Podcast：InfoQ 2026 年 AI、ML 與數據工程趨勢報告

Transcript

Podcast and Panelists Introductions [00:38]

Srini Penchikala: Hello, everyone. Welcome to the 2026 AI and ML Trends Report Podcast. Greetings from the InfoQ AI, ML and Data Engineering team. We also have a special guest today for this year's trend report discussion. This podcast is part of our annual report to share with our listeners what's happening in the areas of AI and ML technologies. I am Srini Penchikala. I serve as the lead editor for the AI, ML and Data Engineering community on InfoQ. I will be facilitating our conversation today.

We have an excellent panel with subject matter experts and practitioners from different specializations in the AI and ML space. I will go around our virtual room and ask the panelists to introduce themselves. We will start with our special guest first, Chip Huyen. Hi Chip, thank you for joining us and participating in this podcast. Could you introduce yourself and tell our listeners what you've been working on?

Chip Huyen: Hello, I'm Chip. I'm very excited to be here today. It's an honor. I know that everyone on the panels has amazing background. I've been following your work for a long time. Today, I was hoping to share what I know, but also hoping to learn from you all. I'm working on a lot of AI stuff, AI agentic, getting AI to work well for workflow. I'm also very interested in making AI just faster and cheaper. We saw recently the TypeSafeAI that came out. I think there's a huge demand for cheap and fast intelligence, and that's very exciting.

Srini Penchikala: Thank you, Chip. Yes, definitely if we have time, I would love to learn more about this Jev model from TypeSafe AI, if you have any experience on that. Daniel, how about you?

Daniel Dominguez: Hi, I'm Daniel Dominguez. I'm the deputy editor at InfoQ. I focus on artificial intelligence, machine learning and data engineering. I'm also a software engineer and really particularly interested in AI agents, agentic software development and how AI is moving from pure digital system into Robotics and Physical AI.

Srini Penchikala: Thanks, Daniel. Arthur?

Arthur Casals: Hi, everyone. I'm Arthur. I'm a researcher and technology strategist, basically. And these days I split my time between consulting, advising companies about how to make AI useful in real systems and applying research basically on multi-agent systems and belief revision.

Srini Penchikala: Thank you. And, Hien?

Hien Luu: Hi everyone. My name is Hien Luu. I mainly focus on ML infrastructures as well as, I guess nowadays, agentic engineering. Currently, I'm helping with a startup on consulting on AI inference and a long-term fan of QCon, also an editor of InfoQ, also a QCon program chair for QCon AI New York.

Srini Penchikala: Yes, thanks, Hien. I am definitely very excited to talk with you all about this topic because you have been part of InfoQ and the QCon family for so long. It's always great to chat on these topics. So I'm looking forward to speaking with you about what's happening in the AI and ML space, where we currently are, but more importantly where we are going at the fast pace as the AI technology innovations keep happening since we discussed this report last year.

Before we start with the actual podcast topics, I want to go over a couple of quick housekeeping items for our audience. There are two major components to these trend reports. One is obviously this podcast, which is an opportunity for you to listen to the panel of expert practitioners and see how the AI innovations are disrupting the industry. The second part of the trend report is a written article that Arthur also mentioned that we can cover a lot more topics in that article.

So that will be available on InfoQ website along with the podcast recording. That's where we actually assess the current technology adoption and we have an adoption graph that we update every year to highlight which technologies are further ahead in the adoption cycle than the other technologies. It's always a great reference for anybody who is following the technology innovations in the AI space. Definitely look for those two parts, the podcast as well as the article. And also I definitely recommend that you check out all InfoQ publications.

In this podcast, we will definitely discuss the important developments in AI technologies that took place since our last report back in August last year. But due to a limited time in this forum, we won't be able to cover all the topics or even get into in-depth discussions on some of the topics, but we will try to follow up on those important topics as separate podcasts or separate articles because they're important for us to learn about. But again, I encourage you to check out the InfoQ website and AI and ML and data engineering community, which includes news items, articles, podcasts like this and presentations.

It's also worth mentioning with all the AI stuff happening that this podcast is not AI-generated or even AI-assisted. It is being recorded with real people who have real-world experience. I just want to get that disclaimer out of the way. So contrary to all the predictions being made out there that AI is going to replace everything. So that being said, let's get started. It's been almost four years since ChatGPT came out back in November 2022. It's been two or three years since RAG Technologies started seeing enterprise adoption and autonomous LLM agents came out. So definitely a lot of things have been happening.

State of AI in 2026 [05:58]

Agentic AI has been the main dominant topic for a long time now. So we'll definitely look at that a little bit today. We are seeing more developments in the quality assurance space with the AI evaluations and harness frameworks and all the other good topics. We'll get into more of those details later, but the first topic is I would like to get a quick take, maybe 30 seconds take from each of the panelists, on how do you see the current state of AI in 2026? Chip, could you start us off?

Chip Huyen: So it's going absolutely crazy. So I'm checking on the money in and out of the AI company. So money in, as in private investment and then money out, acquisitions and IPO. So for many, many years, we see the charts very similar, one of the three numbers because the money in has to come out at some point. But in recent years it's grow like this. The money in grow exponentially and the money out is no money out at all.

It's very exciting. It has the amount of money and it's I think proportional to, I guess I use as a proxy for people excitement. A lot of companies, the number of company size is going down as the average median company size going down significantly. So that means that people can do a lot of things that anyone can build nowadays. Anyone can have a product. So the issue is not really about how to build things anymore. Anyone can build things, but what you build and then how do we as a receiving end filter through all of that is insane.

And I think going through these things like winner-catch-all catch situations because a few products somehow still face and everyone just talks about them and the rest is lost in the shuffle. And it's very, very tricky. It's just like everyone is forced to face the same ideas and at the same time we have a lot of new things coming out, but somehow everyone's talking about the same thing. We don't get it. Yes, it's very, very strange.

Srini Penchikala: How about you, Hien, what are you seeing as AI in 2026?

Hien Luu: Yes, obviously like Chip said, there's a lot of excitements, innovations, product announcements, but if you want to talk about at the beginning of 2026 or the end of 2025, you're talking about the year of agents. I think we are in a phase where the AI agent stack is being built in a way, the different layers. There's a lot of progress there, a lot of exciting, a lot of innovations for all of us, how to learn, how to build AI agents.

And now everybody's trying to figure out how to operate them, how to bring them to production. So that's a very exciting space from an enterprise perspective. I think the other one that I'm sure everybody is aware of is writing code. It's easy now. We're coding agents. So I think there's many exciting ideas going to be built and we all need to adjust in terms of how we build software going forward. So I think that's another area that's happening in 2026 and we'll definitely see more excitement in 2027.

