AI, ML, and Data Engineering InfoQ Trends Report 2025

Transcript

Introductions [00:30]

Srini Penchikala: Hello everyone. Welcome to the 2025 AI and ML Trends Report podcast. Greetings from the InfoQ, AI/ML and data engineering team. Today we also have a special guest for this year's trends report discussion. This podcast is part of our annual report to share with our listeners what's happening in the AI and ML technologies. I am Srini Penchikala. I serve as the lead editor for the AI/ML and data engineering community at 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/ML space. Let's start with the introductions. I will go around our virtual room and ask the panelists to introduce themselves. We'll start with our special guest first, Savannah Kunovsky. Hi, Savannah, thank you for joining us and participating in this podcast. Would you like to introduce yourself and tell our listeners what you've been working on?

Savannah Kunovsky: Of course. Thank you so much for having me. Happy to be here and to continue the conversation from London QCon. My name is Savannah Kunovsky. I am a managing director at IDEO, which is a global design and innovation firm, and I run our emerging tech lab. So basically everything that is a frontier of technology, everything that is a buzzword is in the org that I run. And I get to think about how we might design technology that is more human-centered and also about the evolution of how we design technology and how we do design in general. And I do that in our offices all around the world and across many different domains of expertise and work. And I'm excited to share more and think more about the future of technology, how we create in a way that is really good for people.

Srini Penchikala: Thank you. Looking forward to it. Anthony, how about you?

Anthony Alford: Hi, I'm Anthony Alford. When I'm not writing about AI for InfoQ, I'm a senior director of development at Genesys Cloud Services, so I'm really excited to be here, again. This is one of my favorite things in the InfoQ calendar.

Srini Penchikala: Thanks, Anthony. Daniel?

Daniel Dominguez: Hi everyone. My name is Daniel Dominguez. I'm the managing partner at SamXLabs, an AWS partner company where we focus on outsourcing software from Latin America to the US market. I'm also part of the AWS Community Builder Program in the Machine Learning tier, where I collaborate with the community on cloud and the AI initiatives.

Srini Penchikala: Thank you. And Vinod?

Vinod Goje: Hi, I'm Vinod. I serve as the editor for the AI/ML and data engineering community on InfoQ and I'm excited to be here on the panel. Thanks for having me.

Srini Penchikala: Thank you. Welcome everyone. I am definitely looking forward to speaking with you all about what's happening in the AI/ML space, where we currently are, and more importantly, where we are going with the, I don't know, exploding pace of AI technology innovations happening since we discussed this trends report last year. Before we start with the podcast topics, a quick housekeeping information for our audience. There are two major components to these trends reports.

The first part is this podcast, which is an opportunity for you to listen to the panel of expert practitioners on how the innovative AI technologies are disrupting the industry. And the second part of the trend report is a written article that will be available on InfoQ website. It will contain the trends graph that shows the different phases of technology adoption and provides more details on individual technologies that have been added or updated since last year's report. So definitely look forward to it. I recommend everyone to check out those article as well as when it's published in the next few weeks.

Language Models [03:50]

So let's get back to the podcast discussion. In this podcast, we will discuss the important developments in AI technologies that took place since our last report. Again, due to the limited time in this forum, we won't be able to cover all the topics in AI, but I encourage our listeners to check out the AI/ML and data engineering related publications on InfoQ website. We publish the news items, articles and podcasts like this one and also presentations from different conferences. So let's get started. Since ChatGPT was released back in November 2022, so almost three years, generative AI and large language models or LLMs have totally taken over the AI technology landscape. Major players in the technology space have continued to release newer and better versions of their language models.

Especially in the last year, we also saw other innovations like AI agents and the protocols like model context protocol MCP and A2A from Google, have definitely dominated the news in the space. So we definitely would like to talk more about those. But let's start with the language models. These are the foundation for the GenAI technologies and have seen a lot of interesting developments in the last year. We saw the release of vision LLMs for video analytics and we saw the multimodal language models. We saw the small language models that can be run on the edge devices and then reasoning models and space models, right? State space models. So a lot of things are happening in the language models. So that's a good topic to kick off this year's trend report discussion. Anthony, I know you've been kind of following closely on this topic in this space. Can you talk about what's happening here and what our listeners should be watching out for?

Anthony Alford: It feels like we could have an entire podcast just about this one topic. I think probably the biggest news lately is OpenAI released GPT-5 and I think they may be surprised a few people. They released it and just you're starting to use it and you don't have a choice. I think that took some people aback. They also did something interesting. They have different versions of each model. They'll have the pro or full model, they'll have a smaller faster model and you could pick which one you wanted based on your task. With GPT-5, it just picks whichever one it wants. Now, I actually kind of like that. I feel like OpenAI was struggling a little with naming their models. They had a model called 4o and they had another one called o4. It has logic to it, but it's not always obvious. So I think that was an interesting development as well. What are your thoughts?

Srini Penchikala: Yes, Anthony, also, they claim that it has some built-in thinking capabilities. I don't know what you have seen, but anybody else have any perspectives on this topic? I know it's one of the big topics.

Savannah Kunovsky: Yes, I think that the reason why they're doing that is to simplify the interface. So much of what happens when we're trying to release these new frontier technologies to people is that they get created kind of in the back rooms and R&D departments by engineers and research scientists, which is incredible work. And we, as engineers, sometimes think that the general public is going to understand our vision and understand the way that we're interacting with something and some of those kind of research-led interfaces.

