# The Dodo Digest: Is the End of LLMs Closer Than We Think? > LLMs are great at text but struggle with context, structure, and real-world intent. Yann LeCun just raised $1B to build world models, and we've been thinking about this with Sentra. - **Author**: Rishabh Goel - **Published**: 2026-03-21 - **Category**: Newsletter - **URL**: https://dodopayments.com/blogs/newsletter-mar22 --- **TL;DR:** - I was building a deck for AI Boomi when something felt off - the content was correct, but it didn't feel right. That's when it clicked: LLMs are great at text, but struggle when it comes to combining context, structure, and real-world intent. - This week, Yann LeCun raised $1B to build "world models" - AI that understands reality, not just language. A strong signal that scaling LLMs might not be enough. - As AI starts powering real workflows, context and reliability become critical. It's something we've been thinking about with Sentra - building systems that don't just respond, but actually operate with context. **Hello everyone,** I was prepping a deck for an upcoming event called **AI Boomi** the other day. I usually write my slides in markdown first, just to get the structure and ideas right, and then I asked Claude to convert it into a proper PPT. At first, it did what you'd expect. Clean text, decent flow, everything technically correct. But when I opened the slides... it just didn't feel right. The layout was awkward. The hierarchy was off. It looked like something that made sense to read, but not something you'd actually want to present. And that's when it hit me. LLMs are incredibly good at a few things - language, images, even video now. But each of these works well in isolation. Text LLMs do text really well. Tools like Midjourney generate great images. Video models like Seedance create impressive visuals. Individually, they're powerful. But step slightly outside that - where structure, design, and intent need to come together - and things start to feel off. They're great at generating text. But when it comes to turning that into something that actually works in the real world... things start to fall apart. They generate. They don't experience. And that gap becomes very obvious, very quickly. That thought stuck with me. And it brought me back to something I had read earlier that week. ## What Yann LeCun Sees Yann LeCun, one of the people who helped build modern AI, just raised $1.03 billion to work on something very different from what most of the industry is focused on right now. Not better chatbots. Not larger language models. But what he calls **"world models."** The idea is simple, but quite radical. Instead of building systems that generate text better and better, he's focused on building systems that actually understand how the world works - systems that can reason, predict outcomes, and learn from interaction, not just language. That alone is interesting. But the bigger implication is hard to ignore. The person who helped shape modern AI is now betting that the current direction - scaling LLMs - might not be enough. And when you think about it, that frustration I had earlier starts to make more sense. LLMs are incredibly good at predicting the next word. That's what they're designed for. But intelligence, at least the kind we expect from something we call "smart," goes beyond that. It's not just about getting the words right. It's about understanding _structure_. Understanding _intent_. Understanding how something should _feel_ in the real world. That's where things start to break. These systems don't really understand what they're saying. They don't build a mental model of the world. They don't experience anything - they simulate. Which is why they can feel brilliant in one moment and completely disconnected in the next. You're not interacting with something that understands you. You're interacting with something that's extremely good at pattern matching. ## Why This Matters More Than It Seems This might feel like a small limitation today, but it becomes a much bigger problem when you look at where things are going. We're moving toward a world where AI doesn't just answer questions. It acts. It triggers workflows, makes decisions, and handles real operations. It becomes part of systems people depend on. And in that world, context isn't a nice-to-have. It's everything. Because the moment an AI system forgets something important, the entire flow can break. The action fails. The output becomes unreliable. And most importantly, the user loses trust. And once trust is gone, it's very hard to get back. ## How We're Thinking About This This is something we've actually been thinking about for a while. Long before all of this started becoming a bigger conversation, we had already started building [**Sentra**](https://dodopayments.com/sentra) - our internal AI layer that sits on top of payments and workflows. The idea wasn't to build another chatbot. It was to build something that could **work within systems**. Something that reacts to real-time events, carries context across different parts of the flow, and triggers actions based on what's actually happening. In a way, it's less like a chatbot, and more like something that can observe, understand, and act. And to make that actually work end-to-end, we built an [**MCP server**](https://docs.dodopayments.com/developer-resources/mcp-server#mcp-server) that gives AI a structured way to interact with real payment systems. So instead of just generating responses, it can fetch context, trigger actions, and operate within the system itself. That's what makes the whole setup actually useful - not just impressive. Reading LeCun's perspective didn't introduce a new idea for us - it just reinforced something we were already seeing. We're still early in figuring this out, but the direction feels clear. We don't just want AI that responds well. We want AI that can actually **hold context and operate reliably inside real workflows.** ## Alongside This, We Shipped a Few Things While exploring this direction, we also rolled out a few updates in [**v1.87.0**](https://docs.dodopayments.com/changelog/v1.87.0). We made improvements to core payment flows and overall reliability, handled a number of edge cases across transactions more gracefully, and introduced small upgrades that make the developer experience a bit smoother. Nothing headline-worthy. Just the kind of changes that quietly make systems feel more solid and predictable. ## One Last Thing The more I think about all of this, the more it feels like we're at an interesting point. We've built AI that can talk. Now we're trying to build AI that can understand. And those are very different problems. There's a line I keep coming back to: _AI can cook. But it can't taste._ It can generate answers. It can simulate understanding. It can sound convincing. But does it actually _know_ anything? I'm not sure. And I don't think we've fully answered that yet. What do you think? Would love to know your thoughts. Also, join our loving [Discord community](https://discord.gg/dodo-payments-1305511580854779984)! Best, Rishabh Goel Co-Founder, **Dodo Payments** --- - [More Newsletter articles](https://dodopayments.com/blogs/category/newsletter) - [All articles](https://dodopayments.com/blogs)