Why Video Is AI’s Next Great Frontier

Ben Lorica

Yoon Kim on Video Search, Multimodal Reasoning, AI in South Korea, and Open-Weights models.

Subscribe: Apple • Spotify • Overcast • Pocket Casts • YouTube •  AntennaPod • Podcast Addict • Amazon •  RSS.

Yoon Kim, President and Chief Strategy Officer of TwelveLabs, explains why understanding video takes more than bolting a language model onto a stream of frames. He walks through how TwelveLabs’ two foundation models, Marengo and Pegasus, embed and describe video content at scale, and how the company’s new agent, Jockey, orchestrates them for tasks like security investigations, sports film review, and documentary editing. The conversation closes with a look at South Korea’s AI and semiconductor landscape, and a debate over how much enterprise AI workloads will shift to open-weight models.

Subscribe to the Gradient Flow Newsletter

Interview highlights – key sections from the video version:

Jump to transcript

Related content:

Support our work by subscribing to our newsletter📩

Transcript

Below is a polished and edited transcript.

Ben Lorica. All right. So today we have Yoon Kim. He is the president and chief strategy officer of TwelveLabs, which you can find at twelvelabs.io. The taglines on the homepage are “See the Unseen,” “Know the Unknowable.” Your video contains every insight, every event, every decision that mattered. Extracting it has been impossible until now. And with that, Yoon, welcome to the podcast.

Yoon Kim. Thank you, Ben. Thanks for having me. It’s a pleasure.

Ben Lorica. Let’s start with some of the tech components. You have two models: Marengo, which seems to be an embedding model of some sort, and Pegasus, which is a video-to-text model. Before we dive into each of these individually — for our listeners who are less familiar with this space, if someone were to upload a video to one of these commercial foundation models, would they be able to do video search, like with Gemini or similar tools?

Yoon Kim. Search requires a set of data to be searched over. In the case of YouTube, consumers know this is a massive repository of user-uploaded videos. In that case, Google would use video understanding technology, but not at scale, because it’s very expensive. So in most cases, users upload videos along with metadata — including captions and labels — to describe what the video contains. If a user doesn’t provide that and just uploads a video to a Gemini model, they’ll get explanations and a textual understanding of it. But because Gemini, and most multimodal LLM-based models generally, just bolt on the video modality by taking snapshots — perhaps once a second — and using language models to describe those images, what you get is a rendering, or a version, of the video’s description. If you then want to search across multiple videos — maybe hundreds of thousands of hours — it becomes a scale problem, because it’s not feasible for these language models to generate a sufficient level of text to let you search for very specific moments through text-to-text search.

Ben Lorica. So on YouTube, as you mentioned — if I were diligent, I’d put a table of contents in the description, and YouTube itself has automatic transcription. But as you say, that’s not enough if you want fine-grained search. A high-level search like “when did they talk about video understanding models?” might point you to a few clips, but that’s not enough for most applications, right?

Yoon Kim. That’s correct. Video, as you know, contains not only visual information but audio — mostly speech, but other types too. It can involve textual characters displayed as captions and metadata.

Ben Lorica. For example, a lot of listeners probably watch videos of talks that include slides.

Yoon Kim. Yes. For lectures especially, there’s a lot of different modalities of information contained in the video. To give an example: a well-produced video, like the ones on your Data Exchange website, is carefully edited, with captions, labels, and chapters — very well maintained. Now say a video doesn’t include audio or speech — then you don’t have a transcript available. Or say you have videos that aren’t meant to be consumed by humans but by machines — drones, CCTVs, or other machines that will proliferate over the next ten years. For that kind of video, it’s really difficult for language models to describe or understand the content, or to pinpoint exactly where something happens.

Going back to consumer video — even if I wanted to find, among all your videos, a moment where you asked about video search, or where you and I were having a conversation and laughed, or got into a heated discussion, how do you find that? There’s super-linguistic information contained in video that captions or LLMs alone may not capture. You need the textual context, but you also need the “plus alpha.” That super-linguistic information — search, retrieval, understanding — is what TwelveLabs excels at.

Ben Lorica. And then, in a presentation-style video — this isn’t unusual — someone gives a talk and says, “here’s how the product works,” and plays a video.

Yoon Kim. Yeah.

Ben Lorica. And the video has no sound, other than maybe a robot moving around — it’s all visual. If you’re relying on a transcript, you’re lost.