And I think that what the OpenAI team came to understand is that they actually need to simplify the interface to have a consumer product that is really going to be resonate and be more useful to people. And I think in that way, as we move forward, as the capabilities of large language models and other advanced technologies become the norm and we can say it's a given that they can do interesting things, then the interface layer and how we help people interact with those technologies is going to become the most important differentiator for businesses and one of the most important levers for adoption.

Vinod Goje: And also in terms of naming the models, right, I mean, the most recent model, the OSS, and it has been truly open-sourced. So that strikes with the naming, I believe. GPT-OSS.

Srini Penchikala: To add on to your comments, Savannah, so the multimodal AI where these models are trained on multiple data types, text and images and audio and video, so I think that's kind of what I see as the next big thing, right? So we don't have to pick which model to use based on the data type or use case or the size of the model. We want these models to automatically pick whatever is the right option behind the scenes and do the job and then show us output. So kind of more human experience-based solution.

Anthony Alford: And that's actually what the O in the model names like 4o was meant for omni, right. But then they reused the O with a number after it for something else. So yes, definitely the engineers were in charge on these, I think.

Srini Penchikala: Good point, Anthony. Like anything else, I think they're going through these small language models, large models and the other specific models. I think they will all come together and become that omni language model that we can probably use. Daniel, what are you seeing? I know you have done some work on the models as well, right?

Daniel Dominguez: Yes. I think to me the omni shows that the direction of the models are hidden, so it's not just bigger models, it's just a more human-like in a way that interact across different modalities.

State Space and Diffusion Models [08:47]

Srini Penchikala: Yes, and also, I'm seeing more about these state-space models and diffusion models. Any thoughts on those?

Anthony Alford: I noticed Sebastian Raschka, which he's sort of a popular figure in this space. He had a quote saying, "We still haven't found anything better than transformers". The state space models and also the text diffusion models, they do have advantages. They have properties that make them attractive, but they just don't perform as well as the transformers do at scale. That said, people are still trying and still doing interesting work with state space and diffusion models.

Srini Penchikala: Thanks, Anthony. We can also, start talking about the reasoning models, right? This is where we can segue into agent's topic as well. Agents mainly use reasoning models. What are you guys seeing in terms of reasoning models?

Anthony Alford: Yes, well, and don't let me dominate the conversation, but again, back to OpenAI, their own number, o3, o4. That was their reasoning models. And really what these things do is, they talk to themselves and so they often give you a better answer for something that's a tricky problem. On the other hand, they also use up a lot of tokens doing this, so it's sort of a trade-off.

On-Device Models, Robotics and Physical AI [09:50]

Srini Penchikala: Okay. Thanks, Anthony. So Daniel, I know you looked at some on-device models and robotics. Can you speak about those, how the models are helping running on devices?

Daniel Dominguez: Yes, I think on-device LLMs are going to be a big part of the future, mainly because they balance privacy, latency and the cost, in a way that a closed-only solution cannot do it. So we're already seen this shift with models, for example, with Apple, with the on-device LLMs that were announced for iOS, Meta Llama 3 and the Mistral models are more optimized for edge use, rather than cloud. Even as we were mentioning, the GPT-4o mini from OpenAI, shows that a small model can still provide real value when it's deployed locally. No? So I think the challenge will be the evaluation. So traditional benchmark tell us one story, but real work performance will look very different. So I think what we really need is a more practical evaluation framework that includes things like memory footprint, response latency and domain-specific accuracy to work more with these on-device LLMs.

Srini Penchikala: Yes, definitely. I'm very interested in that edge-based computing, right, Daniel. So that's where the kind of real-time analytics can happen because on the cloud, it's more offline and you have more data available on the cloud, but it takes more time to come back with the answers. Whereas on the edge, on the devices, whether it's manufacturing or any other industry, they need answers sooner than later, right? So that's where the models can help.

Anthony Alford: And that's a situation where these other architectures, the diffusion or state space may have an advantage, they are faster, also use less power, et cetera. So we may see these two things combine.

Srini Penchikala: Yes, definitely, Anthony. I was saying the same thing, like Internet of Things, IoT and small language models, I think they will all really come together, right? And Savannah, it's kind of gets into a little bit of robotics space as well. So what are you seeing from your projects and the clients? What does everyday life robotics mean to people?

Savannah Kunovsky: Yes. Well, I think the conversation about edge is especially relevant when we're talking about coming back to users and what people are going to be excited to have in their homes and kind of in their more everyday spaces. What edge lets us do is design for trust and design experiences of how people interact with technology and how data is processed and where it's stored, in a way that people will hopefully and more likely, feel comfortable than if their data feels like it's being processed and sent off to faraway places that they don't know or understand. So especially when we start talking about Physical AI and capturing data that is inside of people's homes that can feel really precious and intimate.

And it's important that we create products and services that respect the needs and desires of users for us to know that that data is really sensitive and to kind of be careful with it. And I think that with the advances in manufacturing and kind of the race that we're seeing in robotics, there are going to be more devices available that will have advanced technology in them that can support us in our everyday lives. And that's really exciting. And the only way to true adoption of those technologies is to create them, one in a way that's honestly and practically extremely useful if people are going to be paying for these things, but also in a way that people are excited to invite them into their homes, which we all want to keep relatively private.

