Which Route Leads to the “World Model” Endgame? | A Conversation with 黄碧薇, Founder of Aether AI

Summary

• None of the 3 existing world-model routes is the endgame. Ethereal AI founder 黄碧薇 sees video generation, 3D generation, and 杨立昆’s VGAPA as transitional forms, and is betting on a fourth route: a causal world model that learns causal variables, causal structure, and transition dynamics simultaneously in latent space. “Because it understands the underlying laws, it can generalize like a human.” The company has just raised about $20M from top-tier US-dollar funds and currently has roughly 400 GPUs.
• Her scorecard: VLA may top out at 5/10, WAM at 6.5/10, and the endgame is an action-conditioned world model plus a policy model. VLA operates in a continuous action space, making it nearly impossible for training data to cover every state: “Raise the desktop by 2 centimeters and the task might fail.” WAM benefits from abundant video data, predicting the next frame first and then using an IDM to infer the action; it “must be an intermediate route.”
• LLMs have already hit their ceiling on embodied tasks. Language and coding are “very simple modalities”: causal relationships have already been compressed into the surface of language, tokens are discrete, and the data volume is large enough that learning surface correlations is sufficient. Robots, by contrast, are “actually performing interventions” every time they interact with the environment. Her categorical view is that the ceiling for LLMs in embodied tasks “has already appeared”—the arithmetic behind choosing embodied AI over LLMs is that LLMs are already at 90 points, while embodied AI is still at 10, and she would rather push that 10 to 90.
• Current talk of scaling laws is weak and somewhat nebulous. Scaling laws have to be tied to both data quality and the model: if an LLM gains 20% performance from 1M additional data points, a model that understands causality might need only 200K to achieve the same result. The accompanying data flywheel is to collect 10,000 samples but feed the model only the 100 that contain new information; once sufficiently trained, the model becomes its own simulator, generating corner cases and controllable failure cases that may be impossible to collect on real machines and feeding them back into training.
• A model trained only on lift and pick-and-place completed an entirely new stacking task zero-shot, and the first version is expected to use 7,000-8,000 hours of data. The mix will be about 80% simulation, ego-centric, and video data, with 20% teleoperation. This year’s planned real-robot demo will showcase generalization and reasoning over long-horizon tasks.
• The market agrees that causality is the endgame; all the skepticism is about execution. The open questions are how to extract causal variables from raw data, learn causal structure, and learn dynamics. 杨立昆 has recently discussed causality at a high level, and 黄碧薇 sees some of 李飞飞’s high-level ideas as related to it. 杨立昆—one of AI’s 3 giants and a Turing Award winner—has discussed the approach in depth and cited many of her papers, but “we should be the only team” working across all 3 layers.
• A factory robot being returned in early 2025 was the trigger for her founding the company. “It couldn’t truly replace humans and instead became a burden in the factory because it was missing a very intelligent brain.” Five years from now, she says she will not call any route wrong: VLA will have left behind action-head modeling, WAM will have left behind the intermediate product that turns video generation into a world model, but “the current model paradigm will not be the endgame.”

Deep dive

1. “World model” is an overused label; her test is whether it can simulate how the world works

• Corgi’s opening question goes straight at the definitional confusion: video generation, 3D generation, and VGAPA are all now being called world models, even though they optimize for very different things—rendering, understanding the dynamics of physical systems, or the quality of generated video. 黄碧薇 has one hard criterion: the model must understand the underlying physical laws and causal relationships. It must be able to simulate how the world works, moving from the current state to the next state based on time or on different actions.
• The 3 mainstream routes are video generation, 3D generation, and VGAPA, led by 杨立昆. VGAPA removes the decoder to pixel space entirely and retains only smooth-transition information in latent space. Asked which route she favors, 黄碧薇 points to a fourth: “Maybe it is our causal world model”—while stressing that the approach stands on the shoulders of giants, combining early causal models, correlation-based foundation models, and video generation “at every level.”

2. A causal world model has 3 components: variables, structure, and dynamics

• Its core difference from other approaches is that it learns 3 things simultaneously in latent space. First are causal variables: object shape, quantity, velocity, angular velocity, and friction. Second is the causal structure among those variables: when grasping a cup, how grip point, speed, and angle jointly determine success or failure. Third is transition dynamics: what state the system reaches at the next moment under different actions.
• Only with all 3 components can the model genuinely generalize: it can transfer a robot model from one environment to another and from one task to another because it understands the underlying laws, allowing it to adapt quickly to new environments and tasks “like a human generalizing from one case to another.”