Yoon Kim. That’s correct.

Ben Lorica. And even if you have a transcript, what about sarcasm, irony, subtle things like that? I guess the other issue is that language models excel with text, so even if a presentation becomes multilingual — people mixing Spanish and English in the same sentence — language models handle that fine. It’s the mixture of modalities that’s tricky, and, as you mentioned, the scale of things. I’m more familiar with security applications, for example: someone broke in somewhere, here’s 2,000 hours of video, find that moment.

Yoon Kim. Yes. And also find near-crimes, where the same person — or different people — may have done similar things in the past that you didn’t know about. Can you find those too? These are moments and information that can be searched across thousands of hours of data, which isn’t uncommon. Then you get to knowledge: why did that happen, or how did it happen? Can you describe what led to it? Insights are nothing but knowledge, information, and patterns you may not have known existed, that the AI surfaces — things that almost happened but didn’t, or things it notices that you don’t, like vulnerabilities in your security that you may have missed. That’s where “seeing the unseen” and “knowing the unknowable” come in, applied to archival footage.

Ben Lorica. Going back to your foundation models — when TwelveLabs started, was the goal, at least initially, to build on other people’s models? Were there existing models you could use, so you’d just build an application on top? And did you end up finding that you actually had to build your own model?

Yoon Kim. Yes. When the founders started the company, the most advanced model at the time was GPT-3 — not even 3.5.

Ben Lorica. That wasn’t even multimodal, was it?

Yoon Kim. No, it was just a language model — it didn’t have the capabilities of 3.5 or ChatGPT. But the founders were early in detecting that, given the pace of development in data, compute, and people, language models would eventually exceed human capabilities in many respects. Some of the founders had served in a Korean Cyber Command military unit, where they worked with all sorts of data — text and language, but also other modalities, including video. They realized video is extremely rich and complex, and there was no foundation model that could analyze, understand, and process it — only point-solution computer vision tools. So they decided to build a video understanding model from the ground up: taking video natively, turning it into vector embeddings, and from there, since vector embeddings are far more computable, building other capabilities on top. We started with vector embeddings and search foundation models back in 2021.

Ben Lorica. For listeners who don’t follow this space closely — say I have an hour-long video and I’m going to embed it. Is that done frame by frame?

Yoon Kim. Not exactly frame by frame — if it were, there’d be a lot of redundant information.

Ben Lorica. For instance —

Yoon Kim. — security footage, which you mentioned, wouldn’t have a lot of change. So our model understands when to embed dense, new, useful information.

Ben Lorica. Something new. If you’re filming a security camera and nothing is happening, you’re not going to embed frame by frame — that would be a waste.

Yoon Kim. That’s a waste. Our models are efficient at understanding what to embed, and when, and at what scale. The opposite end of the spectrum is something like a football game — high value, high definition, high quality — where, in an instant, something happens. That’s very temporally dense video information. So the way you sample and tokenize video has to differ; it can’t be linear or uniform.

Ben Lorica. So it’s application-specific — a soccer or football game has a lot more action than a security camera feed.

Yoon Kim. That’s correct.

Ben Lorica. Is that a knob you turn during the embedding process, or does the model automatically understand what to embed?

Yoon Kim. It’s the latter — it’s difficult to tune a foundation model to specific things without losing its foundational quality. We have a spatial-temporal signal understander, so to speak, that looks at incoming frames and determines exactly what to tokenize — the pixels, the rates, the blocks of information across time and space that need to be encoded.

Ben Lorica. Suppose I engage TwelveLabs in a domain you’ve never really worked in before. Is there any post-training or fine-tuning needed to onboard a customer like that?

Yoon Kim. It depends on the performance you want. For something simple, like search — finding moments — the foundation model may work well as is. We’ve been pleasantly surprised at how well it performs on data we haven’t trained on. But if you want precision for mission-critical applications, post-training is necessary for entirely new domains.

Ben Lorica. Once you have the embeddings — as I mentioned before we started recording, I advised a company in this space that has since exited, and they did something similar. They had an embedding model, but at some point added a storage system involving a graph, so they could do clustering and things like that. Is that something you do as well?