Anthony Alford: And the reasoning language models seem to be sort of a path towards using them in robotics. I've written several news items for InfoQ where people are taking language models and just asking them to make a plan for a robot to go pick up this thing, bring it there. There's a person named Dr. Jim Fan who's a director of robotics at NVIDIA. I saw a tweet from him recently. He says he believes that the GPT-1 of robotics is already somewhere in a paper. We just don't know which one. And he makes the claim that we will not have AGI, we'll not have Artificial General Intelligence without it being embodied. Basically, without it being a robot. This is actually an idea that's been around for a long time. I remember it in my long ago career in robotics in the 20th century that people were making that claim back then. Now, we still don't have it, but maybe someday.

Daniel Dominguez: I agree with Anthony and Savannah. I think that the next frontier is going to be definitely Physical AI. So it's basically all these embodied of artificial intelligence machines. Following Anthony's samples of this. For example, we have this Tesla Optimus robot where basically it has a lot of AI put into the humanoid to handle repetitive task. Also, we have all the robots from Boston Dynamics integrating these vision language models so the robot can understand now natural comments. Also, for example, the Reachy 2 from Pollen Robotics and Hugging Face, which is a new low-price robot. It's open-source and help developers connect all the LLMs that you see on the Hugging Face platform to the robot.

So I think that the idea is to pair an LLM with a small device with robot or wearable or cars or whatever hardware, is going to move from AI to just start talking to you to an AI that sees you, moves you, interacts with you and with the real world. So I think that's where the things are going to be really interesting, but also, there're going to be a lot of challenges regarding safety, evaluations, alignment because one thing is to have a mistake on the chat window and another thing is going to be have a mistake on a physical work, with one of those devices, no?

Srini Penchikala: Yes, definitely, Daniel. I want to talk about evaluations and safety a little bit later in this discussion, but before we leave the robotics topic, Anthony mentioned NVIDIA as one of the leading companies in the robotics space. I know there are a lot of stuff happening with NVIDIA, right? So are you seeing any new developments here?

Vinod Goje: Yes, definitely. I mean, on the NVIDIA front, right, I mean, we see with CUDA 13.0, we can build robotics application once, stimulate it on high-performance systems like DGX and deploy exact same binary directly onto embedded targets and like Thor without any code changes. So the barrier between simulation and deployment have essentially disappeared. So there are major players in this market like Foxconn and all and XPENG Robotics, they're already using this framework to train their humanoid robots and this is just isn't a research anymore, it's a production ready technology and it can hit anytime. So we'll see a billion-dollar company coming out of this robotics space and we'll soon see robots sitting next to us or maybe walking around us in the park, and it's already happening and we see that.

Srini Penchikala: Yes, definitely. I mean, they can be virtual, right? Savannah also mentioned a lot of these will become the fabric of our everyday lives and we shouldn't even see them helping us, right? So, okay. Thank you all. We can quickly talk about the evaluations. I know with all these language models comes the challenge of, how do we evaluate these. Daniel, there is the LLM Arena. You mentioned about the Hugging Face, the dashboard as well, leaderboard, right? So any specific benchmarks you would like to talk about? Maybe Daniel, you can lead this one.

Daniel Dominguez: Yes. I think one interesting product I follow as you mentioned is the LMArena. So basically, it's an open-source community driving platform where people can compare language models side-by-side, no? So instead of relying on the strategic benchmark of the MMLU or the ERC, the LMArena focuses on head-to-head human evaluations. So basically, it lets user to directly vote on which model gives a better answer to the same prompt. So this LMArena, it's powerful because it includes most of the models that are right now. So from OpenAI to Anthropic, Llama, Mistral, Gemini and all the others.

So all this data comes out from the LMArena helps as to see how the models perform, not just on the academic benchmark, but in real world conversation and functioning task. So for me, I think the big takeaway is that the evaluation shouldn't be just one-dimensional. So I think LMArena is a reminder that user perception does a specific accuracy and overall experience matter just as much as our own benchmark score.

Srini Penchikala: Vinod, are you seeing anything in the benchmark evaluation space?

Vinod Goje: Yes, definitely. I mean, I think LLM evaluation has hit a fascinating inflection point. I believe in 2024 and '25, we are right in the middle of what experts call as an evaluation crisis. The traditional benchmarks we relied on, right, like MMLU, which was the gold standard for years are now essentially saturated. And now, with top models hitting near-perfect scores and making meaningful comparisons nearly impossible. What's real exciting is how evaluation has evolved from simple multiple-choice benchmarks to comprehensive frameworks that does reasoning, coding ability and even agentic behavior. Yes, we are moving from, I believe, from static benchmarks to dynamic evaluation that includes tool use, multi-turn conversations, and real-world coding tasks. So as we move into what we call as the year of agents, evaluation is expanding to include how well models can be treated in improved response, use external tools and maintain context across complex workflows.

Srini Penchikala: Anthony, do you have any comments on the benchmark topic?

Anthony Alford: Just a few observations. It seems like every other week, somebody is creating a new benchmark for some task, which is great. The other thing is, we often see stories of, it seems very likely that this model's training dataset included a bunch of benchmarks, so it's tough to really benchmark things when they're trained on the benchmark. I know that the researchers try very hard to filter out known datasets like that and it's a lot of data. It can be hard to do and everyone always says, the benchmarks don't tell the whole story. It's better than nothing probably, but sometimes you can't really judge the utility of a model to you, from a benchmark and I don't think anybody's figured out how to solve that.

