of new problems that you need to deal with.
[23:46] Like fine-tuning first you need to think of now you have this model,
[23:50] you have fine-tuning and now you need to think of how you host it.
[23:53] SPEAKER_1: And a lot of models are big, they're building the parameters.
[23:56] SPEAKER_1: They are not easy to sell.
[24:03] SPEAKER_0: I actually read this part in the book.
[24:04] You actually go into the details of the problems with the memory size and also
[24:09] you cover the alternatives where you might not need as much memory,
[24:12] but they also bring other trade-offs.
[24:14] SPEAKER_0: So it's trade-offs within trade-offs within trade-offs, right?
[24:18] SPEAKER_1: Yeah.
[24:19] You're starting to solve one problem,
[24:21] but you're going to get a bunch of others and you'll need to decide is it worth
[24:24] my time,
[24:24] effort,
[24:25] resources.
[24:28] SPEAKER_1: Yeah, yeah.
[24:28] So I definitely think that...
[24:29] also when you find the model,
[24:31] you kind of own the fashion model.
[24:34] SPEAKER_1: And now the question is how do you maintain it?
[24:36] Like,
[24:36] anything,
[24:36] we have this whole world of very smart people having new models like this
[24:39] capability increases rapidly.
[24:42] So how long can the fine-tuning model outperform those new models that we're
[24:46] putting out?
[24:47] So you might spend a lot of energy in a \[inaudible 00:24:55\] model and just a
[24:51] \[inaudible 00:24:56\],
[24:51] maybe like some random Chinese company you never heard of release is
[24:54] extremely fast and extra model,
[24:57] right?
[25:02] SPEAKER_0: Oh yeah.
[25:02] SPEAKER_1: So yeah, it is quite challenging.
[25:03] So yeah,
[25:03] five-tuning,
[25:04] you think it is the last resort,
[25:06] not the first line of defense.
[25:12] Yeah,
[25:12] but what I've heard is basically,
[25:14] if I got it correctly,
[25:15] do a structured approach,
[25:18] start with prompting,
[25:20] start with simple,
[25:22] start with you getting responses that make sense,
[25:24] then add more data.
[25:25] You can do this with RAG,
[25:27] you can do it with chunking,
[25:28] keyword extraction,
[25:30] data preparation really,
[25:31] really makes a big difference,
[25:33] which a lot of people don't think about.
[25:35] SPEAKER_0: And then you can go to more advanced things.
[25:37] There's a whole host of things that you could do,
[25:41] but my understanding is you're probably saying you'll probably get there
[25:44] over time,
[25:44] but initially these things will keep you busy,
[25:46] and you'll probably be able to build a pretty good system just with the basics
[25:49] and a little bit of engineering,
[25:51] and most importantly understanding the problem that you're trying to solve
[25:55] as opposed to building whatever shiny technology or approach.
[26:01] Yeah,
[26:01] I feel like the approach can also be different for individual developers or
[26:06] enterprise if you look as a whole organization.
[26:09] So one thing I haven't seen is,
[26:10] especially in the early day technologies,
[26:12] usually enabling new use cases actually bring more returns on this
[26:14] incremental improvement of use cases.
[26:16] So maybe if instead of spending the effort into investing a lot of energy,
[26:22] \[inaudible 00:26:28\],
[26:23] like little bit of performance with fancier complexity,
[26:25] maybe using the same suspect you have had and opening up new applications.
[26:29] So,
[26:30] I said that's why I feel like a lot of companies will take a while to get to the
[26:33] fine-tuning phase.
[26:45] SPEAKER_0: Trust isn't just earned, it's demanded.
[26:46] Whether you're a startup founder navigating your first audit or seasoned
[26:49] security professional scaling your governance risk and compliance program,
[26:51] proving your commitment to security has never been more critical or more complex.
[26:54] SPEAKER_0: That's where Vanta comes in.
[26:54] Vanta can help you start or scale your security program by connecting with
[26:57] auditors and experts to conduct your audit and set up your security program quickly.
[27:00] Plus,
[27:00] with automation and AI throughout the platform,
[27:02] Vanta gives your time back so you can focus on building your company.
[27:04] Businesses use Vanta to establish trust by automating compliance needs
[27:09] across over 35 frameworks like SOC 2 and SO 27001.
[27:10] With Vanta,
[27:10] they centralize security workflows,
[27:11] complete questionnaires up to five times faster,
[27:13] and proactively manage vendor risk.
[27:14] Join over 9,000 global companies to manage risk and prove security in real time.
[27:17] For a limited time,
[27:19] my listeners get $1,000 off Vanta at vanta.com/pragmatic.
[27:21] SPEAKER_0: That is V-A-N-T-A.com/pragmatic for $1,000 off.
[27:46] So,
[27:46] let's say that at my company we decide to build an AI solution and let's take
[27:49] the example of that it's going to be customer service automation.
[27:55] SPEAKER_0: What are typical approaches that I should know about?
[27:58] SPEAKER_0: And you do cover some of these things in your book as well.
[28:02] SPEAKER_0: Some of the kind of more common steps that I'll need to take.
[28:05] SPEAKER_1: So customer support solutions.
[28:06] So I would say the first thing I would look into is what are the bottlenecks for
[28:10] the solution right now?
[28:10] So for example,
[28:11] I have worked with another startup I work with,
[28:13] it actually has this challenge,
[28:14] oh,
[28:14] they had a lot of customer support requests and they don't know what to answer.
[28:18] SPEAKER_1: And the solution is very interesting.
[28:20] It's just like,
[28:21] okay,
[28:21] let's try to drive a lot of questions into the common channel like public discourse.
[28:24] SPEAKER_1: So that means, "Are the users going to help answer the questions?"
[28:27] So as much,
[28:28] and also in the future someone have a question,
[28:31] they can just refer to previous discussions instead of asking.
[28:34] So they try to make a lot of the discussion customer support request public.
[28:48] So another solution was pretty popular in 2018,
[28:48] 2019 is that you do routing to the right department.
[28:51] SPEAKER_1: So the challenge then is they realize the bottleneck is in triaging.
[28:54] SPEAKER_1: So first we get a request, I don't know which department to send it to.
[28:57] So there were a bunch of startups then,
[28:59] it's like,
[28:59] "Okay,
[28:59] let's try to build a system to predict,
[29:01] like should this go to the finance department or should it go to a technical
[29:04] supportdepartment?"
[29:05] SPEAKER_1: Just like routing already reduce the friction a lot.
[29:07] SPEAKER_1: And if you think it's like okay, we need gen AI to do it.
[29:10] So I think it really recommend the frameworks that Microsoft introduced
[29:12] they call is like CRON work/run going from a slightly lower stake to higher
[29:15] stake deployment.
[29:41] SPEAKER_1: So first off the customer support chatbot, right?
[29:43] SPEAKER_1: So maybe initially you can have a human in the loop.
