There is a public disagreement among AI researchers that is worth understanding, even if you never read a paper. It is about whether the kind of AI behind today's chat assistants, large language models (LLMs), can keep improving into something that handles the real world, or whether something different is needed.
You do not need to pick a side. But the debate shapes what AI vendors promise about agents, so it helps to know what each camp is claiming, and what evidence they point to, as of November 2026.
Two terms
An LLM is a model trained to predict the next piece of text, over and over, on a vast amount of writing. Most of today's assistants are built on one, then tuned afterward (see our post on RLHF and RLVR). The term "autoregressive" just means it produces output one step at a time, each step conditioned on the ones before.
A world model is a model that learns how an environment behaves, so it can predict what happens next if something changes or an action is taken. DeepMind describes world models as systems that "use an understanding of the world to simulate parts of it," so that agents can predict how an environment will change and how their actions will affect it.
The case from Yann LeCun
Yann LeCun, a Turing Award winner who was Meta's chief AI scientist until he left in late 2025, has been one of the best-known critics of the idea that scaling up language models is enough. In 2022 he set out his alternative in a position paper, "A Path Towards Autonomous Machine Intelligence." In outline, as we read his public statements and papers:
- Language is a narrow slice of experience. Much of what people know about the world, how objects fall or how to move through a room, is learned by watching and acting, not reading.
- Predicting every detail is the wrong target. Generating the next word, or the next pixel, forces a model to guess details that cannot be predicted. He proposes predicting in an abstract space instead.
- That idea has a name: JEPA, the Joint Embedding Predictive Architecture. In I-JEPA (2023), a model predicts the *representations* of hidden parts of an image, not the pixels. In V-JEPA 2 (June 2025), researchers including LeCun pre-trained a model on over 1 million hours of internet video without action labels, then, with under 62 hours of unlabeled robot video, adapted it to plan simple robot picking and placing tasks in two labs without further training for those tasks. These are the authors' results.
- His stated concerns about LLMs include hallucination, and weak grounding in the physical world and in planning.
In 2026 he became executive chair of a company built around this view, AMI Labs (Advanced Machine Intelligence), which TechCrunch reported in January 2026 as developing world models "to build intelligent systems that understand the real world." Its mission statement, as quoted there: "real intelligence does not start in language. It starts in the world." Press reports in March 2026 put its funding at about $1 billion.
The case on the other side
People who disagree do not all say language models are perfect. The strongest forms of their argument are:
- Language models already learn some world structure. A 2022 paper (Li et al., "Emergent World Representations") trained a GPT-style model only to predict legal moves in the board game Othello and found evidence that it formed an internal representation of the board. That is a small, rule-based game, so it shows it can happen, not that it happens at scale for everything.
- Models keep improving, and training methods keep changing. Reinforcement learning on verifiable tasks has produced gains in mathematics and code, as reported by DeepSeek in its R1 paper. One response to the "dead end" claim is that new training methods and tools have addressed some weaknesses once expected to last. (We covered this in the RLHF and RLVR post.)
- World models and language models can be combined. Rather than a replacement, many researchers treat them as complementary: language for reasoning and instruction, a world model for physical prediction. Google DeepMind's Genie 3, announced in August 2025, generates interactive environments from a text prompt, in real time at 720p and 24 frames per second, with consistency for a few minutes. DeepMind lists limits: only a few minutes of interaction, a limited set of agent actions, and imperfect real-world accuracy, and presents the work as a step toward training agents in simulation. A text prompt driving a world model is itself a mix of both approaches.
- Practical results come first. Today's LLM-based systems are already doing useful office work, which is an argument for building on them while the research continues.
Some commentators who look at the dispute neutrally reach a middle position: the strongest version of LeCun's claim, that scale alone will not give you robust physical reasoning and long-horizon planning, is plausible and testable. The strongest version of the opposite claim, that language models will keep absorbing new capabilities, also has evidence behind it. Neither has been settled. We would be cautious of anyone who says it has.
Why it matters for agents
An agent is software that does a task for you: reads your records, takes steps, calls tools. Whether the agent acts in a screen or the real world matters.
- On a screen, with email, invoices, calendars and records, the world is text and numbers. Language models are already used here, and failures show up as wrong drafts you can check.
- In the physical world, with trucks, tools and people, an agent needs to predict what actions do. That is where world models are argued to matter most, and where we see no documented, broad deployments yet. See our post on physical AI and humanoids.
- In between, agents that plan over many steps, such as scheduling a week of jobs, use something like a model of how your business works: how long jobs take, which customers pay late, what depends on what. That model can be built from your records. This is a smaller, more practical "world model" than the research version, and it is the one a small business can actually build.
What this means for your business
1. Do not buy on the debate. Vendors on both sides have commercial reasons for their story. Test any tool on your own work.
2. Treat claims about agents acting in the real world with extra care. Ask for documented deployments and failure rates, not demos.
3. Expect change. The methods behind AI are still moving. Keep your records and your tests independent of any one model type, so you can switch.
4. For screen-based work, the useful question is reliability, not which architecture wins. See our post on reliability.
5. Build your own business model. A clean record of your jobs, prices, durations and customers is the closest practical thing to a world model of your company.
How this shows up at Humonix
Humonix builds on the language models available today and does not claim a view on who is right in this debate. What we do take from it is the idea that an agent is only as good as what it knows about the environment it acts in. For your business, that environment is your records. Know holds them in one place, Hue works from them, and a person approves anything public, priced or paid. Each feature on a module page is labelled Live, Early access or Coming, and the physical capabilities on our humanoid page are Coming. See The Agentic Enterprise.
Sources
- Assran et al., "Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture" (I-JEPA, arXiv:2301.08243, 2023)
- V-JEPA 2: "Self-Supervised Video Models Enable Understanding, Prediction and Planning" (arXiv:2506.09985, June 2025)
- Yann LeCun, "A Path Towards Autonomous Machine Intelligence" (OpenReview, 2022)
- TechCrunch, "Who's behind AMI Labs, Yann LeCun's 'world model' startup" (January 23, 2026)
- Li et al., "Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task" (arXiv:2210.13382; ICLR 2023)
- Google DeepMind, "Genie 3: A new frontier for world models" (August 5, 2025)
- DeepSeek-AI, "DeepSeek-R1" (arXiv:2501.12948; Nature 645, 2025)
This post is general information, not technical or investment advice.