Srini Penchikala: Anthony, thank you. Yes, we can move on to the next topic, the retrieval augmented generation (RAG). I know this was a big topic last year. It's probably has become more of a commodity lately, right? So definitely in our adoption graph, this will be moving up the adoption curve, right? So anything you guys are seeing? I know there's Amazon Q for business and Atlassian also has their own RAG tool. So anyone seeing any innovations here?

Anthony Alford: Like you pointed out, these things are really becoming commonplace. Maybe not commonplace, but they're certainly in the enterprise software space, they're gaining ground. I think any business that has a big document database, bunch of knowledge articles and things like that, they're going to be looking at this.

Savannah Kunovsky: One interesting shift that I've seen in our design process is that a lot of the time, the way that we try to understand what's happening inside of a business is by going and just talking to a bunch of people inside of that business and doing online research about how they're doing and what they've been creating and kind of where their focus is, and CEO messages and all of those types of things. But because of the existence of RAG, we are now able to build systems that actually allow us to gain a ton of context before we start designing.

If we're able to access more documents that can give us that information, then it kind of gives our designers and our design teams the ability to work from a good set of information rather than starting from zero. So it's interesting to see something that started as, I think, a really technical focused innovation, start to have more of an application that's applicable to non-technical folks and hopefully, will become easier and easier to be created by non-technical folks for that reason too. And I think that there's all sorts of business opportunities and just data availability that they provide.

AI from a design perspective [21:39]

Srini Penchikala: Thanks, Savannah. I think that's a good segue, actually. I would like to also talk about the AI from a design perspective. This is your area of focus at IDEO, right? So could you talk about how your teams are creating prototypes and using AI technologies and what else is being disrupted in the design world using AI solutions?

Savannah Kunovsky: Definitely. I mean, there's a ton going on at IDEO. We have, I don't know, 30 or 40 design crafts. So we have your typical visual graphic designers and interaction designers, people who are designing interfaces or designing interactions between humans and hardware or whatever it might be. But we also have business designers and software designers who have software engineering backgrounds and things like that. And so, the ways that these tools are being used across all of these different crafts is incredibly different. But some of the things that are consistent are that they are allowing folks who are kind of in one craft to be able to express their ideas in ways that they couldn't before.

So for example, we have a business designer named Tomochini who's done a couple of interesting experiments where he was doing research with kids to make children's toys that have more sustainable materials. And rather than just coming in and telling them the story or trying to describe the idea or showing them a picture of what these new materials might look like and what the new toys might look like, he was able to create short kind of trailer videos that illustrate his ideas. And I think that's a lot more powerful, is that we're able to just express our intent more quickly and express our ideas more quickly. He also turned business model creation into a game. So he is now able to vibecode these apps and he created this game where he is able to pin different business models against each other.

Yes, so I think in general, if non-technical folks are able to take the capabilities that were once constrained to technologists and express their ideas and make apps or videos or visuals or work with ChatGPT to kind of come up with different ideas and different angles of a pitch for a business, it sort of does this democratization of a lot of capabilities that were once really specialized skills. And I think that's kind of for better and for worse. There's major questions about what an economic transition would look like when these AI tools really have robust capabilities.

And I think that what we're seeing in general right now is that more than anything, we're just able to add a lot of rigor to our design process because we're able to put a prototype of an app or put a full quickly generated landing page of a website or whatever it might be, in front of users as we're doing research rather than something that's kind of sketch level. And so, the rigor of our work has really increased and I think that's super powerful.

Srini Penchikala: Yes, definitely, we can use these tools for ideation, vision boards, right? So a lot of good stuff can happen there. Anybody else seeing any design innovations using AI technologies?

Daniel Dominguez: Yes, I agree with Savannah. I think one of the biggest enablers that we have right now with AI is all these no-code or low-code prototyping. So there are a lot of platforms which you can just work from idea to a working app. We don't need all this engineering part, no? So it lowers the barrier for non-technical to test the concept in a fast way. So for me, power is in the speed. You can validate any idea in days. Before it was months and take a bigger team and now if it works, then you can invest in the scaling it properly with the engineering team.

AI-powered Software Development Tools [25:25]

Srini Penchikala: Thanks, Daniel. Actually, that's probably a good segue to get into the coding side of the AI discussion. I know we have seen AI-powered software development tools right from the GitHub Copilot to other tools, right? So what are you guys seeing in this space? Are there any new trends happening in the software development process that AI is enhancing or enabling?

Anthony Alford: Well, I was going to say, the trend is, everybody's using them. You see a headline every day of some big software company there. I think Microsoft said something like 40% of their code is produced by Copilot now. So as a former professional software developer, my attitude is, "Well, we'll see if this turns out to be a good idea or not". Roland and I were on a podcast last year talking about what's going to be the long-term outcome of this. Nobody knows. As someone who doesn't write code on a daily basis anymore, they're amazing. If I need to write some kind of tool or script that is just for me, it works great. We'll see how it does in production. I guess, Microsoft has figured it out. But long story short, everybody's starting to use them. Nobody's really sure if they're actually helping us in the short-term or the long-term.

Vinod Goje: I mean, nowadays, we have plugins in every other, IDE like take it for a VS Code or Eclipse or IntelliJ, it's all embedded. And as Srini mentioned, it has become a commodity nowadays. That's no longer a differentiating factor and definitely it's helping our developers, testers and all to produce more with less.