Yoon Kim. Yes, we do a lot to make sure search is efficient at very large scale. Search isn’t the only thing we do, of course. Right now we have two models. One does search well, using embeddings — that’s the vector-search-based embedding model, Marengo. The other is Pegasus, which works more like Gemini — it produces text from video. The reasoning is that while speech-to-text, or even image-to-text, is feasible — because the dimensionality of information in speech, an audio rendering of language, or in images, is easier to describe in words — going from video to text is harder, since video has multiple, very rich layers of spatial-temporal information. So dense captioning generally isn’t as feasible or cost-effective. Pegasus is designed around what the customer actually needs video-to-text for: what kind of video are they inputting? Is it YouTube-style video, or captionless, unlabeled video data — which, for us, is the more useful case?

Ben Lorica. So is the ultimate goal of video-to-text just search?

Yoon Kim. Not necessarily. Pegasus converts raw video data into a set of information and knowledge the user wants, whether for organizing information or for question answering. One example is broadcast and media companies that have decades of video archives and don’t know what’s in them.

Ben Lorica. Let me find the snippet where I interviewed this person about this topic.

Yoon Kim. Exactly.

Ben Lorica. And then you point me to the exact timestamp.

Yoon Kim. The exact timestamp, segmented, with a description that matches what you’re searching for or want to know. It could be a question-and-answer format, or structured metadata — a markup-language-based file with events, people, and other customer-specific schema. I feed that into the model and say: watch the video, and populate this metadata file as if a person were writing it. That’s something we do extremely well.

Ben Lorica. If you upload a video and have YouTube transcribe it, you can start querying it, because it has a transcript — but the timestamps are based on speaker transitions. So if I start speaking at minute one and I’m long-winded, talking for ten minutes, and I mention something somewhere in there, you’d still have to manually find that moment.

Yoon Kim. That’s right.

Ben Lorica. Whereas what you’re describing points directly to that segment.

Yoon Kim. Correct, because we watch and listen to the video content.

Ben Lorica. Interesting. So watching the video is computationally intensive.

Yoon Kim. Yes — that’s why you can’t just use LLMs for it; it’s very expensive, and LLMs, as you know, have limited context, so you can’t rely on context alone in a model like that. Instead, we build up a store of knowledge over time, even across videos. As the model watches your videos, it forms an inter-video archival knowledge of who you are, how you speak, the format of your podcast, and the most interesting bits of information you tend to draw out. As the video library grows larger and more diverse, our models can provide much richer information — to the point where a lot of people now use Marengo and Pegasus for search, and then, after finding what they need, want to analyze it: summarize it, generate highlights, chapterize it — interchanging these models in ways that’s essentially what an agent should do automatically.

So we recently launched, in research preview, something called Jockey: an orchestration, agentic platform that uses our foundation models to handle more complex tasks. For instance, if I wanted to do a thorough analysis across all the videos on your website — or any video archive — and get insights and patterns to help with something, like making short reels, or learning more about you as a person. I actually did that: I wanted to learn about you not from internet text, but from all your videos — how you ask questions, what information and knowledge you try to extract from guests. It gave me a much richer sense of you than transcripts or language-based data alone could.

Ben Lorica. Let’s take a step back. Marengo is the foundation model that embeds video and makes it searchable for downstream applications — because with embeddings, as with text, people use RAG, so you can do RAG on video too. And Pegasus is the video-to-text model, but the text itself is so rich in metadata that it’s like a supercharged transcript.

Yoon Kim. Exactly. It’s a markup language — a hyper-markup language that includes not just the transcription —

Ben Lorica. — what was said —

Yoon Kim. — but also what happened: who, where, when. All of that is embedded into the markup-language file the customer wants.

Ben Lorica. So, going back to the example I gave earlier — a video presentation where I press play to show a product demo with no sound, just visuals of how the product works — the markup language would describe that too?

Yoon Kim. Exactly.

Ben Lorica. What’s happening within that video?

Yoon Kim. Right — and it can also extract snippets that are especially interesting to look at, since seeing is different from reading. It’s all segmented.

Ben Lorica. Are both foundation models multilingual?

Yoon Kim. Yes. We support a lot of languages — I don’t know the exact number, but all the major ones — covering audio as well as any on-screen text, transcriptions, and textual metadata in the video.

Ben Lorica. Neither of these foundation models is open weight, correct?

Yoon Kim. No, they’re proprietary.

Ben Lorica. So they’re accessible through an API.

Yoon Kim. That’s correct.

Ben Lorica. Can you give us a sense of how big these models are, if you’re comfortable sharing?