[29:45] So for example,
[29:46] you can have for every request instead of a human agent writing the response
[29:51] from scratch,
[29:51] you can have AI suggest a few options and the human can choose one or can just
[29:55] choose ones as a starting point and then make short a quick edit and send it.
[30:06] So once you see this,
[30:07] okay,
[30:07] maybe the acceptance rate is getting really high.
[30:10] So first of all,
[30:10] for these categories of queries,
[30:12] maybe the acceptance rate is 90%,
[30:13] so you might feel more confident to roll it out.
[30:14] So maybe roll it out to maybe smaller users or you can run it out in even
[30:17] internal use cases and so then,
[30:18] you give it more automations but reduce the scope of deployment and then
[30:21] after you're really happy with it,
[30:23] you can run out to more users.
[30:25] So yes,
[30:25] that's how I would go about something like customer support chatbot.
[30:43] Nice,
[30:43] because what I was kind of expecting you might say is like,
[30:48] "Oh,
[30:48] just build this AI framework,
[30:49] deploy."
[30:49] SPEAKER_0: I feel that's what a lot of companies are doing by the way.
[30:52] SPEAKER_0: A lot of things is like, "Oh, gen AI.
[30:55] Oh,
[30:55] let's get this model from ChatGPT or Anthropic,
[30:58] let's try to put it there.
[30:59] SPEAKER_0: Let's put it out there."
[31:00] But I really like how...
[31:01] what you're saying sounds like it's not really,
[31:03] what you explained,
[31:05] it's not really specific to gen AI that you went through.
[31:08] Look at the business problem,
[31:10] look at the problems you have,
[31:10] look at the options which include not just gen AI but more traditional
[31:14] machine learning,
[31:16] as you said,
[31:17] "Classifier."
[31:20] And then you'll look at if these tools help your problem and then you'll make
[31:22] sure that it actually solves your problem when you roll it out,
[31:25] don't just blindly roll it out.
[31:26] SPEAKER_0: Which, I mean, all of this sounds to me it's not really new is it?
[31:29] It's not...
[31:30] like,
[31:31] you could have said the same thing two or three years ago before gen AI,
[31:35] except we would have not had those gen AI tools to play with.
[31:40] SPEAKER_1: Yeah.
[31:40] So I think actually before our chat,
[31:41] I should look up one of the talks before and you have this talk about things
[31:46] that haven't changed,
[31:47] things that feel very similar about social engineering and I definitely
[31:50] like dealing with a new technology is one of the things that never change.
[31:54] So I feel like every time there's a new technology that comes out,
[31:57] I can hear the collective sigh of senior engineers everywhere saying,
[32:02] "Noteverythingisnew."
[32:04] SPEAKER_1: People just try to get such technology to work for everything.
[32:07] So I do think that,
[32:07] a very common challenge I see is that people just jump straight into it.
[32:11] SPEAKER_1: They just want to use gen AI for when they don't need gen AI.
[32:18] SPEAKER_1: So, I think there's two different headlines.
[32:20] So one headline is like,
[32:21] "I use gen AI," or as the other headline,
[32:23] "Isolvedtheproblem."
[32:24] So if you want to focus on the first headline here,
[32:27] you use gen AI,
[32:27] but if you want to solve the problem then you need to understand what is a problem,
[32:31] what are the challenges there,
[32:32] what's the roadblocks?
[32:33] And remove the roadblock using the simplest solutions,
[32:35] not the fanciest one.
[32:44] Yeah,
[32:44] I feel that there's a bit of a really strong fear of missing out across most
[32:46] tech companies that everyone knows that this is such a transformative technology.
[32:49] SPEAKER_0: It gives us so many new capabilities that it will be important.
[32:51] I think everyone knows that their company will be using it,
[32:54] but now there's a fear of missing out of,
[32:56] "Oh,
[32:56] what if my team doesn't build it?
[32:58] SPEAKER_0: What if someone else gets ahead of me?"
[33:00] And so a lot of companies,
[33:01] many teams are all building it and trying to just,
[33:03] using a hammer,
[33:03] looking for nails even though they might not need it at the time.
[33:05] I'm not sure if this is a bad thing necessarily because people at least get
[33:07] experience with it and they will need to learn about it.
[33:24] But it's a very interesting time,
[33:26] because rare to see,
[33:27] usually we see a new backend framework come out or something that's limited
[33:30] to the main and people jump on it.
[33:33] But this is one...
[33:34] the first time I've seen that the whole industry is jumping on it and everyone
[33:36] is trying to use it and put it in whether it works or not.
[33:44] SPEAKER_1: I definitely agree with you on this FOMO thing.
[33:46] I do think this is like,
[33:48] everyone's jumping on,
[33:49] it's actually a pretty good thing.
[33:51] SPEAKER_1: I feel like the energy is incredible.
[33:54] SPEAKER_1: I've never before seen so many smart people focusing on the same problem.
[33:58] SPEAKER_1: It's incredible and the progress is amazing.
[33:59] I do think,
[34:00] however,
[34:00] I think there's an irony is just like the more we want to not miss out on things,
[34:04] the more things we will miss,
[34:06] because what I think is feel like if we try to keep up with news,
[34:09] we should try to keep from jumping from one piece of news to another,
[34:12] we will always stay at the surface level.
[34:14] SPEAKER_1: We never really go deep into anything.
[34:22] So I actually don't quite read news,
[34:23] so I find like a little bit distracting.
[34:24] So I think,
[34:24] I try to,
[34:24] feel like my approach is like,
[34:25] okay,
[34:25] pick a problems that you care about and then only care about things just like
[34:28] how you solve this problem.
[34:29] So if there's some news coming out and was just like,
[34:31] "Does this help me solve this problem?
[34:33] If it doesn't,
[34:33] I can't wait," because I feel like if there's something important,
[34:37] it will still be important two weeks from now on,
[34:38] a month from now on.
[34:39] Like I don't drop everything and it's like,
[34:41] okay,
[34:42] let's just go and understand what it is.
[34:44] SPEAKER_1: So I feel like trying to get a more, try to stay bit calmer, yeah.
[35:05] So when you're building an AI system,
[35:06] one of the things that you will come across is you need to evaluate the output,
[35:09] how well it works,
[35:10] does it solve your problem.
[35:11] Why is it difficult to evaluate AI systems and what are common ways to do that?
[35:23] So evaluation I think is billion-dollar question or even a trillion-dollar
[35:26] given how much we're investing right now.
[35:30] SPEAKER_0: Trillion-dollar, nice.
[35:31] SPEAKER_1: Like, we need to go big and if go big, better go really big.
[35:34] So I think this it's challenging because the smarter AI becomes,
[35:39] the harder it is for us human to evaluate it.
[35:42] So before,
[35:43] if AI was incoherent,
[35:43] you can pretty tell if the response is bad,
[35:46] I say,
[35:46] "Okay,
[35:46] doesn't sound good,
[35:47] it'sabadresponse."