3. Why brute-force scaling works for LLMs but falls short in embodied AI

• Corgi’s follow-up is arguably the best question in the episode: if LLMs can reason rigorously without understanding causality, might causality simply emerge once the data is sufficient? 黄碧薇 starts with why LLMs have succeeded in language and coding. Both are “very simple modalities”: causal relationships have already been summarized on the surface of language—“Because the weather is good today, I’m going out for a walk”—tokens are discrete, and the data volume is large. A paradigm that learns surface correlations is therefore enough. But for harder embodied tasks and scientific discovery in biopharma, new materials, and astronomy, “the current LLM paradigm is clearly not enough.”
• Every time a robot interacts with the environment, it is “actually performing an intervention,” which makes the process inherently causal. Her pancake example is straightforward. A robot trained only on video can learn the surface sequence—pour batter, wait, flip—and generate video that is “not bad, watchable.” But in a real kitchen, the pan may be hotter or colder, the batter thicker or thinner, and the oil quantity different. “When the pan is very hot, waiting 30 seconds means it is already burnt.” The physical world imposes far higher demands on action than video generation.
• Her ceiling call is task-specific but categorical: “For embodied tasks, the ceiling of large language models has already appeared.”

4. The data flywheel: feed the model only 100 of every 10,000 samples

• The causal approach makes the data pipeline selective. Of 10,000 collected samples, perhaps only 100 contain information the model needs, so only those 100 should be used for training. On the output side, once sufficiently trained, the model itself becomes a simulator, producing long-horizon data with corner cases and controllable failure cases. These may be impossible to collect through real-robot teleoperation, and can then be fed back into training, creating a loop in which data is the fuel and the model generates higher-quality data.
• The data mix is divided into 4 categories with a clear allocation: simulation data—including physics simulators and the company’s own model as a simulator—ego-centric data, and video data account for roughly 80%; teleoperation accounts for about 20%. Teleoperation is the “last mile,” grounding the physical laws back into the robot. Most data will be generated in-house, since simulation data “can be collected without limit”; the smaller teleoperation component will be partly self-collected and partly commissioned from vendors.

5. The first model: 8,000 hours, 400 GPUs, and a zero-shot stacking signal

• The first model is expected to use 7,000-8,000 hours of data and several hundred GPUs, versus roughly 400 currently available. The goals are 3-fold: genuinely long-horizon tasks; real reasoning—for example, putting bills, as private items, in a drawer while placing books, which are “more public,” on a desk; and generalization to unfamiliar objects and new tasks. The hedge remains important: “The physical laws behind these tasks must have been covered in the previous training data.”
• The confidence signal already exists. A model trained only on lift and pick-and-place was tested on an entirely new stacking task. “If it were merely memorizing skills learned in the past, it obviously wouldn’t work. But our causal world model can complete stacking extremely well,” because stacking is the sum of the physical laws behind lift and pick-and-place. Shared laws are enough to solve the task; if the model encounters a completely unfamiliar physical law, it still has to explore and learn it, “like a human.”
• Results so far are in simulation, with “hundreds of hours” of data, including this year’s ICML papers, “Learning Task Sufficient Word Models by Synergize…” and “Add Diffuser Latent Aware Adaptive Diffusion for Decision Making.” A real-robot demo is planned for this year to show generalization and reasoning over long-horizon tasks.

6. The academic lineage of causal AI: 2,000 years of philosophy, the PC algorithm, and a 3-way split

• Causality has a 2,000-year philosophical lineage, from Aristotle in the West to the I Ching in China. The modern definition centers on intervention: A causes B if and only if changing A changes the probability of B. The earliest application was the randomized controlled trial, but “experiments are expensive, and sometimes you cannot run one.” In the late ’80s, 3 CMU professors—Clark Glymour, Peter Spirtes, and Richard Scheines—developed the PC algorithm, using statistical methods that go beyond correlation to extract causal structure from observational data. The field advanced rapidly from 89 to 97, went relatively quiet from 97 to 06, and in 06 a Finnish scientist identified additional properties under non-Gaussian conditions that made it possible to determine, using observational data alone, whether A causes B or B causes A.
• The field split into 3 schools, which “didn’t get along very well” in the early years: the CMU school of causal discovery; the Judea Pearl school of causal inference, centered on graphs, at UCLA; and Don Rubin’s Potential Outcome Framework at Harvard. The leading figures are all in their 80s, making it “a bit like the Three Kingdoms—none of them accepts the others.” 黄碧薇 is CMU lineage: she studied under Glymour, Spirtes, and Kun Zhang, while her Max Planck Institute advisor Bernhard Schölkopf also came from the CMU school. She says many of her high-level ideas were also heavily influenced by Judea Pearl.
• Her entry into the field was accidental. In 2013, while studying computational neuroscience as a graduate student in Germany, she heard Schölkopf speak about discovery at a summer school. “I didn’t fully understand it at the time, but somehow I felt this problem was especially fundamental.” The next 12 years focused on causal discovery in imperfect worlds: theoretically proving and algorithmically deriving the underlying causal relationships in the presence of latent variables, bias, missing values, and distribution shifts.