Yoon Kim. Without going into detail — when we first tackled this problem, we recognized it would be impossible to process video using pure transformers to understand it, because, as research has shown, it’s very difficult to predict frame by frame. If you can’t predict and tokenize frame by frame, you need clever techniques to make the whole process efficient. That’s what we did, and that’s the key to scale — being able to index millions of hours of video for a customer’s search needs. Pegasus, the video-to-language model, and Marengo, the embedding model, are both anywhere from about 100 to 1,000 times the size of frontier large language models.

Ben Lorica. The frontier model labs are chasing scaling laws. Are you chasing scaling laws too?

Yoon Kim. We chase scaling laws for the language part of what we do — that applies. But we don’t want to reinvent the wheel. Our argument, at a basic level, is that video and language are different kinds of data that need to be treated differently. So we don’t really apply scaling laws across the board.

Ben Lorica. Just to clarify — your strength is video, but there are existing foundation models that excel at language, some of them open weight. Do you leverage those LLMs?

Yoon Kim. Yes. One way we do that is through Jockey. In plain language, Jockey is the eyes — it’s vision intelligence.

Ben Lorica. Basically a chatbot you can use?

Yoon Kim. Jockey is a visual agent that looks at, sees, and watches video, but it doesn’t pretend to be an LLM. This product, combined with something like Claude — we have a Claude Code interface for accessing Jockey. In private beta, people can subscribe to our agent and use the Claude interface to converse with Claude just as they normally would.

Ben Lorica. But it’s pointed at video workflows — their video.

Yoon Kim. Exactly. Claude calls Jockey, and Claude and Jockey collaborate on the language and video sides to handle the user’s video workflows.

Ben Lorica. Say I have 100,000 hours of video and I’m using your products. Where does the data processing happen — in your cloud?

Yoon Kim. We imagine it’ll mostly be enterprises and large companies with hundreds of thousands of hours of video, and they’ll already be storing it somewhere.

Ben Lorica. They’re storing it somewhere, right?

Yoon Kim. Right. There are two ways to do it. One is through our first-party SaaS platform, where the data is sent to us and processed within our infrastructure.

Ben Lorica. They’d have to FedEx it in on hard drives?

Yoon Kim. That’s been a real challenge. We’ve learned that video data ingestion — and moving data from one place to another, per video — is a huge challenge.

Ben Lorica. Yeah.

Yoon Kim. So sometimes we use Amazon’s infrastructure — specifically Amazon Bedrock, where Bedrock models, including ours, can serve customers within their own VPC environment on Amazon. That way the data doesn’t leave.

Ben Lorica. So imagine I have 100,000 hours of video on-prem already, and I use your technology to run it through your two foundation models. I’d need enough storage for the embeddings and the artifacts from your models, right?

Yoon Kim. Right, but that’s a small fraction of the size of the original video. The real challenge for most customers is digitizing these videos, organizing them, and figuring out how to store them — including whether to put them in the cloud. Most of our customers already have their video assets in the cloud, and depending on their data needs, we use either our first-party platform or a third party like Amazon Bedrock. Some customers are very sensitive about data, which is a real challenge, and we respect their data governance requirements. In some cases, neither option works — the data can’t be sent to the cloud at all, for security or privacy reasons, sometimes requiring an air-gapped environment. That tends to be the case with government customers. In those more extreme cases, we provide on-prem solutions, solving the problem at the customer’s site, even on an air-gapped basis.

Ben Lorica. We talked about search — finding a gesture, or a picture, or similar things. I’d imagine another classic use case is analytics: how many times did this happen, compared to the same day and hour last year — a year-over-year comparison. What other things do people typically do, besides those two?

Yoon Kim. I see more people becoming sophisticated — not just with quantitative analysis, like counting or tabulating data, but with qualitative analysis too. That’s where our models also do well: finding patterns, and predicting or reasoning about why something might happen again. We call this video reasoning — a bit different from semantic reasoning, because it’s based on episodic memories of what the model saw, heard, and learned through language and video, drawing on multiple layers of knowledge.

Ben Lorica. Can you make that concrete with an example?