Srini Penchikala: I think we'll get into those topics a little bit more Hien later anyway, so we'll have time. So how about you, Arthur? What do you see at a high level?

Arthur Casals: Yes, I think  - following up on Hien, I mean, I think what we're seeing now is we're changing or shifting our perspective from the parts of the car to the car itself. I mean, you have impressive isolated capabilities and each model will be developed year after year. And now you're starting to realize as companies in production that we need them to work and to be dependable.

So I mean, you can have a pilot of a great agent that can work beautifully, can have created data and that's great, but very good prediction that you have your data is messy, you have different permissions, you have regulations, the system has to be affordable, it has to be recoverable.

So that feels like the defining test for AI that we're seeing emerging right now is not whether it can do something once, but whether a team can trust it in a real workflow to solve real problems. So that's when you see it actually being applied in companies that have been successful.

Srini Penchikala: Thanks, Arthur. How about you, Daniel, what are you seeing?

Daniel Dominguez: I think the biggest change in 2026 is that AI is moving from being something that we interact to something that we delegate. Like Hien was saying, I think most AI applications over the last few years were based on a simple interaction. You will just ask a question and the mobile gives you an answer. Right now we're starting this year, we are starting to see a system that can plan, can use tools, can execute actions, can do their own work and continue working for much longer period of time. So I think this year is less about another jump in the model intelligence and more about turning that intelligence into actual systems.

Large Language Models (LLMs) and Frontier Models in 2026 [10:36]

Srini Penchikala: Thanks, Daniel. Thank you for that. Yes, that's a good segue from there. With AI agents and all the other innovations happening, the foundation is still on those language models. So maybe that's where we can start our main discussion today because there's always a lot of new developments happening there. So in 2026, my observation is the language models have shifted from these massive size models, billions of parameters into the smaller size models, small language models where they can be faster in terms of performance, may not be as accurate, but accurate enough for most of the use cases.

And also we are seeing the rise of reasoning models that can think through math and logic problems before actually responding. So a lot of good stuff happening. So we can start with this as the main topic. Arthur, I know you've been looking at some language model topics and also frontier models. So what do you see is happening here, what our listeners should be aware of or should be learning about?

Arthur Casals: Yes, I think in general, when you focus on specific models, we have better reasoning, we have better coding, we have better multimodal abilities. They're emerging, it's evolving. But then we're going to go back to what I said first, that if I were to pick a headline for frontier models in 2026, I'd say that the model itself is becoming less of the whole story. So you do have the capability, you do have to worry about it, of course, but we're changing the question from which model is best to which model or combination of models gives us the right capability, latency, data boundary, cost for this workload, specifically this workload?

Which, to me at least, the trend seems to be a move from choosing one model to delivering matching models to workloads, which makes it a system design question, not a model choosing question. And I see that as a sign that a field is maturing because the model stops being the whole problem and becomes the one component of the architecture. So we do have to worry about all the specific capabilities of each model, but they have to work in unison, together.

Srini Penchikala: Any other thoughts from anyone else on the models?

Chip Huyen: So I think that model size is a little bit interesting because the definition has changed over time because the model size, number of parameters, it also depends on how many of them are actually active at the same time. It's a trillion parameters model with only a few percentage. And also another is the quantizations. I do think this very, very exciting work on quantization recently. I feel like about two, three years ago, people thought it was crazy to have a 4-bit model, but now they have consistently see people running 1 to 2-bit model, which is a significant reduction in size. And we think of it as a compute requirement, a computer efficiency of these models.

So I think it's interesting to see. Another trend I see from models is that of a release cycle because let's say that you have a very, very strong model. You release this out at the top of every benchmark. And then a few weeks later it released another model much cheaper, smaller, almost like equivalent, but maybe 1 or 2% drop in performance. So I think this is usually a trend of yes, we have a big model and then you distill that into a lot of other techniques, like using a better mixture of expert, making it more sparse, quantization aggressively one time into a smaller and a cheaper model.

Hien Luu: Yes. There was a lot of interest in doing fine-tuning toward the beginning of the year, and there's a lot of innovation in that space with LoRA and all that stuff like that. And all these kind of neocloud inference providers started to provide the ability to do fine-tuning for at least the open source models. I would love to hear from the panel if you have heard successful use cases in the enterprises, they actually do fine-tuning for their specific use cases. I think there's some happening, but I don't hear much about it.

Obviously it takes a lot of effort to do that with the right skill sets, maintenance, operations, doing inference, the whole shebang. It's doable, but whether that is make you reliable and in operations and all that stuff like that is still a lot of work. So maybe things will be easier in 2027 and people do it more, but that's still a question in my mind of how pervasive that is being actually done in enterprises.

Srini Penchikala: That's a good point, Hien. How do you make the model management life cycle similar to software development lifecycle? How do we make it repeatable and streamlined?

Hien Luu: Exactly. Yes. Unlike traditional ML, you start with data and then you train and evaluation and then you bring it online and here you do fine-tuning. So it's a lot of work, but I'm sure there are specific use cases, particularly perhaps these AI-native company that in the domain of legal, Harvey and such. I'm sure they do fine-tuning there. But beyond that, I wonder if enterprises, they actually do it. That's my question.

Arthur Casals: I've seen a couple of companies doing that, but they're very specialized in what they do and they're very, very careful about IP. So they won't disclose that for a few years at least I think. But usually I'm seeing this basically in healthcare and finance. It's concentrated there for some reason. You know the reason, but yes, it's happening, just need the budget, but it's already happening. It's great.

Hien Luu: Makes sense.

Chip Huyen: Yes. I was just about to ask, has any of your company fine-tuned any model? Arthur mentioned it's waiting for budget, it's supposed to be expensive?

Arthur Casals: I think it depends on what you're doing, how you're doing it. For example, when you come to open models, I've seen a small rush of open models and local models and people thinking that, "Okay, it's going to be cheaper". But think about open. When I have the word open, I have open weights, open licensing, transparent training data. I have a lot of things involved and they don't necessarily mean lower costs or commercial license or whatever.

So depending on what you're doing, you need massive data or you want to get trace out of big models to train smaller models. Thing is, I think it's still more of a craft than, an art thing than a process. So people are discovering this. And although you have the process for training models and for training things, applying that to your real world problems is still a process of discovery. So they end up spending a lot of money. I think that comes naturally. It's R&D basically.

Model Distillation [17:09]

Srini Penchikala: Yes. Hopefully some of these learnings from these companies will be shared with the community next year, the next couple years as Arthur you mentioned. Yes, I want to maybe talk a little bit about the model distillation topic. Chip, I know you talk about this in your book titled “AI Engineering”. I highly recommend everybody to check out the book by Chip, one of the very valuable books in the AI space. So you talk about model distillation. Can you discuss how it works and how it can help the overall model usage?

