Season 21, episode 7 of the DataTalks.Club podcast with Micheal Lanham
Links:
The transcripts are edited for clarity, sometimes with AI. If you notice any incorrect information, let us know.
Alexey: Hi everyone, welcome to our event. This event is brought to you by DataTalks.Club, a community of people who love data. We have weekly events. If you want to find out more about the events we plan, there is a link in the description. Click on that link and check it out. You'll see all the events we have coming up. (0.0)
Alexey: Do not forget to subscribe to our YouTube channel. This way, you'll stay up to date with all our streams and future events like today’s. If you haven't joined our community yet, join now. There you can hang out with other data enthusiasts. (19.0)
Alexey: During today's interview, you can ask any question you want. There is a pinned link in the live chat. Click on it, ask your questions, and we will cover them during the interview. (32.0)
Alexey: That’s the usual intro. Now I am opening the questions we prepared. If you're ready, Michael, we can start. (44.0)
Michael: Yeah, sure. Hi everyone. (1:00)
Alexey: Today we are joined by Michael. His career spans two decades, from building game AI with neural networks and evolutionary algorithms to writing more than 10 books on augmented reality, virtual reality, reinforcement learning, and AI agents. (1:07)
Alexey: Today he focuses on advanced deep learning, evolutionary methods, and building intelligent systems that go beyond large language models. We will talk about what makes AI agents different from simple chatbots, how games and simulations continue to shape AI research, and what has changed in the field over the last two decades. (1:27)
Alexey: Welcome, Michael. (1:46)
Michael: Thank you for having me. (1:46)
Alexey: Evolutionary methods that’s been a while since I last heard about those. I remember studying them at university in 2006 or 2007. I haven’t seen much practical use of them since. My view is limited to what I was personally exposed to, mostly software engineering. (1:53)
Alexey: I’m really curious to know more because I remember doing those simulations and they were very fun. But first, can you tell us about your career journey so far? (2:05)
Michael: Over the last few decades, I’ve worked in various data science and machine learning roles. I started in academia, working with a group of Harvard scientists. We built a game to test children for executive functions like ADHD and behavioral problems. (2:36)
Michael: We used simple neural networks and evolutionary algorithms to create tests and patterns for players to go through, then analyzed the data. Since Calgary is an oil and gas town, I joined a small consulting company in that industry. (2:55)
Michael: I worked with them for several years on academic projects involving geomechanics and data science. We developed two products called Stab and Roxbank for oil and gas. (3:15)
Michael: That company was later purchased by a larger oil and gas provider, where I moved up from managing a small group to leading 80 people across 11 cities in seven countries including Scotland. (3:55)
Alexey: You called that a small group? (4:12)
Michael: There was a lot of travel involved. I gained exposure to many data science and engineering techniques while building various applications. I left that position because of the travel and returned to consulting, working on evolutionary algorithms for optimizing numerical analysis related to pipeline corrosion. (4:19)
Michael: Evolutionary algorithms were able to adapt faster and process data more efficiently back then. But they are very computationally intensive and haven’t been designed to benefit from frameworks that improved deep learning. (4:50)
Michael: That’s why they’ve been pushed aside, though I believe they could come back. Along the way, I worked with oil and gas providers, then in cannabis business intelligence, and financial tech. (5:01)
Michael: Currently, I work in a company building AI support assistants powered by multiple agents. We are developing deep research operator agents and other advanced tools. (5:28)
Michael: Around 2015 or 2016, I started writing books. I contacted a publisher and proposed three ideas: geospatial analysis and data, data analysis, and reverse engineering Pokémon Go. (5:45)
Michael: I had kids interested in the game and a background in GIS and Unity game development, so I was able to reverse engineer Pokémon Go. The publisher agreed, and that book became very popular, translated into 14 languages. (6:10)
Alexey: I remember when Pokémon Go came out. People were crazy about it, taking out their phones and seeing Pokémon on the screen like it was real augmented reality. (6:27)
Alexey: It lasted maybe a month from my view as a casual observer, then interest dropped. (6:47)
Michael: One of my kids still plays it. It’s still around and was a big deal as augmented reality was the next big thing then. I wrote another AR book focused more on AI because I saw many benefits from AI even then. (7:13)
Michael: Then I wrote a book about sound that became popular and used in colleges to teach sound design for games. My experience with waveform analysis from the oil and gas industry helped with that book. (7:36)