[35:47] SPEAKER_1: But nowadays, it's pretty coherent.
[35:50] SPEAKER_1: First of all, you ask, should I generate a summary of a book?
[35:53] If the summary sounds convincing,
[35:55] you actually don't know if it's a good summary or not and you might have to read
[36:00] the entire book yourself just to evaluate whether it's a good summary.
[36:13] The math,
[36:13] a lot of time I personally use AI,
[36:15] actually to ask a lot of questions because I don't know the answer and because
[36:19] I don't know the answer,
[36:19] I don't know if the answer is correct,
[36:22] right?
[36:22] So,
[36:22] an example was a lot of people can tell if a math solution should,
[36:25] a first grade question is correct,
[36:26] but very few people can tell is if a fancy equation or proof is correct.
[36:40] So I remember when Avant came out,
[36:42] Terence Tao,
[36:42] he's this amazing mathematician,
[36:43] I think it's one of the best mathematicians of our time,
[36:45] he actually took time to evaluate Avant and he says the experience of using
[36:49] Avant is similar to advising an incompetent but not completely stupid,
[36:54] like a mediocre but not completely incompetent student.
[37:01] But that makes me think,
[37:02] if we really need the brightest minds today to evaluate AI,
[37:05] then we are soon to run out really,
[37:06] really smart people to evaluate AI.
[37:18] SPEAKER_1: So what would be the next step forward?
[37:19] So before a lot of time we use a whole human as a gold standard for AI performance.
[37:24] It's like okay,
[37:24] so humans starting writing out,
[37:25] "Here is how you should respond to this and here's how you should do it," and AI
[37:30] should try to copy human.
[37:32] SPEAKER_1: But now for many, many tasks we have AI outperform human way better.
[37:37] So I think there's several...
[37:40] so I thought about several approach to deal with it and I think that's why I
[37:43] separate the chapter like those.
[37:54] SPEAKER_1: So, initially I had one chapter on evaluation.
[37:56] SPEAKER_1: There's more I write about it.
[37:58] It was like should...
[37:58] there's so much,
[37:59] enough,
[37:59] two pretty long chapters on evaluation.
[38:00] The first chapter is on general methodology and the second chapter is about
[38:03] how to use different techniques to evaluate the AI system.
[38:06] SPEAKER_1: So as methodology, one is that functional correctness.
[38:09] So it evaluates the output of application based on how well it performs a task.
[38:13] So if you say that,
[38:13] "Hey,
[38:13] use AI to save energy," you can see how much energy is actually saved.
[38:15] Or,
[38:16] "Hey use AI to play this video game," you can see how high is the score you can
[38:19] actually get.
[38:20] SPEAKER_1: Or, "A very common use case for this coding."
[38:37] SPEAKER_1: So I do think it's not a code incident, it's a coding.
[38:39] It's one of the most popular use cases because we actually know how to
[38:44] evaluate generated code.
[38:45] We might not know to evaluate generated essays and stuff,
[38:47] but you know how to evaluate generated code because we've been testing code
[38:54] for a long time.
[38:55] So with code you can do use functional correctness to evaluate first whether
[38:59] is this code comply,
[39:00] does it run?
[39:01] SPEAKER_1: Second, does it generate expected outputs?
[39:03] SPEAKER_1: That's what we wanted to do.
[39:05] SPEAKER_1: So that's one approach.
[39:08] SPEAKER_1: The second approach is using AI to evaluate other AI.
[39:11] SPEAKER_1: So we have been using AI to evaluate, to automate a lot of applications.
[39:17] SPEAKER_1: So can we also use AI to automate evaluations?
[39:20] SPEAKER_1: And actually they're doing pretty well.
[39:22] I think we have this like...
[39:23] I think,
[39:23] in many,
[39:24] many,
[39:24] even back in 2023,
[39:24] \[inaudible 00:39:30\] has this report.
[39:26] They saw as the majority of applications they saw already has some sort of AI
[39:29] as a judge or \[inaudible 00:39:38\] as a judge.
[39:32] SPEAKER_1: And I think it's growing.
[39:33] SPEAKER_1: And we do things that it's getting pretty cost affections and useful.
[39:37] But,
[39:37] of course,
[39:38] I see a lot of challenges around using Azure just so we can go into later.
[39:51] But another approach I thought was very interesting is on comparative evaluations.
[39:56] And the reason,
[39:56] as humans it might be hard for us to give an absolute score on something,
[39:59] but if we can give two versions of something,
[40:01] we can tell,
[40:02] "Oh,
[40:02] werealizethisoneisbetter."
[40:03] So we have done a lot of studies showing that even for tasks where AI is outperforming,
[40:08] doing as a level with humans,
[40:09] experts can't really do,
[40:11] we can still tell,
[40:12] detect the differences.
[40:13] So I think this hasn't been a lot,
[40:14] been guiding,
[40:16] not just evaluations but also model development.
[40:31] SPEAKER_0: Sounds like there's just no simple answer, right?
[40:32] You kind of need to go through all these options and figure out in your case,
[40:38] which makes sense for costs for what you can do,
[40:40] can you have a human in the loop so there's no real silver bullet,
[40:42] no one thing that you can just use.
[40:51] SPEAKER_1: Yeah, I don't think there's a simple solution.
[40:52] So that's one thing I'm a little bit skeptical about,
[40:55] evaluation tooling,
[40:56] because a lot of challenge evaluations,
[40:58] or not because we don't know how to evaluate,
[41:00] but because it's required disciplines and hard work and a lot of things is
[41:03] like tools can't really automate.
[41:05] So,
[41:05] for example,
[41:05] one thing for evaluation,
[41:07] we need to evaluate an application based on what users want.
[41:11] And so that means that we need to go and talk to users,
[41:14] we need to look at their interactions,
[41:17] because a lot of things what we think is like we want you to evaluate what matters.
[41:23] SPEAKER_1: We have to measure what matters.
[41:38] So for example,
[41:38] I have several examples of how it's very counter-intuitive thinking that we
[41:41] are measuring one thing,
[41:42] but user actually care about other things.
[41:45] So in the beginning,
[41:46] for example,
[41:46] I have a friend who building a pretty big applications basically building to
[41:50] summarize meetings.
[41:51] SPEAKER_1: And initially they were like, "Okay, we try to measure correctness.
[41:54] SPEAKER_1: Does the summary cover the content of this meeting?"
[41:57] SPEAKER_1: I say, so think about, "Hey, does the model follow the format?
[42:00] Because they think the site users want a shorter summary and they agonize over,
[42:04] dowewantthree-sentencesummariesorfive-sentencesummaries?"
[42:07] SPEAKER_1: And they try to measures though.
[42:09] But actually eventually what they found out that users don't really care
[42:11] about the whole content of the meeting.