Chip Huyen: Yes. So I think model distillation is extremely important. It's very exciting. And I think a lot of us today benefit from model distillations either done by some providers or some application developers. So the idea is that you have a very big model that performs really, really well. But of course these big models are very expensive to run and of course a lot of them are very slow. So can we just get another smaller model to imitate a load from these big models so that you can achieve comparable or most comparable performance, but at a much cheaper price.

So a lot of time you can get, so we care about something called compute efficiency. So we usually think about, "Okay, for this much drop in the compute efficiency, how much is the performance we have to sacrifice?" And this metric has gone really, really, really well in recent years. I think that people are aggressively developing a lot of new techniques and it's working things that I didn't think would be possible even just a few years ago.

Srini Penchikala: Also in the models we have local versus on the cloud models, right? We can run some of these models locally using Ollama or the other LLM engines. And also these are multimodal models for the text and images. So anybody have any comments on those topics running these models on-prem versus cloud or the multimodal models?

Arthur Casals: I think regardless of the objective of defining what you're using, you need to separate a few questions about what we're doing because have you seen people defining on-prem versus cloud as a full thing or really separate, really binary thing. I mean, it depends on where the client runs, where the model runs, where the prompts to outputs, logs go, who keeps them. Everything's involved.

So in the end, both setups tend towards a hybrid approach. They're neither fully on-prem or in the cloud. And thing is, why is it important to define that for the company? What are they trying to solve? Is this a budget problem? Is this an architecture problem? So I mean, okay, you can categorize them, but what for,  what are we doing with it? So which questions are you asking about this? It's not as simple as choosing a category and that's it.

Srini Penchikala: Thanks, Arthur. Yes, definitely a lot of other stuff happening. One thing I just want to mention briefly about the Jev model side, these are what they call System One models. I just started learning about these. I have no in-depth knowledge on this, but yes, I think we will definitely share more about this in the future. But has anybody else done any reading on this, like this Jev model from TypeSafe AI that they talk about very fast inference and very low cost compared to the LLMs?

Arthur Casals: Yes, it's supposed to be my weekend experiment for tomorrow.

Srini Penchikala: Cool.

Hien Luu: I've been watching a lot of videos on this and recently I watched a podcast from Latent Space where they interviewed the founder of TypeSafe. Very, very fascinating piece of technology. One of the key things I took away from that is a different type model, but it's also extremely fast. I think that's going to help unlock an interesting new sets of use cases.

There are a bunch of interesting demos about using that kind of decision to build UI and other use cases that you might have not considered like, "Hey, is that a relevant use case?" But it's a lot for us to learn and figure out how to incorporate this. But I think the key idea is, at least for the founder, is what he wanted to do is to build a model that is for builders and the output is for software, not for human. That's one of the key things he pointed out.

Srini Penchikala: It basically evaluates the state and returns the answers, and is mainly for the software. So we can maybe... Hien, you mentioned about the inference. I heard the same thing from my own experience. 2026 is being dubbed as the year of inference, right? So could you talk about how we are shifting from training models to running AI inference and what does all this mean in the real world usage? What are you seeing in terms of inference frameworks?

Hien Luu: Totally. The inference topics, I mean, it's exploded since beginning 2026. Obviously from one end, you have all these inference stack from vLLM, SGLang, and so on that are helping the rest of us to spin up our own inference service in our own environment. And at the same time, these open-weight models are becoming more and more capable. And that space is initially not a whole lot of people how it works and how to operate it, but now there's a lot more information and resources out there about how to actually use these software, how to optimize them. And at the same time, some of the use cases we talk about the financial companies where you have PIIs.

So the ability to host these open-weight models internally for those kinds of specific use cases are very, very relevant. And I believe that that's going to continue to grow, continue to be the case where a company will do more of this moving forward now that it's a lot more approachable, easier to do. There are some companies that because they want to save cost, they actually want to host internal open-weights models for coding, for example. So for many reasons that inference is going to continue to be a very, very useful, very powerful topic and more and more companies will start doing this moving forward in my mind.

Chip Huyen: Hien, it's very exciting for the work on inference. Do you have a focus on specific chips or hardware architecture to inference on?

Hien Luu: Not so much at this point because AMD is just coming online, but more and more inference providers, at least the neocloud, they want to be able to do inference across chips, Nvidia and AMDs and potentially in the future, Intels and so on. But at the moment, still predominantly Nvidia from what I gather.

Chip Huyen: I'm curious because I think this is a trend of as you get newer and newer chips, more powerful chips, there's many, many, many older generation of chips. And if somebody say, they almost come for free because they are so slow to run and so energy inefficient. So what do we do with them? So I think getting those done on older generation of chips, making things faster could be very interesting.

Having smaller, I think it tied more installations. You got to have smaller, cheaper models that require less complete footprint. You can fit them onto an older generation of chips. And also a lot of upcomers for chip because it's very hard to compute Nvidia on the flagship architecture, but you can definitely compete at a lower tier architecture.

Srini Penchikala: That probably the next area of the competition. I heard companies like Nvidia and Apple are definitely dominating the hardware space in terms of demand.

Hien Luu: Cerebras is an interesting chip player, and I think we'll see more of that maybe later on this year.

Context Engineering [24:57]

Srini Penchikala: Hien and Chip, you mentioned in a previous conversation about agentic loops and harness engineering and context engineering. Can you share, maybe starting with Chip you, can you share what you see happening in the context engineering space?

Chip Huyen: I do think definitely in the last couple of years, agentic coding use cases have been dominating everything. But I think with the TypeSafe launch, it shows that actually people have been so focused on coding. That's a huge tap area for non-coding use cases. For me, coding is definitely very important, but I definitely realize it's not the only thing.

Srini Penchikala: Yes. Arthur, what are your thoughts?

Arthur Casals: Context pollution is actually the biggest problem, the problem worth emphasizing now, because we've seen the past year or so movements toward bigger context windows in general, but they don't really solve the issue if the system cannot distinguish what's useful, what's current, what's alternative information from irrelevant, stale, adversarial material. So I'm bringing something from my own research here, but in general, more context can make an agent slower, more expensive, less reliable.

So the practical question would be, if you have your context window, regardless of the size of it, what enters the context? From which source, with which permission, for what step of the workflow so you don't have this huge pollution. That talks to retrieval quality, source distribution, professions rules, context model in general, context compression. So I mean, there's a huge, huge heap of problems underneath the word context. That's how I see it.