Michael: Underground, waveforms are used to analyze rock composition, which translated well to sound design. After that, I moved into reinforcement learning. Alberta was the capital of reinforcement learning, with Richard Sutton who introduced me to it, along with David Silver, his student. (8:01)
Alexey: Why was Alberta the capital of reinforcement learning? (8:38)
Michael: The roots are definitely at the University of Alberta. David Silver has since moved to California and carried the work forward, so the center of research has shifted. I continued working on various projects and books. (8:45)
Michael: I wrote several reinforcement learning and deep learning books. Then I returned to evolutionary algorithms with a book called Evolutionary Deep Learning, combining deep learning and evolutionary algorithms. (9:09)
Michael: The idea covers hyperparameter search and modifying network architectures. These methods are computationally intensive but work well for optimizing convolutional neural networks. (9:29)
Michael: I started looking at large language models early on, but they weren’t big then. After LLMs exploded, I focused on NLP work. AI agents became popular around 2023, five months after ChatGPT launched, and I embraced the concept. (10:00)
Michael: We built agent pipelines to generate code using AI agents. I also wrote a book on using AI for learning game development, called vibe coding for games, published around 2023-24. (10:35)
Michael: I finished most of the AI agents book in 2024. Initially, no one was excited about agents, but interest surged when it was published. The concepts were based on 2023 principles rather than the cutting edge by release. I'm now working on the second edition. (11:00)
Alexey: You've been productive over the last decades. Speaking of evolutionary algorithms, I think they're also called genetic algorithms, a type of evolutionary algorithm. (11:30)
Alexey: The main idea is to start with a population of individuals that vary in fitness. The unfit die, and the fit reproduce, creating offspring with random mutations. (11:47)
Alexey: Those offspring can be either fit or unfit, and the cycle repeats. You need two functions: a fitness function and a mutation function. (12:27)
Michael: Sometimes there is also a pairing function to determine how parents combine genetics to produce offspring. (12:47)
Alexey: You repeat the process until resources run out or convergence is reached based on fitness. (12:59)
Michael: Exactly. Evolutionary algorithms are popular for many applications but computationally intensive. They explore many solutions without many constraints. Around 2006, they were seen as a promising path to intelligence. (13:21)
Alexey: That was what I studied in my AI class. We used Prolog and other rule-based systems. Genetic algorithms were interesting but I never used them outside class. (13:44)
Michael: Some recent uses include evolutionary algorithms in prompt engineering for LLMs and agents. They optimize prompts by evolving them to find better results. (14:09)
Michael: This is popular but computationally expensive. You can discover prompt variations producing unexpected outputs due to LLM complexity. It has exciting applications if we solve computational challenges. (14:28)
Alexey: I found a GitHub repo called Genetic Prompt Lab with 28 stars. (14:59)
Michael: Yes, it may take about a week of computation to get results. (15:23)
Alexey: The idea is interesting. I do AI and prompt engineering too. With older models like GPT-3.5, tuning prompts made a big difference. Now, with newer models, the impact is less but still present. (15:35)
Alexey: A precise prompt usually leads to a better answer. Writing good prompts feels more like art than science. I experiment a lot, hoping something sticks. (16:00)
Michael: That's how many of these things began. It's not only LLMs. All generative AI using text can benefit. Finding the right prompt can feel like art, where one day you find the perfect phrase that works perfectly. (17:25)
Michael: Generative AI extends to video and image prompts with many applications. I've thought about writing a book on this topic next. (17:58)
Alexey: How many books have you written? (18:13)
Michael: I’ve written 11 books so far. I’m working on the 12th, the second edition of the AI agents book, and possibly a 13th on evolutionary algorithms. (18:19)
Alexey: Is the AI agents book about game development? (18:25)
Michael: No, it covers AI agents in general, for development and any application. (18:30)
Alexey: You mentioned a book using agents for vibe coding for games in 2023, but now agents are big in 2025, which explains the second edition of the AI agents book. (18:36)
Michael: Yes. I was working in fintech on NLP tasks like customer sentiment classification. We found LLMs and LLM-powered agents perform better than traditional deep learning. (18:59)
Michael: Because of my games background, I used agents with ChatGPT in a book teaching Python game development using Pygame. (19:42)
Alexey: That’s a different book then? (20:04)
Michael: Yes, a different book about teaching people to code games using AI. I was using GPT-3.5 and starting with GPT-4, with much prompt iteration. (20:11)
Michael: Now with GPT-5, fewer prompts would likely be needed to produce the game. (20:23)
Alexey: So we’re discussing the AI Agents in Action book, correct? (20:28)