[42:14] People only want,
[42:15] "What is the action item for me,
[42:16] likewhatwehadtodoafterthis."
[42:32] SPEAKER_1: So they start changing.
[42:33] SPEAKER_1: So they don't measure correctness anymore.
[42:34] I mean they just don't make up things,
[42:37] but they focus on,
[42:39] don't miss out on action items.
[42:40] SPEAKER_1: Best fit for the person asking for the summary.
[42:43] So I was like...
[42:45] or the example is some people using chatbot for,
[42:47] so we talk about a customer support chatbot.
[42:49] SPEAKER_1: So I want to go back to that example.
[42:53] SPEAKER_1: So a pretty big tax firm.
[42:54] So it'd be with chatbot,
[42:58] so tax software,
[42:59] you can pretty much tell which company it is.
[43:03] SPEAKER_1: So they just launch a chatbot to help people with tax preparations.
[43:07] SPEAKER_1: And the responses were very lukewarm.
[43:11] SPEAKER_1: They were measuring by the users, how they use it.
[43:14] SPEAKER_1: And it was like people just didn't really seem to use it.
[43:17] SPEAKER_1: And they were like, "Why is that?
[43:19] SPEAKER_1: Is that because it's not, what hallucinate, or what is the challenge?"
[43:26] So they charge you to measure all those kind metrics,
[43:28] but in the end,
[43:29] user didn't use it because seems they hate typing.
[43:31] SPEAKER_1: People don't really like typing.
[43:34] And also because in your face with a domain,
[43:36] a domain that you don't really know,
[43:37] I use the software because I don't know a lot of things about tax.
[43:40] SPEAKER_1: I never know what question to ask.
[43:45] SPEAKER_0: Oh, so they just didn't know what to type.
[43:46] SPEAKER_0: They didn't understand the domain.
[43:49] They went to this tax thing because they wanted to take care of their tax.
[43:57] So I think myself started trying to understand more what type of questions
[44:02] people would ask and suggest that in the beginning and then basically say
[44:07] guide users.
[44:08] So it's kind of education,
[44:10] "Here's a question you should ask and then here's the answers,
[44:12] keepgoing."
[44:13] So I think a lot of that is it's just a lot of understanding your domain as a
[44:16] problem domain,
[44:17] like go talk to users,
[44:17] looking at the data.
[44:19] SPEAKER_1: I do still think that looking at data is very, very important.
[44:21] SPEAKER_1: I think Greg Brockman has a great quote about it.
[44:24] "So you think it's manual data inspections is orders of activities that has a
[44:28] highestratioofvaluetopackage."
[44:40] So that means that's people don't think highly of manual data inspection of
[44:46] data wrangling.
[44:47] Let's give some interns to do,
[44:48] let me think of something fancy or algorithms and stuff.
[44:53] But actually it's extremely high value because by looking at data you detect patterns,
[44:58] you understand how that users use a product.
[45:01] So actually I usually a very good practice,
[45:02] I really highly recommend to teams.
[45:04] SPEAKER_1: It's like don't forget human evaluations.
[45:07] So you use ASA judge,
[45:08] but as a judge has a lot of challenges because the quality of the judge depend
[45:12] on the underlying model and the prompt and also non-deterministic.
[45:17] SPEAKER_1: So things can change over time.
[45:19] But if you have some human evaluations,
[45:20] very consistent,
[45:20] very clear guidelines,
[45:22] every day go in there,
[45:24] look at maybe 50 samples of actual interactions.
[45:26] Or if you have more resources,
[45:27] go as high as like 500 or 1000,
[45:28] so that you can get some kind of picture of how first,
[45:31] how the users are using your product,
[45:32] any changes in behavior based on current events,
[45:34] maybe because of recent administration change,
[45:35] maybe we'll have a lot more questions about that topic for example,
[45:39] or help you correlate with all the automated metrics.
[45:41] First of all,
[45:42] if the AI just score somehow start changing compared to the human's judge score,
[45:45] maybe there's something you to investigate.
[46:07] Yeah,
[46:07] so I guess you cannot,
[46:10] as you said,
[46:12] you can't really skip hard work if you want to get good results and you can't
[46:17] really pull humans out of the loop fully,
[46:20] at least initially.
[46:21] What are some common mistakes you've seen when teams are building AI applications?
[46:26] SPEAKER_1: Common mistakes.
[46:31] SPEAKER_0: I'm sure you did a lot.
[46:33] Yeah,
[46:33] but I feel like I don't want to say otherwise everyone say she is an idiot.
[46:38] SPEAKER_1: Yeah, I think we touched on several.
[46:39] SPEAKER_1: So one very common mistake is like use gen AI when you don't need gen AI.
[46:42] SPEAKER_1: So first of all, this is a startup that came with me with a pitch.
[46:46] It was like,
[46:47] "Oh,
[46:47] we'regoingtousegenAItohelppeopleoptimizeelectricityusage."
[46:51] So when I asked,
[46:53] "So people can tell the chat,
[46:56] the AI,
[46:56] 'Hey,
[46:56] I live here and here are the activities I do during the days that are very
[47:02] energy intensive,
[47:02] maybeinchargingacarordoinglaundryorsomething.'
[47:07] And AI is going to tell you,
[47:09] 'Hey,
[47:09] you should do this activity this time and this time so that you can minimize
[47:12] theelectricitybill.'"
[47:13] And they were like,
[47:14] "Oh,
[47:14] ourreasonsshowthatifcansaveyouonaverage30%ofelectricitybill."
[47:18] SPEAKER_1: It was like free money.
[47:20] SPEAKER_1: Why would anyone not want that?
[47:22] So I was asking them what is this cost saving if I'm just manually schedule the
[47:26] most intensive one during the off-peak hours and it just,
[47:29] it'll go like,
[47:30] okay,
[47:30] just charge the car at 10 p.m or something.
[47:32] They were like,
[47:33] "We haven't done that yet,
[47:34] butwearegoingtotryitandletyouknow."
[47:36] SPEAKER_1: And they never got back to me and they abandoned the idea later.
[47:39] So I feel like a lot of those optimization problems can be solved like readily,
[47:42] maybeevenaspreadsheet-
[47:56] SPEAKER_0: Without gen AI?
[47:57] SPEAKER_1: Without gen AI.
[47:57] Another spectrum is that I see a lot of companies giving up on gen AI because
[48:03] they think that gen AI is not good for that problem because they have tried it
[48:06] and it doesn't work.
[48:07] SPEAKER_1: And a lot of time I got surprised.
[48:08] I was like,
[48:09] "Wait a second,
[48:10] I just talked to another company who just,
[48:12] theyusethatforthesimilarusecaseanditworkedreallywell."
[48:19] And when we look into it,
[48:21] usually because of bad product,
[48:21] because they don't promise well,
[48:23] they don't understand the users,
[48:25] they don't even know what to evaluate well.