Srini Penchikala: Yes. Daniel, how about you?

Daniel Dominguez: Yes, I think the ecosystem on the AI powered software development is getting really, really diverse. So for example, on the coding side, we have two tools as you mentioned. We have GitHub Copilot, we have Windsurf, we have Amazon CodeWhisperer. All of those are focused on code completion and help developer to move faster, no? So I think Copilot is probably the most widely adopted. CodeWhisperer has a strong niche with the AWS environment. Now we have Windsurf and Cursor AI because they don't just autocomplete, it's more like having an AI paid programming inside your IT, which understand the whole project. But now, we're seeing also these by coding tools.

For example, we have Replit, which takes care of the whole project. So it's not just an assistant, it gives you the full flow-based environment where the AI can code, run and even deploy for you. So that's changing the way that the beginners and the small teams are approaching software building.

We also have Vercele's V0, which is pushing AI into the front end world. So you describe what you want and it generates a full react component and designing systems. More recently, Amazon launched Kiro. Kiro, it's aimed to be like an AI powered application builder for enterprise environment. So it's like a competition for AWS, for Vercel and Replit, it's going to be Kiro. So it gives teams the ability to prototype and launch apps quickly with an enterprise creative scalability. So I think what ties all this together is the shift from an AI, just from being an assistant to the AI being a co-creator of the software. So we're just not writing code faster, we're entering a phase where the entire application can be developed, tested and shipped with the AI as part of the development team and not just as an assistant.

Srini Penchikala: Yes, good point, Daniel. I saw the other day Copilot was used to create a end-to-end spring boot application with all the classes and the configuration files. So that was interesting. So going back to Anthony's point, Anthony, the 40% of the code being generated by the tools, I wonder how much of that is quality enough to go to production as is, right?

Anthony Alford: Well, yes, and I just saw the headline. I didn't get a chance to read in more in depth. In my experience, in the experience of everyone who's done software development for a profession, one of my devs told it to me like this, writing the code is not the bottleneck and it may be less than half of the effort. There's design and testing and all the other fun stuff. And in fact, the coding part is the part that a lot of us like to do. So if we're letting the robots do the fun stuff and we have to do the stuff that nobody likes, that seems backwards. So what I do see happening is, people use these to write unit tests. That's the part of writing code we don't like. So hey robot, write unit tests for me or I've changed the code, update the README file, update the documentation.

These things are awesome and I think Daniel, you mentioned some of these tools are even going farther to where it's completely hands off, the so-called agentic mode, YOLO mode, whatever you want to call it, that's kind of cool. People are just giving some instructions to it, telling it to run overnight, these things, write tests, run test, fix the code until tests pass, et cetera. They can put it up in GitHub and do the whole coding workflow. That does seem to be an interesting trend. I think that works good. For example, in the case we're talking about rapid prototyping. Again, my inner skeptic thinks, if you're just launching some unmaintainable code really fast, I don't know if you want to live in that situation for a long time, but it's good to get a prototype. And like Daniel said, decide if you're working on a problem that's a good problem or if you're wasting your time.

Srini Penchikala: Yes, especially if you have to pick a couple of different options and throw away the other options, that's probably a good one to go with.

Savannah Kunovsky: I think that's also the difference between the kind of early stage innovation work that we get to do and some of the later implementation work is that I think sometimes it doesn't matter if it actually works. The kids and the testing of the sustainable materials didn't care if the unicorn had five legs in these videos. But I think that once you get to more production spaces, then those are more important considerations. So I think that sometimes we have this, is AI good conversation that feels very black or white, but it's actually really nuanced to which part of a development process you're working in and who actually is going to be using the thing and why and for what purpose. And so a lot of the time when I talk about this, I try to just get people to think about their own context and what it is that they're trying to accomplish, and then what's appropriate right now and what might be appropriate in the future.

Anthony Alford: Definitely.

Srini Penchikala: Usually software development is iterative process, right? So the code that you write on day one may not be that code that will go to production. So we have to wait and see.

Anthony Alford: And ultimately, that I think is the answer. If this helps us iterate faster, then it's a win.

Srini Penchikala: Also, I'm seeing this AI DevOps getting some attention. There are companies that are specializing in automating the DevOps tasks like CI/CD pipelines and other observability type of tasks using AI tools. Maybe that can be a good complement for this code development tools to kind of make sure that the output has been evaluated and make sure is good to go to the next step in the process, right?

Anthony Alford: If the AI can handle the pager, then I'm definitely behind that. AI is on call and not me.

Srini Penchikala: Yes, that would be a good area that we can delegate to AI agents, right? Because nobody wants to be on call and get a call in the middle of the night. So I think AI agents can probably do some of those things, right? Or they're already doing it. So let's maybe segue into that. So AI agents, we have seen maybe 2025, I think it's called, year of agents, I forgot. I'm losing track of which year is called what in the AI space. So definitely, there are a lot of AI agent discussion going on, the Claude, Anthropic recently announced Claude subagents and also Amazon announced Bedrock, Amazon Bedrock agents. Right? So a lot of stuff is happening here. So Daniel or Anthony, I know you guys probably have spent a little bit more time on this. So what are you seeing in some of the agentic AI space?