AI Agents [26:36]

Srini Penchikala: Thanks for saying that, Arthur. So yes, in the interest of time, I do want to move to other topics. I want to focus on the agents and the evals a little bit more in this discussion. So we can jump into the AI agents topic. Daniel, I know you've been looking into this space for a while now, so this continues to be one of the major trends in the AI space. Agents, they have gone, like you mentioned earlier, from just coding assistants and now they can act on several different aspects.

Daniel Dominguez: AI agents are probably one of the biggest trends of 2026. What I find interesting is that agents are moving through several levels very quickly. So we started with personal agents helping one person, then we start seeing agents working with teams, and now companies are thinking about agents operating across entire organizations. So I think the important change is that we're moving from beyond just a chatbot model. Instead, we are simply generating an answer. An agent can have goals, can use tools, can interact with software, can make decisions, can keep working until the task is complete.

So I think that changes AI from interface into something closer to an execution layer. We can see this mainly in software development. Software development is probably where this transition is more visible. So we move from autocomplete to copilots and now we have coding agents. We have tools like Claude, Codex, Cursor, Replit, and others that can navigate repositories, can modify multiple files, can run comments, can execute tasks, can inspect failures, and can iterate on their own. So the interesting question is no longer, can AI write code? We already know that AI can write code.

I think the more interesting question is how much of the software development loop can we delegate to an agent while we still maintain quality, security and control. That is starting to change the developer's job. So I think developers are spending less time producing every line of code and there is more time defining intent, giving context, reviewing decisions and validating what agents can produce. So I think the recent developer surveys also show how quickly agent coding has become part of everyday development workflows.

So I think there are a lot of discussions right now on the interface disappearing, the agents and enterprise integration, multi-agent systems, all the challenges that that is bringing. So definitely this is something that we're going to continue to see probably next year and a couple of next years, all this agent and agentic phase working across organization and on a daily workflow of not only developers, but in other areas of the work.

Srini Penchikala: Yes, thanks, Daniel. I think as the AI tools generate more and more code, I don't think that will be any different than humans doing it. Maybe they'll be faster, but in terms of the defects, I think I'm sure the code will not be perfect right off the bat. So how do we see these generated code, how do we make sure that it's safe and secure and in compliance with the corporate standards? Arthur, how do you see if we connect AI agents into their software delivery platform and the SDLC process?

Arthur Casals: That's a big question, isn't it? First, we need to separate coding assistants from coding agents. Assistants would be something like an agent helping a developer navigate a code base, writing a test, drafting a patch, something like that. Agents, on the other hand, means the system can run comments, modify a repository, open pull requests, and affect deployment in any way. And for that, we need operational design. So for operational design, you need all the steps. You need authority levels, permissions, you need to audit your workflows, see what's happening there.

And the thing is, it's really interesting to see how the tools that work with that are emerging because the first ones are obviously based on the human flow. And we just had the discussion about Jev, right? We are seeing more and more tools that are realized that, "Wait, an agent is anthopomorphized here. Okay, but it's not really a person, it's a piece of software. So we need to write tools for the software to use, not for a person to use". We are seeing this change right now. I'm curious to see how it's going to affect this operational design for greater permissions for structuring what agents can and cannot do from the coding agent's perspective, not from the coding system's perspective.

Srini Penchikala: Yes, definitely. That will be a big thing that's going to keep watching, right? Thank you for that.

Hien Luu: Yes. One thing I want to add to discussions of agent is there's two things. One is harness engineering. I think that extensions of context engineering. But the thing that I see is quite interesting is relevant to a workshop that I'm going to be giving is these agent SDKs start out with just helping us to write the loop. But now this SDK has been expanded to provide harness engineering capabilities, including memory management for context, including co-execution, including some of the out-of-the-box tools for doing into actual local file systems.

So that's another trend that you can see that these agent SDK from LangChain or Anthropic or OpenAI, they all move in that direction of adding more capabilities as part of the harness engineering. As anyone building agents moving forward, that's something to look out and leverage. And there's a whole other trends of managing these agents. So these companies are providing a runtime environment for you to host your agent. Whether you want to do it or not, that's a separate question, but that's what they're seeing. "We able to build, we are able to build it well, and then how do we operate and run it at scale?"

Daniel Dominguez: I think there is an architectural shift happening underneath all of this. So for agents to become useful inside enterprises, they need to access tools, data, APIs, and talk to other agents. So all the protocols like MCP or A2A are going to become more and more different. So I think the question would be whether agent protocols become a new enterprise integration layer.

So for example, today we connect applications to APIs. Tomorrow we may increase and have agent discovery capabilities that we see in task coordinating work on top of those APIs. So I think all of these architectural layers are also changing the way that agents interact with each other.

Srini Penchikala: Anybody else have any thoughts on quickly how all this agentic AI systems are changing the team organizations, team hiring and everything?

Arthur Casals: One thing that I can say is that not developing with AI systems these days is unpractical because (it is) very expensive. I mean, people are more and more qualified. So people that are qualified to use AI, they carry that on their price tag, let's put it that way. So you have to use it anyway. So once you get there, how do you hire people to fit your own AI-assisted workflow? I think that's the biggest question because again, we go back to the point that we don't have real standards yet. We're experimenting in companies with different companies and different sizes. It's very different to work in an AI-focused company and in a bank or a retail supplier that has its own IT department using AI for something.

One common point that I've seen is people are focusing on how well a new hire can actually describe or structure the system verbally because, I don't like this, but we go back to prompt engineering. You have to have a full understanding of what you're doing because most people are not really reading the code anymore. I mean, AI generates all this boilerplate and it's very hard to do a code review these days the way you're supposed to do a few years back. So with the lack of a process or tools to deal with that, people are sticking to what they know or they perceive which is valuable at that time.

And the one common point is that new hires that can actually specify the system and get something useful out of AI, something that works within their workflow. And then you have the variance. So if I use a workflow that's based on, I don't know, I have one API model or I have my own workflow pipeline and then I have the specifics. But the common point is that we are skipping the junior level and we're going to the architecture design level because for a new hire to be able to really specify a system, they need a certain baggage of knowledge. So kind of generate the problem how we're going to train your new hires if you are not really hiring junior people, which is part of the mess we're seeing these days. But it's a big discussion.

Hien Luu: I do see one thing that's changing as part of the interviewing process where companies are now asking candidate to demonstrate their competency in terms of using AI tools. So it's one of the new modules. Not all companies do this, but I wouldn't be surprised if more and more companies will incorporate this in the future.

