# Xu Huazhe's Organizational Bet

Xu Huazhe's Organizational Bet

Full article: https://haiguangboy.com/posts/xuhuazhe-poke

## Core Approach

↳ What's welded shut isn't just the division of labor, but three "don'ts"
① No robotics-style single-task optimization—traditional robotics can pull off backflips, but that's a dead end; tasks can never be exhaustively enumerated
② No autonomous-driving-style small-scenario closed loops—perfecting one stretch before expanding doesn't yield generality in embodied AI
③ No stacking small models—"never expect 100 small models stacked together to become one big model; small models are on the wrong track from the start"

## Key Results

The boldest bet is on training tempo: rejecting "perfect one task before scaling" and insisting all tasks progress in parallel from mediocrity to excellence—"going from most tasks mediocre to most tasks decent is hard to predict," a cost he explicitly accepts.

Organizational resources: RMB 100-200 million per year at current data volumes, scaling to foundation-model-level investment once data grows; the milestone is getting robots to 'play around' at home within a year, with real users by early 2028. He estimates the window for startups is 18-24 months before big companies jump in—tied to their core businesses, they can only station small labs to probe first.

## Comparison with Related Paths

- Opposing path · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Path bet: behavior/action must be a unified model, rejecting modular assembly contradicts joint co-training on heterogeneous human-robot data as structurally suboptimal
- Opposing path · [sunday_blog_20260717](https://haiguangboy.com/posts/sunday_blog_20260717) `sunday_blog_20260717_2026_07`: Path bet: rejecting 'perfect one task before scaling'—tasks improve in parallel from mediocrity contradicts path bet: home-scenario research-market fit—single point crossing deployment line starts the flywheel
- Opposing path · [wx_界面新闻_20260605](https://haiguangboy.com/posts/wx_界面新闻_20260605) `wx_界面新闻_20260605_2026_06`: Path bet: the economics of general embodied AI—'build a human' rather than specialized/pervasive deployment contradicts path bet: build the brain, not the body—betting on foundation-model identity, rejecting the motion-capability track
- Same path · [No need to imagine the future at inference—robots still hit 91.8%! Fast-WAM dismantles WAM's core assumption](https://haiguangboy.com/posts/fast-wam) `fast-wam_2026_07`: Training video objectives matter more than test-time future imagination
- Same path · [π0.5 shudders in place when grabbing a spoon fails, but Orca goes further with physics intuition gleaned from watching videos](https://haiguangboy.com/posts/orca) `orca_2026_07`: The key to a world model is readable state

## Boundaries

· All from a single media interview, no third-party verification
· Founder himself admits the path is 'hard to predict'; model architecture and embodiment form are undecided
· 7 tasks tested 900 times with 100% success, but not yet integrated into one system
· At interview time, the office 'was still barren, just a floor'

## Author's Assessment (not paper content, cross-paper synthesis)

This bet collides head-on with two other known paths. In the same camp of 'believing scaling inevitably yields emergence,' Sunday chose to punch through the deployment line at a single point, letting deployment itself start the flywheel before solving the next; he chose the opposite—refusing to fully nail one thing first. The two most committed scaling believers are diametrically opposed on the sequencing of tactics. And for companies that weld 'incremental deployment first, general brain on the sidelines' (benchmarked against Momenta) into their organizational strategy, his triple negation is almost open fire: accumulating data through vertical-scenario commercial loops to reach general embodied AI is precisely the path he opposes.

What will be settled in six months isn't just whether he's right, but which of these two logics for allocating organizational resources cracks first.