Daniel Dominguez: So I think all this big shift that we're seeing on the agentic space, so instead of having the chatbot that we were just to interact, now we're having this AI that can help us to book meetings, to update databases, to launch cloud resources, to do a lot of things. So in that space, for example, Amazon Bedrock Agents are interesting because they let us build production-ready agents, so on top of any foundation model. So without the need to manage the infrastructure. So they can chain tasks, AWS services, and interpret data securely.

So it's basically bringing the agent paradigm into the AWS ecosystem, make it easy for companies to move from experimentation to real world applications. So it's not only AWS, the platform that allows to create a production-ready agents. I know that Google also allowed to create those agents, Azure as well. So there are a lot of platforms that NAN let you also create agents from scratch. So I think where it good to see these agents moving faster because all of these easy and ready platform to deploy it.

Anthony Alford: Yes, I really think that this agents are, as like everything, a double-edged sword, extremely powerful and useful, but also kind of dangerous. The base example is, it's an LLM that can call a tool. And so, some of the tools they can call are things like file system operations. Well, it might try to RM minus RF and this happens to people, but these things are, they're very useful. And once I've started using Amazon Q command line, which by the way Q is, there is no CodeWhisperer anymore, I don't think, it's Q.

They're just calling everything that, but you can chat with it and say write a shell script to do whatever, find all the image files in this folder and enlarge them and it'll do it. It's been very helpful to me because I can never remember shell script syntax, things like that, but we've been talking about AI safety for years now. It's important already, but it's extremely important when the AI can erase your hard drive, access your bank account, all this other stuff. Yes, it can make our lives easy, but it could also make our lives very unpleasant. And again, I don't think anybody has an answer for it.

Srini Penchikala: Yes, good points. Vinod and Savannah, any comments on the AI agents topic?

Savannah Kunovsky: I think the only thing that I have to add is, it's interesting when this type of language kind of meets the consumer market because at the end of the day, your grandma who is using the product that you're creating, doesn't really care to understand what an AI agent is. And so I think that using that type of language is important for the technologists to face what the technology is capable of and try to make sure that we're designing the systems in a way that's secure and trustworthy, and all of that stuff.

But when it comes to how we're talking about the products that we're creating using technologies like that, I think that we really need to focus on how we are delivering value to consumers and how we're creating products that are truly trustworthy and safe, and talking about them in more human terms rather than these technical terms that just feel more confusing and the future is coming, but what does it actually mean? And it makes all of us ask all of these questions. So yes, I guess the takeaway is using human language.

Vinod Goje: And one thing that we're not sure about, I mean, companies like for enterprise adoption, we are seeing AWS, Agentic Core and Databricks, they're developing platforms for enterprises to adopt and build agentic models and to make it seamless and adaptable and easier to manage.

AI’s Role in Human-Computer Interaction (HCI) [36:37]

Srini Penchikala: Yes, thanks, Vinod. So going back to Savannah, you mentioned about the human side of this, right? So maybe we can quickly jump into that topic, the human-computer interaction, HCI. Again, this is something that your teams focus on on a daily basis. Would you like to share your thoughts on what are the use cases that you see AI being used here and what are the different interfaces and how is interaction with software is changing in everything we do every day?

Savannah Kunovsky: Yes. Similar to my comments that I just made about AI agents and consumers at the end of the day, if we talk about something being AI or not, actually there have been studies that if you market your product as using AI, then people trust it less. And so, I really think that what we need to do is focus on how we are designing and delivering value. And at the end of the day, the ways that we interact with software and the way that we interact with computers is really contrived. When I was in school then we were learning how to type and it was like a typing bootcamp of a teacher who was walking around the room and yelling commands at us to learn how to use keyboards. And that's such a contrived interaction paradigm with a computer. And I really do believe that a big utility of emerging technologies are to make technology easier for people to use and easier for people to access.

And if we get lost in the sauce of what the technology can do and all of these kind of buzzwords floating around it, then we end up just kind of creating technology for technology sake rather than creating technology truly for consumers. So I think that AI is interesting or large language models specifically are interesting and the technologies that we're laying on top of them are great because it does just mean that we can interact with this really powerful technology in a way that is more human, which is through natural language. And I think that the interfaces that these allow us to create also can be more human rather than having similarly, when I was first becoming a software engineer, we had to learn how to write a really good Google search. And that's a skill that maybe through these technologies is becoming more or less antiquated.

But right now, in order to use a large language model, you have to know how to prompt it and talk to a computer rather than talking to it actually like a person, which is what we're used to. And I think that's a detraction from adoption from these technologies for people. One last comment is that there was a lot of talk in the design community when Apple released their Liquid Glass new kind of design system and there was a lot of pushback against it because it was kind of like, "What is this?" And it's just sort of evolutionary rather than revolutionary. But I think just in the name, Liquid Glass, what they're trying to show us is that we can think of our technology as something that has an interface that is a lot more fluid. There's a group in the MIT Media Lab that's literally named Fluid Interfaces. And I think that the reason why these Liquid Glass and Fluid Interfaces and just trying to think about rather than these incredibly structured kind of mechanical ways that we interact with the lines and boxes of the internet.

What if we actually just had more interfaces in our homes and on our computers and in our lives that let us move information around a screen where we actually needed it to be or access information about cooking on a cooking surface rather than balancing your laptop on top of your microwave, might just be me. But I think that what technology allows us to do is embed information in the places that we need it to be. And as you're walking down the street, it doesn't make sense that you have to take your phone out and stop and pull over and respond to a message. Why can't you just do that as you continue going about your day? So in general, when we talk about human computer interaction with emerging technologies, what we should be driving towards is taking all of our research and engineering goals and mapping those to true human needs and understanding how these technologies fit into the context of people's lives and designing for that.