Daniel Dominguez: For example, in my experience right now we stop hiring junior developers. I mean junior developers are no longer needed because of also the front end. Front end, we stopped hiring front end developers, front end is no longer needed. So what we're doing right now with the developer that we're hiring is that, as Hien said, we look for people who know how to use AI tools, who know how to implement all the AI into the workflow. And the developers are becoming more orchestrators, agent orchestrators, rather than seeing the code and the daily things that are happening.

So for example, in our case, we develop a lot of agents for our platform to do a lot of things. So the developers, what they do is to make sure that those agents work fine, that we don't have any, I don't know, data lake or that we control the expenses of the coding, we control all these things that we can control. But the developer, the code, the source code, all that stuff is completely out of our equation right now. And I think that's where the industry is moving. It's not that developers are going to disappear, it's that the developer role is going to change. Mainly it's going to become an orchestrator of all those agents and make sure that the agents work as they supposed to work.

Srini Penchikala: Thanks, Daniel. Also, Hien, and I just want to quickly cover the self-learning and self-improving agents that you mentioned in the offline conversation. So can you share any thoughts on that? That sounds like a very fascinating topic.

Hien Luu: Yes. Obviously this topic is very interesting and new at the same time. I think it's a progression. We know how to build agents, but how do we fine-tuning? How do we make it better? How to make it improve overtime? There's a lot of research out there in terms of techniques on how to build these agents that they can self-improve overtime. The startups are building products where you can give them agents and they will help fine-tune your agents. So it's a very natural trend. I think we're still in very early days of that, but the idea is pretty fascinating to me because at the end of the day, if you're writing an agent, most of your logic's going to be in the prompts, in the system prompts.

And that you have tools and whatnot, but the core logic of your agents is in the prompt. And now we learn doing evaluations and tracing becoming very, very valuable as part of the agent and how to iterate. So I think these bits are coming together of one day we just got to be expressing what we want and then the system can help fine-tuning it and make the agent better over time. But we are still in the early days, but it's a very exciting space to watch.

Srini Penchikala: Yes, thanks, Hien. Also, you shared that link, the article from Anthropic that is very good. Actually we will include as part of the references. So it goes from their transformation from person-computer and then the AI solution to agent working with workers, automated, basically AI building itself. So very interesting.

Chip Huyen: And Hien brought up the point about self-improving for agents. And I think the term is a little bit ambiguous. What does self-improving mean? On the one hand you can just say, "Okay, it can tell the agent to never use Rust", for example, and it remembers to never use Rust in the future. Is that considered self-improving or does it involve more of architectural change? Yes, what does self-improving usually mean in terms of agents?

Hien Luu: Yes, good question. One example, if you think about Hermes as a personal assistant that has that capability to improve overtime once it knows you better, it knows how you work and stuff like that. It's one form of self-improvement, but that's not the only form, but that's an example of how... It's like a human, once they're working with you more and more, they know what you want, they want... the way, the style. So therefore they can be a lot more effective at helping you, your tasks or what you want to do.

Srini Penchikala: Right Hien, and it could include agents capable enough to build and train models themselves with little or no human involvement. So it's the whole life cycle. The whole process is controlled by agents.

Hien Luu: That's another form of self-improvement. So there's many forms of self-improvement.

Daniel Dominguez: Something I would have here is that larger context windows don't eliminate the context problem. So I think agents still need to decide what information is relevant, what can be discarded, what should persist as memory, what information is authoritative. So I think context engineering is becoming almost like memory management for the AI system. So giving an agent more context is not necessarily better if that context is noisy or is stale or is contradictory.

Srini Penchikala: Okay. Yes, let's switch gears slightly. And Arthur, you mentioned you are working on multi-agent systems. Definitely that's another very interesting area. So whatever you can share with the audience, can you tell us how these multi-agent systems work? What are some good use cases?

Arthur Casals: Actually, that's a bit of a sensible topic because when you think about multi-agent systems as a topic, it's been there for a few decades already before LLM agents. So people from the academic community are seeing whatever comes with LLM agents and after transformers and is related to the entire multi-agent systems trying to reinvent things that were there in the past. So my research area is on classical multi-agent systems, so before LLM, which doesn't mean that it can't incorporate or use LLM systems as part of it. But these days I've been working on belief revision. So this is a big part of solving two problems actually in the AI landscape.

One related to hallucination and the second one related to information accuracy and provenance after being processed by the agent. It's between computer science and philosophy, a lot of formal logic, but yes, it's really interesting. One thing that we've been working with, I think beginning of this year, end of last year, formal logic-based profile definition that could be used. It's similar that we can use with the concepts around Jev that we were discussing before. So have a structure input, a structural output and how we can make the AI not the agents refrain from using agent, but actually working from here to there.

Spec-Driven Development and Software Factories [42:58]

Srini Penchikala: Thanks, Arthur. A quick one though, back to you, Hien, you mentioned about Spec-Driven Development and software factories. And I've been seeing this in a couple of companies that are doing this. They're creating dedicated teams to create the software factory model. Can you share more about that Spec-Driven Development, how it works, especially for those who may be new to it and what's happening in this space?

Hien Luu: This probably started maybe four or five years ago. And in terms of with the coding agent now is pretty powerful. But as you build a project, typically you want to have a plan in advance in terms of what you want, the capabilities and whatnot, and not jump to the coding agent immediately to build. So it is borrowing what we know as a good practice to do, but now codified it in a form for the coding agent that you work with. So if you build a brand new, like a greenfield project, for example, you build a spec up, what is it about? What are the capabilities that you want?

And then you have a brainstorming session with your coding agent to provide a context that it needs before you actually go and build it. I think there's a lot of good practices there that people have built around this. And I think most people recognize that's a good practice. I don't hear much more about it now because as people actually practicing it and the AI factory is another that's probably in the last few months of really how to infuse AI into the various steps of your software development life cycle.

The typical use case is, "Hey, somebody raised a bug. How do you get AI to take that from beginning to end of actually take the PR, do an analysis, trying to recreate the problem, come up with a fix, submit a PR, and then you go through the rest of the steps of review and testing and so on?" That is as we all have too many things to do, too many tasks. And I think if we can help automate that, that will accelerate another dimension to accelerating the software development life cycle.

Srini Penchikala: If the AI agents fix the bugs they created in the first place.

Chip Huyen: I want to ask this question about multi-agent system. When I say use multi-agent system, which pattern do usually go for? So for example, is this one orchestrator? Is it speeding up multiple workers? Or do you have different agents working in parallel? And if that's the case, then how do you do the communications across different agents working on the same complex task?

Arthur Casals: Basically, the formal definition goes back to agents working together in a social capacity in terms of their interaction with each other to solve a common problem. And in that sense, they can have a common goal, shared goals, distributed goals, and they can use orchestrators, they can use different architectural patterns. So for multi-agent systems per se, the definition is very broad in terms of how they can appear.