Yoon Kim. Sure. This might not be the most relevant example, but say I want to analyze Ben’s speaking style — for reviewing your videos, but also for your own benefit, to see how you speak, how you enunciate, and what patterns show up in your gestures around certain topics, so you can improve as a speaker. I’d feed all my YouTube presentation videos into the model and say: watch these, identify the portions where I’m speaking — interview format, presentation format — and give me your own analysis of the strengths and weaknesses of my presentation style, along with opportunities to improve, and anything that might be annoying to the audience. That’s something it can do across multiple videos.

Ben Lorica. If I applied this kind of technology to sports analytics — Moneyball-style, for basketball, soccer, or football. Say I’m a quarterback: look at every snap I’ve taken, and start asking questions. Is that the kind of thing you mean?

Yoon Kim. Exactly. If you extend it to sports, we have a demo video of where we’re heading. The vision is a coach in a film room full of raw footage —

Ben Lorica. Very raw, and from different angles, too.

Yoon Kim. Absolutely — different angles, multiple cameras on the same play at different points in the game, different perspectives from different players. Same with golf — lots of perspectives. When an NFL coach goes into the film room, they spend hours analyzing what AI can probably do better than most coaches.

Ben Lorica. Actually, the coaches themselves don’t usually do it — they have video coordinators, often more than one.

Yoon Kim. Right, they have staff — scores of people doing this analytic work and bringing the footage to the coaches. Even after that curated footage is delivered, the coach still has to work with it, asking questions: can you put these together into a set of clips I can review, or tell me what happened there, or compare what this player did in Game Eight versus Game One — what’s improved and what hasn’t. These are things coaches want, but right now it’s not easy for them to have a copilot to help. They’re experts in football but need help navigating video, so it becomes as easy to search and comprehend as text.

Ben Lorica. And the vision is that the coach can just interact with it directly, like a chatbot?

Yoon Kim. Exactly.

Ben Lorica. So essentially you’re eliminating the need for all the video coordinators.

Yoon Kim. I think video coordinators on each team can focus more on football-related work, rather than the laborious parts. There’s a lot of footage, and, like security footage, most of it is uninteresting. We help these coordinators do their jobs better — not just the coaches. Extend that to security, or pop culture, or ad marketing, where teams want to review the ads they’ve produced alongside what the rest of the world produces — getting both a snapshot and a thorough analysis, drawing on inter-video knowledge. It can analyze ads, provide analytics, and even offer recommendations on what makes a good hook and how ads could improve. That goes beyond analytics — it’s qualitative in nature.

Ben Lorica. Would this work for a case like this: I’m a documentary filmmaker, I’ve interviewed 100 people, generated 2,000 hours of footage that I need to distill into 90 minutes — the classic problem of what ends up on the cutting room floor. Would this help me understand everything I have and compose the story better?

Yoon Kim. Yes — you can compose the story better, from different angles, perspectives, and narratives, generating multiple versions for you to review. We’re good at turning very large amounts of raw footage into short, interesting, valuable pieces of video. As a matter of fact — not for this exact purpose, but researchers with the Jane Goodall Institute, who have decades of video of Jane interacting with primates, mostly uncaptioned, used our technology not only to generate highlights but also to do deep research, and may learn new things from it.

Ben Lorica. It may actually help advance science.

Yoon Kim. We believe so.

Ben Lorica. So you’re only as good as the video itself. A classic example of poor video, unfortunately, is home security cameras — they look fine in principle but can get blurry. Assuming the raw video quality is limited, I suppose you could apply advanced video-enhancement technology, but that’s added compute. Are there use cases where you need to improve the raw video itself?

Yoon Kim. Great question. You’re right — the source’s resolution, precision, and clarity matter. But if you have multiple angles of the same scene, you can use correlation across those streams to improve precision, detection, and recognition. That’s another capability we have that other models don’t — processing across multiple simultaneous video streams.

Ben Lorica. And that’s something your technology can do in a much more automated way?

Yoon Kim. Yes, that’s correct.

Ben Lorica. All right — I’m going to do a hard pivot and spend the last few minutes picking your brain on South Korea. It’s a powerhouse in technology, and right now certain segments are booming, especially memory chips — I’ve read stories about people at memory chip manufacturers getting huge bonuses. Memory chips are one of those feast-or-famine sectors, and right now, with the rise of AI, it’s going well. But South Korea is also strong in other areas, like robotics. Can you give us a high-level overview of what’s happening in South Korea in AI and tech generally?