[48:28] So first of all,
[48:29] I was working with these companies that does basically extracting resume information.
[48:34] So there's a person,
[48:36] I get a resume and it tries to map out where does the person work before and
[48:40] create a summary of that person's life.
[48:42] SPEAKER_1: And they have two steps.
[48:43] First from resumes they try to extract on the text and then after own the
[48:47] extracted text,
[48:47] they extract the organizations from the extracted text.
[48:50] SPEAKER_1: So by the way, the resume is a PDF, not pure text, right?
[48:53] SPEAKER_1: And then I asked them... so it worked terrible.
[48:55] SPEAKER_1: They got the organizations wrong about 50% of the time.
[49:11] SPEAKER_1: And then I was asking them, "Where in the process does this fail?
[49:13] Is that in from the PDF to extracted text or from the extracted text
[49:17] organizationextraction?"
[49:18] SPEAKER_1: They were like, "Oh, we don't know.
[49:19] SPEAKER_1: We didn't do that."
[49:21] It was like if you can't pinpoint,
[49:23] you can't localize where it fails then how can you fix it?
[49:27] So a lot of times it seems like common sense,
[49:28] but somehow,
[49:29] I don't know,
[49:29] it's just something is always puzzle me a little.
[49:38] Or another is just statue complex for first of all,
[49:40] jump \[inaudible 00:49:48\] bases or fine-tuning.
[49:41] Or another common one nowadays it's like when you see a fancy agent framework,
[49:44] just let's use this framework,
[49:45] let's just try it.
[49:46] SPEAKER_1: And I think that eventually attractions are really, really cool.
[49:49] SPEAKER_1: I think very grateful for many attractions that make my life easier.
[49:52] But I do think that the attraction should encode best practices and should be
[49:55] heavily tested.
[50:13] But I think we still in the phase,
[50:14] we're still learning best practices and also a lot of attractions can
[50:19] introduce unnecessary very,
[50:20] very painful bugs.
[50:22] So when I was going through the code basis of a lot of those frameworks and I
[50:26] found out something interesting.
[50:29] A lot of those frameworks are some deformed prompts you can have,
[50:33] you get started,
[50:33] because it just make very easy for you to begin.
[50:35] But then every single of those prompts and look have some type typos and
[50:40] they're just like chains.
[50:42] So you have somebody submit a quick PR to fix the typos,
[50:44] but it's not part that release anything.
[50:47] So if you're using this framework,
[50:50] using one of the deformed prompts,
[50:52] and then suddenly the performance applications just change,
[50:54] eventually don't quite know why is that changing?
[50:57] SPEAKER_1: Because prompt was changing under the feed.
[50:59] SPEAKER_1: So those are very interesting.
[51:00] SPEAKER_1: Those are just patterns that I've seen.
[51:14] It's interesting,
[51:15] because what you mentioned,
[51:17] it sounds like if I'm collecting these,
[51:18] it's like using this technology when you don't really need it,
[51:21] giving up on it without just for common sense reasons.
[51:24] You could have just fixed some easy things,
[51:26] using a new framework when it's just not really high quality,
[51:29] and it doesn't really have the best practices.
[51:32] This kind of all sounds stuff that we could just replace gen AI with a new
[51:37] technology or a new stack and we will probably hear similar things,
[51:40] right?
[51:40] It's typically these things,
[51:41] because it's changing all the time,
[51:44] there's no best practices,
[51:45] no one really knows how to use it.
[51:46] Whoever tells you they're the expert,
[51:48] they're still just,
[51:49] they have a maximum year of experience with it.
[51:51] SPEAKER_0: It's not really new, is it?
[52:04] SPEAKER_1: Yeah, I think I definitely agree with you.
[52:06] I do think that even though technologies change over time,
[52:09] so systematic thinking,
[52:11] it's like systematic approaches to problem usually don't change.
[52:13] If you want to solve problems,
[52:15] you first start by breaking down the problems,
[52:18] seeing where the challenges are and go through different solutions to do that.
[52:21] SPEAKER_1: It seemed common.
[52:22] But I think a lot of time,
[52:24] a lot us get formal,
[52:24] I think formal get in the way and was like,
[52:26] "Okay,
[52:27] we know this is the right thing to do,
[52:28] but I also feel like I just need to check this thing out first and keep doing
[52:33] thatthreetimesaday."
[52:35] So day's gone and you just never really get time to sit down and think really
[52:39] deeply about what is that you're trying to do.
[52:55] Yeah,
[52:55] so I guess we're going to see a lot of the mistakes that are with new technology happen.
[52:58] Plus if some of the listeners have adopted new technology,
[53:00] you can probably use some of that approaches.
[53:02] I mean just localize it for gen AI and see if you can avoid some of those.
[53:05] Speaking of new technology,
[53:06] as someone who's learning gen AI,
[53:07] a software engineer who wants to get into AI engineering,
[53:09] what would your recommendation be to learn,
[53:12] if things do change so fast?
[53:13] SPEAKER_0: You did mention the importance of fundamentals.
[53:16] SPEAKER_0: What would you focus on?
[53:27] So I have a lot of thoughts on learning because I like learning a lot and I think
[53:32] over time I just try to just observe some patterns.
[53:37] And by the way,
[53:39] the way I learn things might not be the same as the way you're learning.
[53:43] SPEAKER_1: People have different learning style.
[53:45] SPEAKER_1: But in general, I think of learning has two different approach.
[53:46] SPEAKER_1: One is project-based learning and the other is structured learning.
[53:49] Project-based learning is like,
[53:50] okay,
[53:51] you choose a project and you work on it,
[53:53] really go and try and solve every problem in that project and finish it.
[53:57] But structure-based learning is more like when you take a course or you read a book,
[54:03] it's just like somebody else laid out,
[54:05] here's the things that you want to do.
[54:07] And I think there's quite a bit of a debate on...
[54:10] somebody told me recently that,
[54:12] a friend,
[54:12] a very good friend,
[54:14] and he said he thinks it's a problem nowadays with people who want to become engineer,
[54:20] is that they spend too much time learning and not enough time doing.
[54:29] SPEAKER_1: And he was "Just forget all the courses, forget all the books.
[54:32] SPEAKER_1: Just pick a project and just work on it."
[54:33] SPEAKER_1: And I do think this project-based learning is very, very valuable.
[54:35] But you think of here is a set of the skills and knowledge I want to do,
[54:39] I need to have to become really good at something.
[54:41] Project-based learning can have you hit a lot of these points,
[54:44] but it doesn't always have you hit all the points and it can sometimes get the confusions,
[54:49] whereas also sometimes students to complement the structured learning.
[54:57] And another thing that struck my project-based learning that a lot of people
[55:01] usually do it follow some tutorials.
[55:03] SPEAKER_1: People like here someone has this picture how to do this.