But in practice, what we see is mainly related to orchestration and communication. So since they have to work together and they have to do that in a social manner, which means communicating, so FIPA standards and things like that, everything revolves around orchestration and communication. So I have different protocols for communication, different architectures for orchestration. A lot of work on just these two.

Chip Huyen: Yes. So I'm very interested in agent-to-agent communications. So actually recently I use multi-agent a lot. So I have some complex tasks. I would get a bunch of agents to work together. So I actually build this protocol to let them like, "Okay, when you start working on a task, it's a part of big system. So as an agent working on the same task, so we can message them to make sure that you guys can coordinate and don't overlap". So what happened after that is that my task didn't get done at all. I keep getting a lot of messages and it turned out one agent started forwarding tasks to the other agent.

And then that agents forward that to another agent and then none of them did the work. And I'm spending half an hour just checking, "Okay, where does this go? We end up doing it?" And nobody did it. So I feel like communication is one thing, but for agent to realize, "Okay, this task is relevant in my name and I should do it", versus, "Okay just going to farm it out to a bunch of other agents". So you need to be aware of what other agents are doing and the scope of that. How do you manage that?

Arthur Casals: You try. You don't do it. You try to do that because thing is, depending on how you're tackling problem. For example, one of the simplest ways of doing that is using a blackboard messaging to see what everyone's doing, so they're just posting back and have the shared, let's say, context for that. But what if you have two multi-agent systems talking to each other? The problem starts that you have different IDs, you can have different ID schemas for agents, and you have different vocabulary, you have different definitions for the tasks. You don't have a shared states. You don't have shared states, period.

And then you have research, you have a lot of things. So for the past decades, when you see all the work that was produced, they're trying to tackle specific problems in specific setups. So for example, how do you tackle the problem of one multi-agent (system), so have shared environment, shared everything, trying to... You have to delegate a task that you can't do for some reason. So an agent starts doing something and for some reason wasn't able to keep doing that. You have to find another agent that's willing to do that.

Just for that, you need a setup. Are your agents benevolent? Do you have to put a premium on that or no? Let's take the happy path. All of them are benevolent, are trying to help. Okay, so here comes the big question. How do you define the equivalence of a task to achieving a goal? So I want you to get from point A to point B. Okay, so getting there by car or getting there by bus, it should be the same or not, they have time restriction. One thing that you'd find when you go deep and deep and deeper on that is that if you don't have a very specific shared setup or very specific shared conditions, the difficulty of the problem increases exponentially.

So one of the problems of the multi-agent systems area that we have and we know that we have is that we have a lack of standards. So we have standards for communicating, for example, have messaging standards. We do have some protocols in place, but you don't have a standard architecture or standard state description or standard environment description for that sense. That's part of the problem. You have two multi-agent systems with two different setups, two different structures, data structure, descriptions, whatever, and then you have to start tackling the particulars. So how do you do that? Then going back to your problem, how do you solve that?

You have to start by compartmentalizing your problem like, "Okay, this is my box. So I'm trying to solve this problem in this very specific box, these very specific conditions and set up and descriptions and good luck from here on". As long as you have an alien agent coming up or a different multi-agent system, so you have different boxes. So start over again. That's the usual path. Not sure it actually answers your question, but it is what it is.

Chip Huyen: No, it's a very hard problem. I'm wondering if any of you, also would love to hear from Hien and Daniel, have you used the box orchestrations ability of different models? First of all, in Claude or in Codex, you can ask an agent, "Okay, spin up my workers to do this task to make things faster". Have you ever used them and do you have any impressions on them?

Hien Luu: I haven't done that extensively. Maybe Daniel.

Daniel Dominguez: No. What I do is I work with them like an ecosystem of specialized agents. So for example, you could have one agent which is responsible for coding. We can have another which is responsible for testing. We can have another which is responsible for security, another for deployment. And perhaps you can have a supervisor agent coordinating all of them. So at that point, all of the challenge is to looking like a familiar to a distributed systems engineering.

So you have an identity, you have permissions, you have observability, you have orchestration, you have failure recovery and that communication between components between all those agents. So that's why I think agent engineering will increase becoming more like a system engineering discipline rather than just prompt engineering or orchestrate those agents individually.

Chip Huyen: Yes. So when you mentioned that made me think about the sandbox system, I did see from the trend from this year and last year is a lot of sandbox companies and I've been in sandbox for agents. So do you only using sandboxes for agents? For example, each agent had you work in an isolated work tree or something like that or it's dependent on the task?

Daniel Dominguez: Yes, for me it mainly dependent on the task.

Chip Huyen: So something I've been wondering about, so we have compute architecture, so AWS, we have GCP, we have Azure, and usually a company is a shop right now. "Hey, I'm a GCP shop. So one of my things is run on Google Cloud", or something like that. I wonder will the same thing happens to AI? Because a lot of companies have subscriptions or they have contract with multiple providers. Do you see a future where people will just stay with one provider, maybe OpenAI can have owners this model? Or do you think that people would have a multi-provider system?

Daniel Dominguez: Yes. In my opinion, I think it's going to happen the same that is happening with cloud computing. I mean, you are not tied to just one cloud provider, you work with multi-cloud providers. So with AI, it's going to happen the same. You are going to be tied to multi-AI providers working on different things of your workflow or different task. But I don't see just being attached to only one provider.

For example, I work with AWS, but I have multi-cloud providers working on different states of my workflow with GCP and Azure. So with AI, I think it's going to happen the same. You can say, "Okay, I prefer Anthropic, but in the end we'll end up working with OpenAI stuff, with Grok, with any other providers". So that's going to happen.

AI Evals and Guardrails [53:47]

Srini Penchikala: Thank you. Yes, I would like to touch a couple more topics and we can wrap up. It's been a great discussion, but I still have a couple more topics. I want to make sure that we share our thoughts with the audience. The first one is the evaluation frameworks. Definitely this has been a big topic this year. I even heard in one of the conference presentations that 2026 is the year of AI evaluations.

So I know there are a lot of different ways of doing this. There are harnesses and the LLMs being used as judges, LLM-as-Judge. So how do you measure AI effectiveness, AI agentic effectiveness in general? What kind of metrics do you see in your own experience? How do we make sure that the end-to-end process has enough checks and balances? Maybe Hien can start us off on that. What are you seeing in terms of agentic evaluation space?