Michael: Yes. (20:35)
Alexey: I looked it up. It’s by Ming, published in March, and you’re working on the second edition because the field moves fast. Last year we did a course about LLMs, but this year I had to redo many parts. (20:42)
Michael: AI agents have come up really fast, but we've used them differently over time. The way I teach AI agents now embraces minimalism, keeping agents as lean as possible. (20:49)
Michael: I’ve seen people bulk their agents with many tools and instructions, but I recommend breaking up the workflow into tasks assigned to individual agents. These agents can run in a flow like an assembly line or through orchestration where one agent calls the others as needed. (20:57)
Alexey: I have a pet project — a single agent that writes code to create websites. It reads and writes files, lists files, executes bash commands, and grabs data. It plans and executes the plan well. (21:52)
Alexey: I’m thinking to split it into multiple agents: a requirements agent to clarify what the user wants, a planning agent to create the execution plan, and an execution agent that carries out the plan. Does this approach make sense? (21:59)
Michael: Yes, that’s a good way to break it down into tasks. You could have a flow or an orchestration agent that reviews output, compares it to requirements, and decides whether to rerun or continue. (23:18)
Alexey: Could you explain the difference between a flow and orchestration? (23:42)
Michael: A flow is like an assembly line of agents where each does a task and passes output to the next. For example, requirements, planning, and execution agents. (23:48)
Alexey: So the flow is sequential: requirements, plan, execution. But usually there’s back and forth, especially in the requirements stage to clarify details. This is like a conversation until agreement before passing to the next agent. For users, it looks like one chat, but under the hood multiple agents handle parts of the process. (24:08)
Michael: Yes, communication patterns vary, but flow is sequential agents. In orchestration, a front-facing agent manages interaction and calls planning and builder agents. It inspects outputs, providing feedback loops to meet requirements. (25:20)
Michael: Orchestration is more complex since it must understand other agents’ actions. Some platforms like AutoGPT use collaboration, where agents communicate directly with each other in feedback loops. This is powerful but expensive and not good for real-time responses. (26:25)
Alexey: When is collaboration a good pattern? What are the use cases? (27:40)
Michael: For very complex problems where you know input and output but the process is intricate. Collaboration agents can work together, iterating and refining solutions like evolutionary algorithms. (27:45)
Alexey: To make it concrete: in coding agents, one could be a designer, another front-end developer, another back-end developer, collaborating by sharing output and feedback. (28:46)
Michael: Yes, collaboration means agents talk to each other by sharing outputs that become inputs for others. They use a shared message channel to exchange context continuously. (29:05)
Alexey: So collaboration allows parallel work, unlike the sequential flow where tasks happen one after another. For example, front-end and back-end development can run in parallel, sharing completion messages. (29:35)
Michael: That could be either collaboration or orchestration. You may have a manager agent orchestrating parallel development tasks, switching between workflows, ensuring requirements are met. (30:09)
Alexey: Do you describe all these patterns in your book? (31:10)
Michael: Yes, the second edition covers architectural and communication patterns, with chapters detailing management and control of agent workflows. (31:17)
Alexey: Do you use a library for this or show how to build from scratch using APIs? (31:24)
Michael: I like using the OpenAI Agent SDK it’s simple and includes features like guardrails and handoffs, making it good for demos and teaching. (31:31)
Michael: It also integrates with MCP servers for reasoning tasks. For example, each agent can talk to a sequential thinking server that acts like a workspace to plan and reason through tasks. (31:51)
Michael: The sequential thinking server stores plans and thoughts. Reasoning agents use it to write out their thought processes, which can be seen by users as the agents work, even though they interact primarily with one agent interface. (32:05)
Alexey: Can this be done with just the Agent SDK? (32:59)
Michael: Yes, alongside an MCP server. The sequential thinking server is a scratchpad for reasoning and planning. (33:08)
Alexey: So it’s not for inter-agent communication but for helping agents think through problems? (33:20)
Michael: Exactly. It’s like prompting agents to think step-by-step, improving answers by allowing them to give feedback to themselves as they generate output. (33:25)
Michael: This reasoning is separate from the output passed between agents. Usually, agents don’t share their reasoning steps with each other, just the results. (34:03)
Alexey: Which chapter covers this topic? (34:29)
Michael: Chapter four in the second edition. Earlier books mention reasoning and evaluation patterns, but this is a deeper treatment of reasoning and planning tools like sequential thinking servers. (34:34)