7. The route scorecard: VLA may be a 5, WAM a 6.5, and the endgame is an action-conditioned world model

• VLA’s ceiling is “possibly 5 out of 10.” It adds an action head to a VLM, but “people are still trying to approach the problem from the language angle, when action should actually be viewed through the lens of time series.” The action space is continuous, and training data is “very difficult to cover” across all states: “Raise the desktop by 2 centimeters and the task might fail.” That is the core reason VLA lacks generalization.
• WAM “may reach 6.5,” but “it must be an intermediate route and cannot reach the goal we ultimately want.” It benefits from abundant video data, predicts the next frame first, and then uses an IDM to infer the action between 2 frames. But this “is not a natural state-transition process.” A natural transition should take the state at time t, apply an action, and observe the state at t+1.
• If the causal route can be implemented at every level, it would be a 10. But the company’s strategy is to add causality layer by layer; it “will not aggressively charge straight at that 10.”

8. The market’s biggest question is implementation—not direction—and scaling laws are “hollow”

• 黄碧薇 offers an unusually candid account of the competing view: the market agrees, without question, that causality is the goal to be achieved. All the skepticism is about implementation—how to extract causal variables from raw video, time series, and sensor signals; how to learn causal structure simultaneously; and how to learn a causal system that changes over time. “There are not many people in either the market or academia who truly understand causality.”
• Her scaling-law critique deserves separate treatment: “The scaling law people talk about now is also relatively weak and hollow.” It must be tied to data quality and the model. If an LLM gains 20% performance from 1M additional data points, a model that truly understands causality and the core underlying laws “may need only 200K data points to achieve the same performance.”
• She is encouraged that causality has entered the mainstream conversation. 杨立昆 has recently discussed it at a high level; she sees some of 李飞飞’s high-level ideas as related to causality; and 杨立昆 has affirmed her approach in emails. DeepSeek previously included “causal” in its model names. But most teams either engage with causality only at a high level or implement it narrowly in simple areas, such as predicting the future from the past. “We should be the only team” working across the full stack of variables, structure, and dynamics. 杨立昆—one of AI’s 3 giants and a Turing Award winner—was “very taken with” the causal model, discussed it with her in depth, and “cited many of my papers.”

9. The founding trigger: a factory robot that was sent back

• The turning point came in early 2025, when a friend described factory automation: “The robots had already gone into the factory to work, but after being tested for a while, they were sent back because they couldn’t truly replace humans. Instead, they became a burden in the factory—they were missing a very intelligent brain.” Her reaction was immediate: “Isn’t that exactly what causal AI is best at? This is the right time for me.”
• The release of ChatGPT prompted “deep reflection.” If causal AI was the more rational approach, why had LLMs captured all the attention? “I had not realized one thing before: stacking data is useful.” The conclusion was to combine both approaches: use the foundation of massive data and apply deeper methods to extract deeper information. Causality offers LLMs 2 paths, inside and out. Externally, causal discovery can identify relationships among variables and feed them in through RAG or prompts to reduce hallucinations. Internally, the architecture can be changed so the model genuinely learns causality. But OpenAI and Anthropic have “path dependence” and “have not truly made it onto the road of causality.”
• The arithmetic behind choosing embodied AI over LLMs is simple: LLMs in language and coding have “already reached 90 points,” while embodied AI is “basically at 10.” She would rather push that 10 to 90; LLMs are improving too quickly with each generation, so the incremental gain from moving further in that direction is “very limited.”
• The company has raised about $20M for compute, data, and talent. A small portion came from friends as “friendship sponsorship”—“I obviously can’t take too much money from them.” It positions itself as a frontier lab and is hiring across 3 areas: causal algorithms, video-generation model training, and robotics full stack.

10. Advice to PhDs and a 5-year lookback: no route is wrong, but none is the endgame

• Her view of the PhD has never changed: “Only people with a genuine desire to do research should pursue one.” If the goal is merely a degree, “you don’t need to spend 5 or 6 years on it; you may miss many opportunities to make money.” Idealists may be better suited, but they also need to “identify which things outside are noise”—even if a classmate goes to OpenAI and earns $30M a year. The good news is that “the gap between industry and academia is narrowing… choosing one route does not mean you have to follow it to the bitter end.” She divides research barriers in the AI era into 2 parts: the barrier to truly original research is unchanged, while the barrier to simple research has fallen sharply. “Embrace AI, but do not depend on AI—the core creative idea has to be your own.”
• Looking back 5 years from now at the deafening drumbeat around world models in 2026, she refuses to frame the routes as right or wrong: “You can’t strictly say they are wrong. Each step will leave something behind, but the current model paradigm will not be the endgame.” VLA will leave behind its approach to action-head modeling; WAM will leave behind the intermediate product that turns a video-generation model into a world model. “Their existence still makes sense.”
• The closing question turns unexpectedly philosophical: Is causality objectively real, or is it simply a way humans understand the construction of the world more effectively? It is like asking whether time exists—something “people are not always completely certain about.” Corgi responds: “If causality does not exist, we would collectively fall into a giant existential crisis.” 黄碧薇’s pragmatic conclusion is to assume, in daily life, that time exists, the physical world is real, and causality is real; arguments about it can wait for the tea table.