[55:07] And I think tutorials is really cool and I think I'm so personally do it a lot,
[55:10] but I still notice that it's very easy choose just mindlessly clicking one
[55:15] cell to another and just run the cell,
[55:17] run another,
[55:17] and don't really stop to ask why is this being this way?
[55:25] SPEAKER_1: Why is this library being important?
[55:27] SPEAKER_1: Why is this code written this way?
[55:28] SPEAKER_1: Why is the batch size is 16 instead of 64?
[55:30] SPEAKER_1: Why is there expert in the mixture of expert?
[55:33] SPEAKER_1: So some people, it's very easy to not stop.
[55:34] SPEAKER_1: There's no mechanism to force you to stop.
[55:36] You just want to run to the end and see what the output and make some changes by
[55:39] the best guesses.
[55:40] SPEAKER_1: It's something funny.
[55:41] Like,
[55:41] when I was working on this open source project and it was like,
[55:45] it didn't want to do a market research and see who is using this framework.
[55:48] So it was a framework with iBiz,
[55:50] and I thought it's like...
[55:51] I knew that if you wanted to use this framework,
[55:53] you need to import iBiz.
[55:54] So I went through on a GitHub and I searched for all the repos that like import iBiz.
[55:58] SPEAKER_1: And I found repos that I went to it.
[55:59] It has import iBiz,
[56:01] but then the code base does not have iBiz anywhere else.
[56:03] SPEAKER_1: It's not used at all.
[56:04] SPEAKER_1: And I was like, "What is happening?"
[56:06] SPEAKER_1: So I realized that's a lot of those repos, copy from a tutorial.
[56:10] SPEAKER_1: And that tutorial use, import iBiz.
[56:28] SPEAKER_1: And that's my mistake.
[56:28] SPEAKER_1: And then like everyone else.
[56:29] So maybe the original developer import iBiz and then deleted the code
[56:33] because they didn't use it anymore and then everyone could copy it just like
[56:36] it's the same thing.
[56:37] SPEAKER_1: So I feel like that is something a little bit dangerous.
[56:39] Tutorial based learning is great,
[56:42] but I do think it's very important to be able to stop and ask questions.
[56:45] And sometimes structured learning can help you ask the right questions,
[56:47] like think things through.
[56:48] So yeah,
[56:48] I think if before was starting,
[56:49] I would recommend maybe a mixture like yes,
[56:51] choose some project you want to work on.
[56:53] It doesn't have to be big fancy project,
[56:54] just try to pick one and then at the same time compliment it with some
[56:58] structured learning.
[56:59] SPEAKER_1: Pick whatever, maybe a book or doing a course with a friend, read paper.
[57:02] SPEAKER_1: I think read paper is a bit interesting because reading paper is a skill.
[57:07] It can be quite time-consuming and you need to know what you want to get out of it.
[57:32] SPEAKER_1: But yeah, so start a project, compliment with structured learning.
[57:34] At the same time there's an exercise that I felt very,
[57:38] very useful,
[57:39] at least for me initially,
[57:40] is that for a week I try to observe what I do and I try to make note of what I do and
[57:45] try to think of what percentage of that can be automated by AI,
[57:49] what could be done by AI,
[57:50] and then I try to use AI to do those.
[57:52] SPEAKER_1: And this just gave me a lot of ideas on the use cases.
[57:54] Just think about it,
[57:54] what matters to me,
[57:56] and it'll be an application that can solve the problem.
[57:58] SPEAKER_1: Its great already.
[58:09] Yeah,
[58:10] I think that's an unconventional way,
[58:12] but it is a good way to look at it because in the end...
[58:15] also I think it might help you get ahead of this dread of what would AI do for me?
[58:19] SPEAKER_0: Because you realize what happens when you automate things.
[58:22] SPEAKER_0: Which actually leads to my next question.
[58:24] There's a lot of fearmongering around,
[58:26] oh,
[58:26] AI will mean the end of software engineering because AI is very good at coding.
[58:28] SPEAKER_0: It's a lot better than a lot of other areas.
[58:30] What is your take on as AI gets better will it actually end software
[58:34] engineering or it will change it or it's not going to much change actually?
[58:45] SPEAKER_1: I think it goes back to the question of what software engineering is.
[58:48] SPEAKER_1: So maybe you can get an analogy, maybe you can help explain it better.
[58:53] So the writing,
[58:54] so we tend to confuse the most salient activity of something as the job itself.
[58:58] SPEAKER_1: So, first of all, writing.
[58:59] Writing in the past,
[59:00] writingmeansasaphysicalactofputtingsomewordsontopaper-
[59:11] SPEAKER_0: On a paper or-
[59:12] SPEAKER_1: Yeah.
[59:12] And back then people think of writing as that and people actually took pride
[59:15] in the calligraphy like,
[59:17] "Oh,
[59:17] you have beautiful handwriting,
[59:18] you must be smart,
[59:19] youmustbeintelligent."
[59:20] But then we had computers and now writing doesn't refer to the act anymore.
[59:25] SPEAKER_1: Writing refers a process of arranging ideas into a readable format.
[59:28] SPEAKER_1: And I think the same thing is coding.
[59:31] So now people think,
[59:32] so engineering thinking it is a physical act of putting code on a VS Code or Vim
[59:38] or whatever software that you use,
[59:40] but that's not what engineering is.
[59:42] SPEAKER_1: So engineering is about solving problems.
[59:45] SPEAKER_1: Here's the problems.
[59:46] SPEAKER_1: How do I come up with executable programs to solve this problem?
[59:49] Coding itself is just a physical act of it,
[59:52] and I do think it's like,
[59:53] yes,
[59:53] maybe AI can have you automate coding,
[59:55] but I don't think it's going to fully automate problem solving because you
[59:58] still need to know what problems is and only you can understand what problems
[1:00:03] you're facing.
[1:00:11] And also AI has the problem like coding really,
[1:00:13] or software engineering really,
[1:00:14] is like,
[1:00:14] yes,
[1:00:15] you need to solve problems,
[1:00:17] but what I don't think we say is you need to do it very precisely.
[1:00:20] The reason the job software engineer or programmer exists is because it is
[1:00:24] very hard to be specific to speed the computer's language,
[1:00:27] because you know if you move that "If statement" somewhere else,
[1:00:34] or if you change a variable,
[1:00:36] suddenly the program crashes because now you have a stack overflow exception,
[1:00:40] which you,
[1:00:40] of course,
[1:00:41] understand if you're a software engineer,
[1:00:44] but if this is just a business user who says,
[1:00:47] "I want to show if you resize the window,
[1:00:50] I want the button to move over," it's easy to say.
[1:00:54] But then as a software engineer,
[1:00:56] you know the edge cases,
[1:00:57] you know the environment,
[1:00:58] you know what needed to worry about,
[1:00:59] you need to worry about system events,
[1:01:02] et cetera.