Srini Penchikala: Yes, definitely. Thank you, Savannah. Yes, I actually see the way we communicate with other humans is voice mostly, right? So why can't we do the same thing with these AI tools? Right? So I think the voice interaction is going to be probably even more prevalent going forward.

Daniel Dominguez: I think that the OpenAI is pointing to that direction with the IoT device that they're building with GenAI. They're trying to build something AI native device. So instead of a screens and keyboards, it should point to voice first, multimodal, context our interaction. So the idea would be that AI feels like a natural extension of ourselves and not as they mentioned that the current devices that we use nowadays are not built to work with AI.

Model Context Protocol (MCP) [41:20]

Srini Penchikala: Okay, a couple more topics I would like to make sure that we cover before wrapping this up, right? So MCP, Model Context Protocol. I know we've been hearing this about a lot since it came out in November last year. Anthropic released this, right? So it's a kind of interface that can help isolate the clients like the AI agents or AI assistants talking to the backend systems. So what are you guys seeing in the MCP space? I think, is it a good thing or another constraint? And so also there's a Google A2A, Agent-to-Agent Protocol. So how do we compare these solutions and what are the pros and cons?

Anthony Alford: MCP has definitely taken off and been adopted. All of those coding tools that we talk about support MCP. MCP is probably the key technology for making agents happen. It's not a panacea. I think we're already seeing headlines of security problems with MCP servers, but they're certainly useful now. I think their utility might be somewhat limited by the context windows of the models because everything that happens with MCP, the input, the output, goes into that model context. And so, you can only do so much. So I think we're definitely seeing things like Playwright, MCP server for running tests. That one seems to be a big winner. There's one for Figma. So you can tell your coding agent to just go look at their mock-up in Figma and go create that. So I think that it's definitely got promise and potential, and it does have utility. But it remains to be seen what the limitations are that we're going to run into.

Srini Penchikala: How about you, Daniel? Are you seeing any real world examples of, in your work using MCP?

Daniel Dominguez: Yes. I think for me the exciting part of MCP is the interoperability. It's for example, with MCP, you could have, I don't know, Anthropic Claude model using Google Search or you can have OpenAI using your company's data. So everything will be working under the same protocol now. So that’s what makes that scalable and is the idea of having multi-agent systems possible. So different agents from different companies of different source working together, doing their best. So I think that's what the exciting part is going to be.

AI Security and Safety (43:30)

Srini Penchikala: Sounds good. One last topic I would like to discuss before we start the wrap-up comments. So AI security, right? The safety, security, privacy, the responsibility of these AI solutions. So what are we seeing in terms of innovation versus the regulations and then the compliance? Savannah, maybe you can probably chime in here and what are you seeing in terms of how do we make these powerful solutions also safe and secure and responsible?

Savannah Kunovsky: Yes, I think that the thing that I see the most is that tech companies and technologists kind of release technologies and just expect people to be on board with adopting them. And sometimes, that is the case is that consumers will freely kind of receive a new service and just be like, "Oh, yes, this is super useful and I really want it". But a lot of the time what also happens or happens instead is that people will receive a new service and they'll be like, "Hmm, this comes from this company that I don't know if I totally trust or it kind of feels like it's off for me".

And I think that the thing that we need to remember is that we're not necessarily entitled to the trust and the time and the dollars of the people who are receiving the technology that we're making. And that really, we need to design experiences that do feel trustworthy and do feel like they prioritize the well-being of consumers in order to get the type of adoption and create the businesses and create the services that we want to create. Because at the end of the day, it really is about creating new and interesting things for people to make their lives better or easier or to make their work more streamlined or whatever it might be. And so I think that just making sure that we're designing experiences that really do feel like they're trustworthy and genuinely helpful for people, is that foundation.

Srini Penchikala: Any other observations in this we know there, Daniel?

Daniel Dominguez: Yes. I think AI security is becoming a real concern. So the models are becoming more and more powerful, but those models can also generate harmful or unexpected content, no? So I don't know if you recall that in the recent weeks there was this chatbot from Elon Musk, the Grok, that started producing answer and response reference to MechaHitler. So there was that kind of output that shows the risk when the word phrase fails. It's not just embarrassing because the chatbot was referring to that, but it also can be dangerous. The system is a customer facing or something like that. So I think this way alignment and the re-teaming and all these AI security layers are just important for the modern innovation. And the challenge is not going to be just to make the model smarter, but making more secure and safe and trustworthy when they're being deployed at scale.

Vinod Goje: Yes. And as we see this new technology getting integrated into human's life, right? I mean, definitely, that option is going to grow. And at the same time, there are, while we see opportunities, there are also risks and we can think about already we started how to mitigate those risks in terms of prompt injection, jailbreaking and security is going to be really a matter of concern.

Srini Penchikala: Yes, I also saw, Vinod, the OWASP Project, the Open Web Application Security Project, which has been pretty popular in the community. They have recently released or launched AI testing guide. So they're also starting to look into this AI side of the things and see what are the risks and vulnerabilities that need to address, right? So good stuff is happening there. And also the sustainable AI. How do we use AI to not only help with the environmental impact but also doing the computing itself? How can we make it more green, right? So are the good things happening there. Okay. Thank you. We can start wrapping up.