Hien Luu: This is a very important topic for companies that are bringing their AI agents to production, particularly the one that for their customers, not so much about internal ones. If you look at example from DoorDash, they build an AI assistant for ordering food. You have other companies that are building for customer support agents for external users. And we all know that LLMs are not deterministics and building agents, they can take the wrong turn, they can make the wrong tool calls and stuff like that. So evaluations is a way to help us to make sure that the agents do the right thing for the right tasks at the highest level is that.

And without evaluation, how would you know that it's gone rogue or do something that's unexpected, not according to your company's policies and all that stuff like that. And there's methodology behind that. And by the way, we're going to have a sessions on that at QCon SF. One of the experts on evaluation is going to be there. And companies not only have to do this, but they have to incorporate this as one of their core steps in building AI agents. And a lot of company ending up building something called evaluation harness, which help drive the whole evaluation flow from getting the data, running the evaluation or tests, as we all know tests, right? And then assessing the quality at the end.

And that help drive the decision whether it's a go or no-go for release. So anytime that you're building serious customer-facing agent, you have to incorporate and spend time and effort of doing evaluation right. If you build an internal agent, you're probably not going to spend a whole lot of time on it, but if you are building a real serious agent for your company, then definitely a must. And you see many examples across companies from LinkedIn, DoorDash, Ubers and such, and many cases study on that.

Srini Penchikala: Thanks, Hien, definitely. I think this area is going to see more it all comes together. So any other panelists, any other thoughts from anybody else? How do you evaluate your own AI systems?

Daniel Dominguez: Yes, I think I completely agree with the idea that this year is becoming the year of evaluations for models and agents. So once an agent can perform an action, evaluating only the model is not enough. So you have to evaluate the whole system, you have to evaluate the model, the tool, the context, the agent loop, the environment. So here is where the metrics changed as well.

So it's not only whether the answer was correct, so it's whether the agent completed the task, how much cost, how long it took, how many interventions were necessary, and whether it did something unsafe along the way. So I think all these evaluations were just starting to see how are we going to evaluate all this workflow regarding agents and regarding LLMs and regarding the interactions that they do in our job.

Physical AI [57:49]

Srini Penchikala: Yes, good point. All those metrics are more relevant now with the overall. Yes. One more topic I definitely want to highlight in this discussion is Physical AI. Daniel, you've been looking at this for a while. We talked about this last year as well. How do we bring intelligence into the real world? So this AI agent is a software system, how do we bring that into the hardware, whether they're robots or whatnot? Now they're being called collaborative robots, cobots, and then even self-driving cars. Daniel, maybe you can start us off on this. What do you see here, what's happening and what we should be paying attention to?

Daniel Dominguez: Basically it's saying, "How do we bring AI into the physical world?" So over the last few years we have spent the time putting intelligence into software. Now we're starting to connect that intelligence to machines that can perceive and act in a physical world. So what you said, we are starting to see robotics, we're starting to see autonomous vehicles, drones, and industrial systems are now part of this transition. What I think makes physical AI fundamentally different from software agents is that mistakes have physical consequences on the physical AI.

So one thing is that in software, an agent can make a mistake generating their own file, but a robot can move their own equator. So it's going to be a whole different level of mistakes. So the reliability, simulations, safety, verification, evaluations are going to become more and more important in the physical world. So based on that, all the foundation models for robots that are happening right now, another important shift is that robotics are starting to benefit from more of the same foundational model approach that transformed the language and vision.

So instead of programming every individual behavior manually, we're increasingly seeing models that combine vision, language, perception, and actions. So is that the robot can potentially understand instructions like, I don't know, "Pick up that object and put that on the table", and translate that into a high level intent of a physical work. So as you mentioned, there are a lot of projects regarding the physical model like Nvidia, like Google robots, where all these multimodal models are taking language and visual information and producing action for robots.

So that connection between language, perception, actions is what makes the space of robotics and physical AI more interesting. So we're going to start seeing things like simulation for digital twins robots, all the things that are going to happen for open robotics, all the things that are going to happen from AI agents to physical agents. All of this is just going to start moving faster and faster over the next year because right now the AI is here. Now we have to put it in our physical world.

Robots and Vision Language Models [59:59]

Srini Penchikala: Yes, that's going to be definitely one of the applications that's going to see a lot more attention, a lot more developments, whether it's manufacturing or any other physical industrial use cases. Any other thoughts? Chip, robots and vision language models, how do you see them working?

Chip Huyen: Yes, I think it's very, very exciting. Definitely a lot will say that it's the next frontier to get the models into the real world and we see a lot of progress. And I think the only challenge is that there are several challenges. We keep saying that robotics is like cars, like LLM to say, "Okay, we know we have a lot of robots in the real world". We have Waymos everywhere. We have Zoox. I'm not sure Zoox is out. So I have never driven one, but I see the advertising for it. So we see a lot of robots in the real world.

So people keep telling me that, "Oh, we know how to train massive LLMs, multimodal stuff. We know how to train robots setup in cars, we should know how to train robots to work with real world". But I don't think it's quite the same. So one big difference between robotics, models and LLMs, it's that the challenge of backtracking. For LLMs, when you're training on a virtual, on the computer, if it makes a mistake, you can run back. You can reset the state and continue. So you can do a lot of explorations and pick the best state.

You can't really do exactly robotics. You can't get a robot, "Okay, there's a staircase that's there". The robot falls and it breaks. You can't just run it back and get the same robot as before. So it makes exploration extremely, extremely hard. And another challenge is data. And so I have been looking into a lot of robotic setups. So a lot of them had to raise a lot of money. So having five of these companies raised what, billion dollars? I don't remember off the top of my head. We have a bunch of companies and a Generalist, Physical Intelligence, Sunday, Skild AI, which I'm very excited about.

If you look into it, the main difference is the approach on achieved data. So I think that people are still betting on what kind of data would help robot models to scale. Something tricky with AI models is that sometimes when it's small, they don't quite work. A model only works really well when they're big enough. For a long time with other language models, they generate trash and they suddenly boom, you have GD3 and GD4 and it's suddenly everywhere. So a lot of the bet is like, "Okay, the models are not working well yet. Is this because the approach is wrong or is it because you do not have enough data yet?"

So people had to make a bet on, "Will this kind of architecture, will this kind of training algorithm, will this kind of data be the right approach to scale?" So we see a lot of these companies making a big bet. So there are some companies that say, "Okay, we need real robotics data". So in China, the Chinese government is investing a lot of money into having a robot training center. We have humans operating real robots and just collect a lot of this data and it's extremely expensive. So a lot of people were like, "Okay, we do need this type of data. Maybe you can just go with some kind of video".