Yoon Kim. It’s been amazing to watch the journey and the level of interest in South Korea, especially since ChatGPT came out — innovation after innovation. Even before that, South Korea was always strong in semiconductors, especially memory chips, with companies like Samsung Electronics and SK Hynix. That hardware advantage came at exactly the right time. When I served as CTO at SK Telecom, which owned SK Hynix at the time, we foresaw that compute, AI, and machine learning services would drive huge demand for chips. What we didn’t know — and probably what SK’s chairman didn’t know either — was how fast that demand would grow. Infrastructure-wise, SK Telecom, Hynix, and Samsung have a very strong position in the ecosystem. It’s also interesting that South Korea is home to remarkably innovative, creative mobile services — search, social apps, entertainment, content, gaming — at the top of the stack. So think of it like a sandwich: chips at the bottom, services at the top. Those are areas where South Korea is extremely strong. It’s no surprise that Koreans, both consumers and enterprises, are avid users of foundation models and services like Anthropic’s and OpenAI’s.

Ben Lorica. Is there much discussion in the country around AI sovereignty? The foundation models are mostly in the U.S., and they could cut you off — like what the Trump administration did with Fable, for example — and now China is starting to talk about export controls on open-weight models too. Is that a concern?

Yoon Kim. Yes, those concerns were already present a few years back. The South Korean government has been very active in promoting and supporting sovereign models — foundation models of Korean origin — to prepare for exactly that scenario. I think the country has since evolved from focusing solely on sovereign models to using whatever models are available, so Korea can be a genuinely interesting player in the AI ecosystem by combining models and services. I think models and services both need a lot more innovation right now. It’s a bit funny to me that, after all these years, the most widely used AI-driven application is still a chatbot. I don’t love that term, because ChatGPT and Claude can do far more than a typical chatbot. But I think a lot of imagination and exploration is still needed to create things that haven’t existed before. Our video models are one example — can AI help with video workflows the way it helps with language workflows? What applications will developers build to create value over time? That’s a critical question for the industry right now, and I think South Korea has contributions to make there too, beyond just chips.

Ben Lorica. And they have homegrown champions in search, and in social media, with their own platforms. So there are real distribution channels, and as foundation models become more commoditized, you can innovate higher up the stack.

Yoon Kim. That’s right. Naver is Korea’s native search engine, and Naver Ventures — its U.S. venture arm — was one of the investors in our latest round, our Series B. We wanted those ties because, like me — I’m a Korean American, born in Korea, but my career accelerated in the U.S., here in Silicon Valley — TwelveLabs was also founded by Korean founders, and we still have teams building models in Seoul, but our headquarters is here. We’re accelerated here in Silicon Valley, thanks to the capital, the people, and the business ecosystem this area enjoys.

Ben Lorica. And a much larger population of customers.

Yoon Kim. Exactly — a broader scope of customers than Korea alone offers. We’re very happy with our footprint, and video is such a vast domain that we’re still just scratching the surface of helping our customers across various domains.

Ben Lorica. I’m going to throw out a prediction I’m starting to come around to, and see how you react. I think we’re increasingly going to see open-weight models handle much of the workload in enterprises — you’ll still have the major clouds and frontier models, but maybe 20% of compute goes to the more complex work. A lot of enterprise use cases are routine; you don’t need a giant model that speaks 500 languages to do one narrow thing. So I think there’s going to be increasing pressure on Anthropic and OpenAI from downward pricing pressure, given that open-weight models are already good enough for most things.

Yoon Kim. Indeed — especially the ones from China, which are extremely strong.

Ben Lorica. And increasingly there will be strong Western ones too — Nemotron from Nvidia, and Thinking Machines released an open-weight model, not meant to compete directly with Claude or others, but as a base layer that you post-train on top of.

Yoon Kim. That’s right. As enterprise and consumer customers become more sophisticated, they understand their workflows better and understand exactly what value these models bring. So it’s natural that they’ll look to cut costs and be efficient while still running their AI-driven business. Right now, they might reasonably say that cutting-edge frontier models, especially for language, are too expensive to use for everything. So I think there will be a delineation of tasks: frontier models won’t be used for some, while open-source models that are just as good will handle them. That means frontier labs — OpenAI, Anthropic, and others — need to keep proving they offer value those other models can’t. There will still be use cases where frontier models are worth the premium, but that’s the challenge ahead. I agree with your assessment.

Ben Lorica. And with that, thank you, Yoon.

Yoon Kim. Thank you very much.