# This statement is missing a premise: whether your data pipeline is connected

This statement is missing a premise: whether your data pipeline is connected

Full article: https://haiguangboy.com/posts/liangwenfeng-world-model

## Core Approach

This statement looks like a bombshell. Jike Technology has written the world model into its corporate positioning, Xinghai Tu has Fast-WAM, Xingdong Era has iterated to the fifth generation, and Poqiao has abandoned VLA in favor of video-action world models; Li Feifei's World Labs and LeCun's AMI Labs are built entirely on this.

## Key Results

But looking at his full ranking: CoT (last year) → Agent (this year) → **continuous learning** → gradual singularity → embodied intelligence. He also explicitly stated: **"The endpoint of intelligence may all be embodied. Because for a normal person, what he needs is not a computer, but human labor."**

So he is not denying embodiment—he is ranking it: "not on the main line" means not on the main path to the next step (continuous learning).

**What that statement truly lacks is a premise: whether your data pipeline is connected.**

The LLM pipeline is connected—internet text is nearly infinite, and next-token prediction directly converts data into capability without any intermediate conversion device. Under this premise, the world model is indeed a detour.

The robot pipeline is not connected—action labels are scarce and expensive (Xinghai Tu calculated: real robot data costs 200-250 yuan/hour), while videos and human demonstrations are abundant but **lack action labels**. The world model here does the work of a converter.

There is quantitative evidence: LAWM's experiments show that using **unlabeled-action** videos for world model pretraining **outperforms** supervised pretraining with real action labels—action labels bind the model to a specific embodiment, while predicting the next frame learns transferable dynamics.

**So the world model does not change the ceiling of intelligence, but the cost of reaching it.**

Can this reframing hold? Compare it against every judgment in the library betting on world models: all 9 strong correlations are supportive or extensions, **zero contradictions**. The conflict disappears—those companies are building data efficiency devices, not "mechanisms to raise the intelligence ceiling."

A harder supporting evidence comes from Xinghai Tu's Fast-WAM: the world model's capability mainly comes from **representations learned during training**, not from imagining the future at inference—cutting out video prediction entirely at inference leaves performance unchanged and speeds things up 4-5 times. On this point, Liang Wenfeng and Xinghai Tu are actually aligned.

## Related Route Comparison

- Opposite route · [wx_界面新闻_20260605](https://haiguangboy.com/posts/wx_界面新闻_20260605) `wx_界面新闻_20260605_2026_06`: ★Liang Wenfeng's AGI ladder: CoT→Agent→continuous learning→gradual singularity→embodied intelligence, explicitly stating "the endpoint of intelligence may all be embodied" contradicts route bet: build the brain, not the body—betting on the foundation model company identity, rejecting the motor capability track
- Same route · [LAWM: Why action labels become a burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why action labels become a burden
- Same route · [latepost_xuhuazhe_202603](https://haiguangboy.com/posts/latepost_xuhuazhe_202603) `latepost_xuhuazhe_202603_2026_03`: ★Liang Wenfeng's AGI ladder: CoT→Agent→continuous learning→gradual singularity→embodied intelligence, explicitly stating "the endpoint of intelligence may all be embodied" validates entrepreneurial three signals: RL single-point tasks solved + scaling law ceiling high + Agent framework reusable
- Same route · [π0.5 trembles in place when failing to grab a spoon, but Orca goes further with physics intuition learned from watching videos](https://haiguangboy.com/posts/orca) `orca_2026_07`: The key to the world model is readable states
- Same route · [Qianxun Intelligence Han Fengtao: Embodied intelligence enters the "speed competition" stage](https://haiguangboy.com/posts/qianxun-general-brain-data-loop) `wx_中国财富_20260719_2026_07`: ★Continuous learning is the common gap between LLM and embodiment, but embodied companies are giving general capability timelines while continuous learning remains unsolved validates ★time point prediction: embodied intelligence's 'ChatGPT moment' is between end of 2027 and early 2028—highly convergent with predictions from two other companies in the library
- Same route · [Sudu Technology's WAIC debut: Reality is the endpoint for robots, not the boundary of training](https://haiguangboy.com/posts/sudu-third-position) `wx_晚点latepost_20260718_2026_07`: ★★Core judgment: "The world model has nothing to do with the intelligence ceiling" implies the premise "data pipeline is connected"—LLM is connected so it's a detour, robots are not so it's a bridge validates precise formulation of simulation-first training route: pretraining with almost no real robot data, but post-deployment failure retries plus a small amount of real robot RL later

## Boundaries

Closed-door meeting, content collected from multiple media sources, the original text states "some words and sentences may differ slightly from the original," no recording, no official confirmation. Whether the most critical statement carried qualifiers (such as "under our resource constraints") cannot be determined.

## Author's Judgment (not paper content, cross-paper comprehensive view)

💡The real disagreement lies elsewhere

The only strong contradiction edge in the comparison is not on the world model—it's Liang Wenfeng's "the endpoint of intelligence is embodiment" colliding with Zidong Robot's bet on "building the brain, not the body." **The real question is not whether to have a world model, but whether you can only build the brain.**

There is also one not resolved: LeCun believes the entire LLM path is wrong, and the JEPA-style world model **is** the path to intelligence. That is a real disagreement that definitional clarification cannot dissolve—the verdict may come even later than 2028.