Alexey: Is it available in Meep already? (35:04)
Michael: I think it was added recently or will be soon. We’re watching for updates. (35:09)
Alexey: Thanks for sharing. I’m also interested in game development. It seems closer to hardware and more complex because it involves languages like C++ prone to errors and segmentation faults. (35:17)
Alexey: How do coding agents deal with this complexity? Are games created with them easy to maintain and play? (35:42)
Michael: Agent patterns can do powerful things in game development. I like playing Spider Solitaire but was annoyed by ads. I asked various LLMs to create a Spider Solitaire game. (36:29)
Michael: Recently, GPT-5 Pro built a complete Spider Solitaire game with card images from GitHub. It required some bug-fixing but the game is fully functional from end to end. (36:58)
Alexey: Was it a React app? (37:41)
Michael: Yes, just a HTML/JavaScript React app. (37:46)
Alexey: Did you do this in the web interface or a coding agent? (37:51)
Michael: GPT-5 Mini works well in VS Code and Cursor, but I prefer GPT-5 Pro despite its slower pace because it generates high-quality code. (37:56)
Michael: I’ve asked GPT-5 Pro to build microservice agents an interface orchestrator, backend image creation agent, web search agent all deployed with Docker Compose. It delivered a zipped package that runs with a single command. (38:07)
Alexey: And it just worked? (38:38)
Michael: Yes, after a couple bug-fix iterations. (38:45)
Alexey: That’s impressive. (38:51)
Michael: You can do a lot now. Returning to game development, a few years ago I wrote a book called Generating Reality about the impact of generative AI on gaming. (38:57)
Michael: Generative AI will change gaming by creating real-world models. The next platform may not be code but generative AI producing whole games without traditional artists or designers. (39:17)
Michael: The games could be infinitely playable and embed discovery, far beyond simple current games. They might look like watching high-quality video, the best graphics ever. (39:48)
Alexey: I spent a lot of time playing Diablo I when I was younger. It’s an RPG where you have a character and fight monsters. I imagine if all the levels, quests, and challenges were automatically generated every time, it would be really cool. (40:13)
Alexey: Every time you play, it would be different, so you could play forever. (40:51)
Michael: You could prompt level one and get a version of level one, or prompt level 99 for that level. It might not happen next year, but that’s definitely the direction we’re heading. (40:57)
Michael: Not just fully generative AI applications, but generative AI producing both front and back end work. Eventually, we might have competent AI opponents in games, something gamers have wanted for a long time. (41:14)
Alexey: I was given a test to implement Space Invaders. It’s not very complex, but you need to model bullet physics, collisions, and manage things moving at the same time, which needs some multithreading. GPT-4 couldn’t solve it fully, but with 10-15 prompts, I got it to work. The same with Claude One, though it needed many iterations. (41:42)
Alexey: Hearing your story about Spider Solitaire made me want to try it again. Maybe I’ll see how far it can go with Space Invaders. (42:46)
Michael: You might be surprised. These models have become quite capable. (43:06)
Alexey: So GPT-4 Pro is required? I need to upgrade? (43:11)
Michael: I’m not trying to sell Pro, but yes. (43:18)
Alexey: I have a paid version but guess it’s not enough. (43:24)
Michael: I spend a lot on AI service subscriptions every month. (43:29)
Alexey: Is this available through API? (43:37)
Michael: Yes, with high reasoning settings. (43:43)
Alexey: So it’s paid separately from the subscription? (43:50)
Michael: Yes, and you have to verify your identity with OpenAI. (43:55)
Alexey: I tried to access reasoning tokens and got a message about security checks—like being a suspected spy. (44:01)
Michael: They hide reasoning outputs so others don’t train on them. It’s likely they scraped outputs and trained on them, but no one knows for sure. (44:23)
Alexey: We tested some new Chinese models. (44:37)
Michael: I’ve tried some vision and image models from China, but not many language models because we need to be careful with corporate usage and Azure. (44:42)
Alexey: For professional use, it’s better to avoid those for now. (45:07)
Michael: Yes, but smaller, task-focused models are coming. Most current models are huge and general-purpose, but future models will be more efficient and specialized. (45:14)
Michael: People will want models running locally because paying for hosted models and bandwidth gets expensive. GPUs are becoming affordable enough to run these models themselves. (45:40)
Michael: We’ve worked with the OpenAI 120 billion parameter open-source model, which is very capable. (46:14)
Alexey: That’s open source? (46:21)
Michael: Yes, we use it with Brock Labs’ GR OQ, a great provider offering many models. It has low latency 1 to 2 seconds compared to 4 or 5 seconds of GPT-4 or slower experience with GPT-5. Latency is a big issue. (46:28)
Alexey: We’re still streaming, but recording stopped due to disk space. Ignore the glitch, we’re still live. (47:05)
Alexey: I tried a new Chinese LLM from Z AI. It outperformed previous agents I used. (47:30)
