Physical AI founder bets hardware first: when foundation models arrive, pure models thin
When Liu Songming founded LiberAI, he predicted foundation models like GPT-6 Astra would consume System 2—the perception and planning layer of embodied AI. From day one, he chose hardware and product as the moat instead of building a pure model company.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Technical judgment, not industry connections, drives early-stage capital decisions
Industry convention pairs young technical talent with a veteran industry executive. Liu chose differently. At this stage, the company's biggest expenditure is compute and data collection. Deciding where that money goes and which technical paths to pursue requires technical judgment, not just resource coordination. Without it, capital efficiency collapses.
— Liu SongmingFoundation models will consume the perception layer, squeezing pure robotics companies
Liu maps robot-world interaction into three layers: System 2 (perception, planning, reasoning), System 1 (converting instructions to action sequences), System 0 (low-level motor control). Through 2023-24, venture capital believed embodied companies could own all three layers and scale toward AGI. Liu predicted from day one that foundation models would inevitably capture System 2. GPT-6 Astra's arrival proved it. This makes pure-model robotics startups dangerously thin. LiberAI chose full-stack from inception—using hardware and product to thicken the narrowing moat.
— Liu SongmingGlove data with 25% teleoperation outperforms 100% teleoperation alone
Experiments show glove data combined with 25% real-robot teleoperation data outperforms models trained on 100% real-robot teleoperation alone. Teleoperation isn't perfect: operators watch sideways while robots move forward, creating a viewpoint mismatch with the robot's camera. Lack of force feedback makes operators slow. This motion pattern diverges from smooth human actions in internet videos. Gloves worn on human hands align naturally with human motion and extract knowledge from internet videos more effectively.
— Liu SongmingHardware constraints go deeper into performance than training recipes can fix
The biggest surprise wasn't training—it was hardware. Different morphologies and sensor choices affect final performance far more than expected, and this impact can't be recovered through later training approaches. System 0 motor control algorithms couple deeply with System 1 action generation. Motor control couples tightly with the robot's joint structure, DOF, and motor parameters. This demonstrates hardware companies' structural advantage over pure model companies—but it also means iteration speed is constrained by hardware's long development cycles.
— Liu SongmingFoundation model labs lack the character for robotics' unglamorous delivery work
Consider rumors of OpenAI or Anthropic acquiring Physical Intelligence, and potential entry by carmakers or Huawei. Liu argues that real competition only emerges at delivery scale, where every customer is a hardened negotiation. This work demands patience that OpenAI likely lacks, and the incentives don't align: the same engineering effort applied to software yields 20x growth. Why chase service and manufacturing margins? Liu judges the real threat is non-standard automation specialists and prior-generation delivery firms.
— Liu SongmingPhysical robotics will eventually consolidate into an Apple-scale giant
Liu expects physical AI applications to transition from today's fragmented commercial and industrial niches toward consolidation. Eventually a mega-company emerges at Apple's scale—among the greatest industrial products in human history. This company likely emerges from firms already pursuing full-stack, hardware-first approaches. This is what LiberAI positions as its highest-probability wager.
— Liu SongmingBuild organizations that question decisions made too fast with too little information
Liu reflects on past errors: mistakes often stem not from bad option selection but from deciding too fast on incomplete information. People unconsciously filter incoming data through existing expectations. This trap is especially dangerous in familiar domains, where strong priors tempt you to skip verification. The remedy: deliberately preserve diverse voices and cultivate people who can stand up and audit—to coordinate information gaps and misaligned goals—rather than relying on one founder's final call.
— Liu SongmingIn their own words · checked verbatim
You can't have it both ways anymore: you can't stay clean, do this long-term, and compete, because the objective conditions have shifted.
就是说你不可能说 我又不dirty 我又能把这个事情做得长远 又能够竞争 因为这个客观事实已经改变了
Liu Songming35:23
My goal has always been clear: what I really wanted to do was entrepreneurship itself.
我的目标一直很明确 其实我一直就想做的事情 就是创业这个事情本身
Liu Songming41:29
I don't think OpenAI has that DNA, and I don't think it's particularly motivated to do it.
但这个事情我不认为OpenEye有这个基因 而且我也不认为它很有动力去做这个事情
Liu Songming1:36:56
The shift from fragmentation to consolidation is definitely a trend. When consolidation arrives, I think an Apple-scale giant will emerge. This is bound to happen.
我觉得从碎片走到集中 一定是一个趋势 到集中的时候 我觉得会有一个对标苹果 特别大的公司出现 我觉得这是必然的
Liu Songming1:48:01
Sometimes you make the wrong choice not because you chose poorly, but because you rushed your choice when you didn't have enough information.
有些时候你做错误的选择 不是因为你选的不好 而是你过于着急的做选择 你的信息不够
Liu Songming2:00:06
Figures
| Liber-0 pretraining data | 2.2 million hours | 56:36 |
| Glove data replacement for teleoperation | 75% | 1:05:39 |
| Zero-shot grasping task success rate | 60-70% | 50:36 |
| Zero-shot non-grasping task success rate | ~30% | 50:36 |
| Teleoperation data collection cost | ~1000+ yuan per hour | 1:06:40 |
| Gen-1.5 pretraining data | 500,000 hours | 57:36 |
| Dollar2 pretraining data | 1 million hours | 57:36 |
Glossary
- VLA
- Vision-Language-Action model: model architecture that end-to-end converts perception and language instructions into robot actions
- WAM
- World Action Model: model that predicts world state changes and outputs actions accordingly
- System 0/1/2
- Layered control architecture: System 2 handles perception and planning, System 1 generates action sequences, System 0 handles low-level motor control execution
- Zero-shot
- Model completes new tasks without additional training or examples
How to listen
Investors and engineers tracking embodied intelligence and physical AI—anyone wanting to understand how hardware companies defend their moat as foundation models expand.
The first 25 minutes covering growth history and Tsinghua entrepreneurship background can be skipped. The core technical insights are in the second half.