If humans can learn from watching video, can robots do it? So it was collecting a lot of this in videos. We have people paying people in different countries, just wearing the camera on their head and collecting videos. So it's very interesting, but since a lot of companies go into middle ground, like Physical Intelligence, I think it's Generalist, only Sunday they have gloves. So they were like, "Okay, videos are not put in for haptic information". Because a lot of time you can't judge from a video how much force it should put into squeezing an egg or a strawberry.

So you need real haptic information so that you have some gloves, people have to wear them and charge money for the object and get data to train the robots. So it's the middle way because you don't have the full robots, it's a lot cheaper, but it's also not just videos. But anyway, I ramble on. I think it's a very exciting space.

Srini Penchikala: Thanks, Chip. Yes, we are running short on the time remaining in this meeting. I know there are a lot of topics. We have talked about the topics that we think are definitely important to be aware of, but there are other topics that we can probably... We won't have time today, but I just want to quickly mention you can follow up offline as separate podcasts or articles.

So there's AI governance platforms that are coming up and then there's AI security and regulations is an important topic. AI observability is an evolving topic as well. And definitely the AI SRE and AI ops are also getting a lot of attention. So do you see anything else happening at very high level before we wrap up? Arthur, do you have any comments?

Arthur Casals: Yes, I think we could relating multi-agent systems and AI engineering because not as hopeless that might have seen when it's reship, but there's a lot of things that we could bind together and bring together to correlate to how we solve problems that we have using work that has already been developed. So that would be a great topic.

Wrap Up and Predictions for Next Year

Srini Penchikala: We can start wrapping up. So this is my favorite part of the podcast to request each panelist to make a prediction for next year and we'll take a checkpoint next year and see how these predictions have been realized. So yes, we can start off with Chip, what do you see? What's your prediction? If there is one thing you'd like to see happen or you think is going to happen in the AI space, what would it be?

Chip Huyen: I really want to see US robot companies that have robots ready to be bought. I know it's just very painful. I feel like in China you can just buy a robot, in the US you just cannot get from an American company yet. So I really want to see that shipping. I feel like we've been waiting for OpenAI and Anthropic IPOs for a while. I think they delay. Very curious to see how that's going to impact the market because if the IPOs go well, is the market going to go well, but if IPOs don't go well, what implication would it have for the funding and the AI market? I would expect to see a lot of cheaper and faster models, the TypeSafe style of model would be very interesting.

Srini Penchikala: Thank you. How about you, Daniel? What's your prediction?

Daniel Dominguez: I think we're going to start seeing limitations on AI. So for example, the phrase is going to be the energy. So all the AI infrastructure is becoming extremely power hungry. So I think the access for electricity, the capacity, cooling data center locations, that is going to become a biggest constraint for the growth of AI. So we may start seeing energy crisis for AI. The second prediction that I would like to share is like a futuristic prediction, but I think it's going to happen that all this AI infrastructure is going to start moving beyond the traditional data center and it's going to move into space. So I think putting AI into space is going to be more efficient, probably it's going to take a little longer, but it's going to be more efficient than keeping the AI here on earth.

And finally, I think the AI competition between the United States and China is going to increasingly look like a technological Cold War. So it's not about who has the best model anymore or who has the best chips. So it has to be with who has the manufacturer, who has the energy, who has the data center, who has the open models, the supply chains, and the access to compute. So we are already seeing that semiconductor access and domestic chip productions are becoming strategic issues in that relationship. So I think that for 2027, some of the biggest AI stories may actually happen outside the models themselves. And it's going to be more about the energy, the chips, the infrastructure, and on the geopolitics.

Srini Penchikala: And Hien?

Hien Luu: Just two things. Every AI agent will incorporate some form of the Jev model into the AI agents. And more and more of the communications are going to be between agent and agents moving forward as we figure out how to build, operate and bring all these agents to production.

Srini Penchikala: And Arthur?

Arthur Casals: I'll circle back to what I said at the beginning. I think and I expect that the most successful AI deployments that we will see will be much narrower than what we see right now in terms that rather than one agent with access to everything, we'll see a lot of things using smaller specialized agents for mixed models, for well-defined tasks. And the big difference will be in the surrounding system.

So you have trusted data, leave the permissions tests, and does it not back your visibility, but a way to see what the agent did and whether it's actually helped. Teams can not only do that, but can actually measure the cost and quality of a completed task. You can have ROI for tasks, not just the number of tokens of agents that use it. I think that this will be a great leap.

It aligns well with what Hien said on Jev and what Daniel said on the shifting away from big models, but I think that's where I put my chips in.

Srini Penchikala: Thanks, Arthur. Yes, I think I'll take a quick one minute to wrap up my comments. AI is a significant technology shift. And as we all know by now, after almost four years, it is not a fad or some temporary blip on the technology radar. It is here to stay. It is one of the major technology waves in the last, I don't know, 20-plus years. So my advice and request to our audience is we should take advantage of this technology transformation, not be afraid of it. Find ways to use AI tools to automate any workflows in your day-to-day life, whether it is in the office or in your personal life.

I mean, this is a tool that we can use to find more time to do the important things in our lives, let the agents and the tools do the routine tasks. So my prediction for next year is definitely, I agree with all of you, AI will be even a bigger part of our daily lives. And I always say this as a prediction, and I would like to see whenever we don't talk about AI anymore, it has become part of the fabric of our lives. I think then we can say we have really arrived. And so I think maybe more evolution, more development, but I think the niche areas will be settled where AI will be really valuable and where human intelligence is still required.

So thank you all for your time. We took a little bit longer, but there are so many topics and we can go on forever, but I think we need to wrap this up and we will follow up with any other topics later. So thank you for joining today. Have a great one. Until next time, thank you.

  Mentioned:

• TypeSafeAI Jev Model
• Mixture of Experts (MoE)
• Open-weight models
• Model distillation
• AI Engineering book by Chip Huyen
• Multi-modal models
• Latent Space podcast with TypeSafe founder Diogo Almeida
• Agentic Loops
• Harness Engineering
• Context Engineering: Context Pollution
• Spec-Driven Development
• Model Context Protocol (MCP)
• Agent2Agent (A2A) Protocol
• Anthropic article When AI builds itself
• Multi-agent systems
• Agent Orchestration
• Belief Revision
• Digital Twins Robots
• Vision Language Models
• Physical Intelligence
• Generalist
• Sunday Robotics
• Skild AI

About the Authors

Chip Huyen

Show moreShow less

Daniel Dominguez

Show moreShow less

Hien Luu

Show moreShow less

Arthur Casals

Show moreShow less

You can keep up-to-date with the podcasts via our RSS Feed, and they are available via SoundCloud, Apple Podcasts, Spotify, Overcast and YouTube. From this page you also have access to our recorded show notes. They all have clickable links that will take you directly to that part of the audio.