[1:01:04] SPEAKER_0: And then you'd write code for all of those things.
[1:01:07] And I'm sure we'll get to a point where the AI will be able to generate some of that,
[1:01:09] but it might not.
[1:01:10] And you will still at some point need someone who understands that code and
[1:01:13] can figure out where the gap is.
[1:01:15] Because English as a language is not as precise as a programming language.
[1:01:19] Programming languages were invented to be very precise and unambiguous and
[1:01:23] very easy.
[1:01:24] You can go from assembly code to the programming language,
[1:01:27] because it's a one-on-one mapping,
[1:01:28] and then from English to a programming language,
[1:01:30] that's very fuzzy,
[1:01:30] right?
[1:01:36] SPEAKER_1: Yeah, definitely.
[1:01:39] SPEAKER_0: I think that profession will not go away.
[1:01:40] For users it might work for some kind of more hobbyist use cases of like you say
[1:01:45] something and you get something roughly that's it,
[1:01:47] and you try multiple times and it generates something else and you're happy.
[1:01:52] But for a business or a professional use case,
[1:01:54] you will need those people who will be able to guarantee that you get exactly
[1:01:57] what you want.
[1:02:00] Yeah,
[1:02:00] I'm actually really excited about AI can automate part of coding because it
[1:02:02] actually enables software engineers should do software much more complex.
[1:02:05] SPEAKER_1: So I go back to the analogy of writing.
[1:02:06] Before when I had to do the manual copying words onto papers,
[1:02:08] all the books back then were very small.
[1:02:09] I think it's 5,000 words or 10,000 words considered big because it took a long
[1:02:12] time for people to copying things.
[1:02:13] But now we have books like a hundred of thousand words and I think it just makes
[1:02:16] things a lot easier.
[1:02:17] SPEAKER_1: I do think software engineering.
[1:02:18] If you don't have to manually write code,
[1:02:19] it quickly can turn ideas into snippets,
[1:02:21] executable programs.
[1:02:22] I don't think the \[inaudible 01:02:42\] running much,
[1:02:24] much more complex software.
[1:02:45] And maybe one software engineer will be able to command a lot more software
[1:02:49] debug or maintain a lot more complex system by oneself.
[1:02:51] Because right now there's a reason that for a million lines of code,
[1:02:54] usually there's several engineers.
[1:02:55] SPEAKER_0: It's rare to have just one engineer.
[1:02:57] SPEAKER_0: If a company broke that, I'm not talking about dependencies.
[1:03:00] SPEAKER_0: So that will be interesting.
[1:03:01] So what other use cases are you excited about for that AI could bring outside
[1:03:03] of just coding?
[1:03:16] SPEAKER_1: Let's see.
[1:03:16] SPEAKER_1: I think I'm excited about education.
[1:03:19] SPEAKER_1: So I do thing like AI can help people learn.
[1:03:21] SPEAKER_1: We want to learn a lot faster.
[1:03:22] So I think one thing I realized is that nowadays if you know the questions,
[1:03:26] finding the answer is actually quite easy.
[1:03:28] You can ask AI and it usually give you pretty,
[1:03:30] a lot of,
[1:03:31] at least it can give you a lot of references for you to go and read more about it.
[1:03:35] But then what's still hard,
[1:03:36] it's how should you come up with the right questions and I think it's like
[1:03:39] education needs to focus on forcing students,
[1:03:40] like create the habit of asking questions and understanding.
[1:03:44] SPEAKER_1: So I do think it help learning to become a lot more efficient.
[1:03:46] People can learn a lot of things and I feel like I do believe that if we can learn
[1:03:50] better and faster,
[1:03:50] then we can actually do more things.
[1:03:53] SPEAKER_1: So I'm very excited about what that would look like.
[1:04:31] SPEAKER_1: What other use cases I'm excited about?
[1:04:32] SPEAKER_1: I'm excited about entertainment.
[1:04:33] Yeah,
[1:04:33] I think it's sometime we always think of entertainment and education as
[1:04:35] separate things,
[1:04:36] but I don't see why we can't have games that help to learn things more.
[1:04:39] We have some strategy games like teaching girls about negotiation,
[1:04:42] so that could be really fun.
[1:04:43] SPEAKER_1: Or there's something more intellectually stimulating content.
[1:04:46] Movies or shows don't have to be about mindless...
[1:04:47] but I know other people watch that because it's help them escape.
[1:04:50] But I also think that genres,
[1:04:51] it's maybe think a little or understand more about different fields.
[1:05:26] So I do think that now AI can assist us in creating content that is both
[1:05:31] entertaining and intellectually stimulating.
[1:05:33] SPEAKER_1: So I think it could be a lot of fun.
[1:05:35] For some boys,
[1:05:36] it's like now that we have a lot of medium adaptations,
[1:05:38] so you have a book,
[1:05:38] you can convert it into a movie and sometimes you have a movie,
[1:05:41] you convert it into a game.
[1:05:44] SPEAKER_1: Or you have papers convert into a podcast.
[1:05:46] So now if we have some content,
[1:05:48] AI can help us convert,
[1:05:50] adapt different medium,
[1:05:51] that could be very,
[1:05:53] very exciting.
[1:06:05] Yeah,
[1:06:05] I think it's...
[1:06:06] there are a lot of small problems that I'm also interested in.
[1:06:09] SPEAKER_1: I think I haven't touched on any of the enterprise.
[1:06:11] SPEAKER_1: I feel like that's where most money is still.
[1:06:13] I do think there a lot of...
[1:06:15] I think enterprise or company organization structure is going to change.
[1:06:19] SPEAKER_1: So first, what does that mean?
[1:06:20] If you think about a lot of organizations,
[1:06:21] what is the job as a middle management is to first just upgrading information
[1:06:26] from the reports and transmit it up to the executives and is the other way they
[1:06:31] transport it,
[1:06:32] transmitting information directions from executive to lower layers.
[1:06:37] But information aggregation is actually really,
[1:06:41] really,
[1:06:42] and something AI can do really,
[1:06:43] really well.
[1:06:44] SPEAKER_1: So I do think this, companies can be a lot more efficient.
[1:07:07] SPEAKER_0: So let's close up with some rapid questions.
[1:07:08] SPEAKER_0: So I'll just shoot some questions and you tell me what pops to your mind.
[1:07:12] What programming language did you use most when you built AI applications or
[1:07:15] did ML engineering and why?
[1:07:20] SPEAKER_1: Python and JavaScript.
[1:07:22] SPEAKER_0: JavaScript as well.
[1:07:22] SPEAKER_0: How come?
[1:07:24] Yeah,
[1:07:24] definitely I think it's a huge part of building products and I have to build
[1:07:29] demo quickly,
[1:07:30] so I think that's very,
[1:07:31] very handy.
[1:07:32] SPEAKER_1: I'm not very good.