Panliests’ Predictions for Next Year [47:02]

Usually, what we do is, as a last question in the podcast, we ask each panelist to make a prediction based on what they're seeing or what they would like to see happen in next 12 months, next one year. And then we can actually take a look on this discussion a year from now, to recap from last year, we had these few predictions, hype around LLMs is going to sober up. It may have happened a little bit, but I don't think it was, I think the hype is still there and AI in robotics, embodied AI is an expecting. I think we're starting to see that with the robotics. And then Anthony, you mentioned about AI winter coming.

Anthony Alford: I'm sticking with that. It's going to come, I'm going to predict four out of the next three AI winters.

Srini Penchikala: Yes, yes, I think it's coming, that's for sure. And then we talked about AI agents, which definitely has seen a lot of innovation. So let's talk about what do we think for the next year, right? So maybe Anthony, we can start with you. What's your prediction? What would you like to see?

Anthony Alford: Besides AI winter, I feel like the agents and the coding, that's very exciting. I expect that to keep going. I don't expect that to slow down. I do expect, we'll see. And again, this is like AI winter, something bad or embarrassing is going to happen, beyond the MechaHitler. I think something very financially disastrous will probably happen, sadly. I hope it's not to me.

Srini Penchikala: How about you, Savannah? What do you like to see or any prediction that you can make?

Savannah Kunovsky: I think that if Anthony's AI winter predictions come true, then we're going to see the things that are truly useful to people as the things that set the new precedent for the next foundations of the internet.

Srini Penchikala: Makes sense. How about you, Vinod?

Vinod Goje: Yes, I mean, if I had to bet my money on, I think my prediction would be on video RAG. We have seen short form videos like one minute, two minute videos. But I think by the end of this year or maybe starting next year, we'll have long form videos and it'll be truly hard for us to differentiate between a human generated video and AI generated video. And there is lot of potential, I believe, in that space and which is easily achievable. We are generating text now and it doesn't stop us from generating long form videos.

Anthony Alford: I, for one, cannot wait for AI generated TikTok dances.

Srini Penchikala: Or even movies for that matter. Okay. Daniel?

Daniel Dominguez: I think that will go further the AI winter. We'll start talking about AI bubble. So I think we're going to still see a lot of examples, even Wall Street is right now, all the hardest narrative. It's everything regarding AI. So everything that has AI primed or AI label, it's where the one is. But if we look back, I was recently looking at one interview with from Warren Buffet. He said that you should be very cautious because everything feels like the early days, the dot-com, where all these valuations and all this money were for the promises of the business rather than the actual business result. So I think we're going to start seeing that all this money that has been born with all these models and companies and infrastructure, it's going to be very similar to a bubble. So I will start thinking of AI as a bubble. Obviously, not that the technology is not going to work because the technology is here and the technology is going to stay, but more regarding the industry itself.

Srini Penchikala: Yes, thank you. I think for me, yes, I think AI will continue to become more and more a subtle part of our lives, right? Savannah, you mentioned, why should I take my phone out on the street and then respond to a message? Why can't I just talk to the phone without even taking it out of my pocket, right? So those kind of things, I would like to see those more context based and more the behind-the-scenes type of action or integration, right? And also, I heard about the quantum AI, which is kind of in early stages. So maybe the quantum computing is beginning to intersect with AI, so maybe that will be one of the things that we'll see more. They are saying that it will help in the fields like drug discovery, material science and energy, so all important aspects of human life. So we'll see where we are and we'll definitely revisit these predictions next year. So let's wrap it up. Any final comments, Anthony? Any concluding remarks?

Anthony Alford: I maybe came across as a little pessimistic. I'm actually very optimistic. This is a great time to be alive and be in the business. I use AI every day. I love it, and it has made my work easier and better, so I'm actually an AI proponent.

Srini Penchikala: We know that, right? So how about you, Daniel? Any concluding remarks?

Daniel Dominguez: Yes, I don't think the AI is a fad. I mean, the technology is here to stay and obviously, will transform our lives, but everywhere of innovation, I think there is a period of speculation. So right now, everything is labeled as AI. So I think we need to be more cautious about what really the technology is going to be for and all these expectations to be more realistic.

Srini Penchikala: And Vinod, any final comments?

Vinod Goje: Definitely this new technology is going to make us feel more productive and leave us with enough time to enjoy what we do, besides exploring our passion and our hobbies.

Srini Penchikala: Okay. On that note, I want to thank all the panelists for participating in this discussion of the 2025 AI and ML Trends Report podcast and what to look forward to in this space for the remainder of this year and next year. And to our audience, we hope you all enjoyed this podcast. I really enjoyed it. Hope this discussion has offered a good round-up update on the emerging trends and technologies in the AI space. Please visit the infoq.com website and download the trends report with an updated version of the adoption graph, which we'll be publishing in the next few weeks. Other than that, yes, I hope you join us again soon for another podcast episode. I know Anthony and Roland, they do host AI podcasts regularly.

Anthony Alford: Any day now.

Srini Penchikala: Any day now. There you go. And check out the previous podcasts on various topics. We definitely did not have time to cover a lot of other topics, but we do have podcasts on those specific topics on the website. But check them out and thank you for joining and have a great one until next time. Thank you.

Mentioned:

• Hugging Face leaderboard
• LMArena
• OWASP AI testing guide
• Anthony/Roland Podcast