Alexey: A simple one-agent system completed complex steps from requirements gathering to execution. (48:25)
Michael: More focused models will emerge. OpenAI’s big models are general purpose, but smaller, task-specific efficient models will do the same or better work. (48:40)
Alexey: I see more questions. One asks about your career before AI. You mentioned development work with oil and gas companies. (49:13)
Michael: Even back then, I was doing machine learning, data science what some call AI now. Before that, I did game development courses involving 2D and 3D graphics. This gave me an edge in understanding the math behind AI and deep learning. (49:31)
Alexey: It’s common for game developers to move into deep learning because under the hood it’s matrix math and tensor operations, similar to graphics calculations. (50:01)
Michael: Exactly. (50:18)
Alexey: You mentioned earlier work diagnosing ADHD with games. Did you continue research in that area? (50:25)
Michael: I worked with Harvard professors in the early 2000s on the application but moved on later. I believe they continued with it. Using technology to diagnose children is very capable, and some games help in this. But gaming is fun, and not all kids engage with the same themes or types of games. (50:43)
Alexey: I also worked on cognitive tests at Boston University as a bachelor student. They weren’t entertaining though more like routine tests, unlike games like Mario Kart or Minecraft that kids enjoy. (51:44)
Michael: Yes, the theme matters because if kids don’t like the game theme, they’re not interested, which can skew experiments. The idea was good for the time, but maybe now we could do better with open source or new tools. (52:24)
Alexey: In our tests, students took tests for educational credits, not real money. Maybe kids would prefer rewards like cookies to make it less boring. (52:49)
Michael: If there was a formula to make a game everyone wants to play, it would be overused because the game industry is highly competitive even more than AI, though AI is rising fast. (53:18)
Alexey: For Albin’s question, could you recommend keywords to research this field? (53:36)
Michael: Look into cognitive testing and executive function, which covers brain functions like planning and reasoning and relates to ADHD and other conditions where focus and planning are challenging. It aligns well with reasoning and planning in AI. (53:47)
Alexey: A fresher with a master's in data science asked whether they should focus on machine learning, AI, or LLMs to secure a job. Some data scientists struggle to find work compared to ML engineers. What would you advise? (54:20)
Michael: I suggest focusing on AI engineering. LLMs embed a lot of knowledge and can perform many data science tasks, like merging spreadsheets and creating graphs. (55:16)
Michael: Data science itself might not be the best continued path, but data science skills evaluating experiments, working with data are useful in AI engineering. (55:57)
Alexey: When AI emerged, I was confused by RAG (retrieval-augmented generation), but then realized it’s essentially search and information retrieval plus prompt engineering. Past experience with search and recommendations is very helpful now. (56:24)
Alexey: Data scientists have valuable skills that remain relevant with LLMs, just requiring some new twists and continuous evaluation and experimentation. (57:28)
Michael: Exactly. One core concept I teach is building evaluation and feedback mechanisms. You want to assess agent performance for consistency and understand variance in outputs. (57:39)
Michael: For production applications, controlling variables and explaining agent behavior is vital. Build evaluation pipelines and use tools like Arize Phoenix to monitor LLM communication and prompts, applying data science techniques for success metrics. (58:04)
Alexey: We cover Phoenix in our LLM course. For those interested, check out LLM Zoomcamp. (58:38)
Alexey: Before wrapping up, about your second edition of AI Agents in Action if someone buys the first edition now, will they get the second automatically? (58:50)
Michael: I need to confirm with the publisher, but I think the second edition will be provided free or as part of the package since it’s releasing within the same year. (59:07)
Alexey: Maybe to be safe, we can wait for the Meep announcement, which might come soon. (59:30)
Michael: If viewers already understand agents well, they should wait. The first edition covers fundamentals, but the second has more advanced and valuable material. (59:44)
Alexey: We’ll follow you on LinkedIn for updates. Is your name common or unique? (1:00:04)
Michael: My name’s spelling is unique, but some others have it too. (1:00:10)
Alexey: You’re recognizable. If someone sees a person with a big beard named Michael, it’s probably you. We will include your LinkedIn link in the description so people can find you easily. (1:00:23)
Alexey: Thanks so much for sharing your expert knowledge and experience. I enjoyed this conversation and I’m sure the listeners did too. (1:00:40)
Michael: Thank you very much for having me. (1:00:48)
Subscribe to our weekly newsletter and join our Slack.
We'll keep you informed about our events, articles, courses, and everything else happening in the Club.