[1:07:32] SPEAKER_1: I think I've always been scared of JavaScript, but I'm very grateful.
[1:07:35] SPEAKER_1: AI actually helped me getting started a lot easier nowadays.
[1:07:42] SPEAKER_0: And which one is your favorite LL model right now and why?
[1:07:50] SPEAKER_1: I don't really have a favorite.
[1:07:52] SPEAKER_1: I use it for different things.
[1:07:53] So I still use ChatGPT out of habit because I have a bunch of prompts that I use
[1:07:59] already and I have little things like rebuilds prompts for ChatGPT I'm still using.
[1:08:06] I use Clocks sometimes for grid writing because I think it's sometimes less cliche.
[1:08:11] SPEAKER_1: I'm reading on DeepSeek R1 and who's not reading about R1.
[1:08:14] SPEAKER_1: So just showing up.
[1:08:16] SPEAKER_1: I don't think I have a favorite.
[1:08:17] I think I did some of Lava,
[1:08:20] which is the version of Lama before,
[1:08:21] for some kind of interesting use case from a screenshot to code.
[1:08:24] SPEAKER_1: Just trying to test this out.
[1:08:25] SPEAKER_1: I'm having fun with it.
[1:08:26] SPEAKER_1: But yeah, I'm not emotionally attached to any of them.
[1:08:40] SPEAKER_0: And what's a neat AI tool that you've used and that you like?
[1:08:47] SPEAKER_1: So I have some things that I built that's really helped me do research.
[1:08:49] So one thing is when I saw a link,
[1:08:50] like of papers,
[1:08:51] it's a lot of links and I try to get a...
[1:08:53] So I realized when I read the links,
[1:08:55] I do through the same process.
[1:08:56] I read the abstract,
[1:08:57] I look up the authors,
[1:08:58] use a bit work,
[1:08:59] ask questions,
[1:08:59] and also when I was out,
[1:09:00] check the citations.
[1:09:01] SPEAKER_1: So I have to do the tool just go and scan so I have a link.
[1:09:04] SPEAKER_1: Just give me all the information on like a high level-
[1:09:19] SPEAKER_0: So that's a tool to scratch your own itch.
[1:09:23] SPEAKER_1: Yeah, I think that's the beauty of AI now.
[1:09:26] SPEAKER_1: You can build, it's taking very small amount of time.
[1:09:28] SPEAKER_1: You just build a tool just for you.
[1:09:29] Before it could take me weeks,
[1:09:30] but now I can just do one of that,
[1:09:31] I don't know,
[1:09:32] like in hours.
[1:09:40] SPEAKER_0: This is something to be excited about.
[1:09:41] SPEAKER_0: I agree with it.
[1:09:42] SPEAKER_1: Yeah.
[1:09:43] SPEAKER_1: Yeah.
[1:09:45] SPEAKER_0: What are one or two books that you've read and would recommend?
[1:09:50] Oh,
[1:09:50] I recommend a lot of books,
[1:09:51] but I feel like recommending books when I feel like forcing people to...
[1:09:55] do what you enjoy.
[1:09:55] So I like books that's helped me get a new perspective or give me insight into
[1:10:00] some topics I don't know a lot about.
[1:10:02] SPEAKER_1: So I realized the book, first of all, Complex Adapted Systems.
[1:10:06] It's a very interesting book about a system thinking how to design,
[1:10:08] like,
[1:10:09] how should design social dynamics so that to get people to work toward the
[1:10:14] goals that you want to work on.
[1:10:16] SPEAKER_1: It's a very interesting book, forced you to think about systems.
[1:10:30] I like the book The Selfish Gene because you understand more about free will
[1:10:34] and makes you question a little stuff,
[1:10:38] but it's the idea of you can live on either through genes or through ideas.
[1:10:41] SPEAKER_1: There are two ways that they can live on.
[1:10:44] So genes could lead you to live on with their offspring and reproductions.
[1:10:49] The other thing is if you have ideas and the ideas can also replicate,
[1:10:51] like the one like idea of memes,
[1:10:52] like genes and memes.
[1:11:00] SPEAKER_1: I like the book \[inaudible 01:11:02\].
[1:11:02] SPEAKER_1: So the ideas thing.
[1:11:04] SPEAKER_1: I think the author is a very interesting character.
[1:11:06] SPEAKER_1: I just read several of his books.
[1:11:08] SPEAKER_1: Actually, I really like them.
[1:11:09] SPEAKER_1: Yeah, I like that book.
[1:11:10] SPEAKER_1: Yeah, I think there are a lot of books that I like.
[1:11:20] SPEAKER_0: No, thank you for the recommendations.
[1:11:23] SPEAKER_0: So thank you for being on the podcast.
[1:11:26] I mean,
[1:11:26] AI Engineering is such a new field and it was great to hear from yourself who
[1:11:30] has clearly gone very broad and also very deep and has been in this field even
[1:11:33] before it was called AI Engineering.
[1:11:35] SPEAKER_0: So thank you for this.
[1:11:39] SPEAKER_1: Thank you so much for letting me ramble on on the show.
[1:11:43] SPEAKER_1: Yeah, I really appreciate it.
[1:11:44] And I think that one thing for me I really enjoy from writing or talking is that
[1:11:49] I get feedback because somebody was like,
[1:11:53] "Oh,
[1:11:53] I think I'm less interested.
[1:11:54] SPEAKER_1: I agree with you."
[1:11:56] I mean it's great to hear,
[1:11:59] but it can also a little bit not good for the ego,
[1:12:01] but I also really like maybe a little pushback like,
[1:12:04] "Okay,
[1:12:04] you didn't think about this,
[1:12:05] you forgot this,
[1:12:06] orIdidn'ttakeintoaccountthisone."
[1:12:07] SPEAKER_1: So I would love to get those feedback.
[1:12:09] So if there are anything that you felt like I missed out on,
[1:12:13] do let me know.
[1:12:14] SPEAKER_1: Really appreciate it.
[1:12:15] SPEAKER_0: Thank you to Chip for this conversation about AI Engineering.
[1:12:17] To get in touch with Chip,
[1:12:19] including to get feedback on her book,
[1:12:20] you can find her contact details on her website linked in the show notes below.
[1:12:24] I've also found her book AI Engineering to be a broad and deep overview of this
[1:12:30] important field.
[1:12:31] SPEAKER_0: It's a book that focuses on the fundamentals that are unlikely to change.
[1:12:35] SPEAKER_0: So if you want to learn these, it's a good book to have.
[1:12:39] In the Pragmatic Engineer,
[1:12:40] we've previously done several AI and ML related deep dives.
[1:12:44] SPEAKER_0: Check them out, also linked in the show notes below.
[1:12:46] If you enjoyed this podcast,
[1:12:47] please do subscribe on your favorite podcast platform and on YouTube.
[1:12:50] SPEAKER_0: Thanks and see you in the next one.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.