刘松铭 | 晚点聊 LateTalk Summary

1. Key Themes

Foundation models will eat "System 2," so a pure-model embodied company is thin

Liu argues that once frontier labs (he cites GPT-6 Astra) can perceive, plan and reason about the physical world, the moat for an embodied-AI company shifts to the layers below. He says he expected this at founding and built the company around it. As he puts it: "Today we found that Astra has come in. For the embodied field, your AI model has become thin. So from Day 1 of our founding, we have been a full-stack company." [00:00:00] He adds: "If you build a pure foundation-model company in this field, it's actually very dangerous." [00:00:13]. His reasoning for expecting it: "capability keeps rising... it will 100% find scenarios to penetrate. Because capability rising means it has the ability, and capex is also increasing, which means it has the motivation." [00:31:35]

The moat is System 1 and below: contact, tactile, motion control and hardware

LiberAI deliberately positions beneath the frontier labs' reasoning layer. Liu says: "we want to cooperate with it, we can complement each other. They solve the parts they're good at, we solve the parts they're not good at. That's the action-execution layer going down from System 2, interaction with the physical world, including processing like force and touch, then further down your motion control, your hardware, even how you deploy in long-horizon scenarios." [00:38:23] Their own tests of Astra as both System 2 and System 1 showed a clear weakness: "it can get to the general range, but it never quite touches... the last millimeter problem isn't solved very well" [01:43:10], while its strength is semantic understanding, decomposition and recovery from failure [01:43:10]. His summary of the division of labor: "Astra represents this reasoning mode, it can't eat everything. There's still a lot that needs intuition, fast, reflexive thinking." [01:44:05]

Hardware is coupled to the model far more deeply than expected

Liu says the biggest surprise in building Liber-0 was hardware. "In my expectation I could foresee that your hardware would have some impact on the model, but in our practice I found this goes even deeper. Even your different configurations, and your sensor selection, would have a very large impact on the final result, and this impact you can't solve through later training recipes." [01:19:47] He sees this as double-edged: it separates them from frontier labs, but "it slows your iteration speed" because hardware iterates far slower than software [01:20:33]. On why this protects them: "if assuming it could fully cross embodiments, with no differences between embodiments, then you wouldn't need vertical integration." [01:12:11]

Human data is the scalable path: the glove curve is the steepest, and 25% real-robot data plus gloves beats 100% teleop

Liber-0 is pretrained on 2.2 million hours of human experience (internet video, ego video, glove data, teleop, cross-embodiment). Liu's finding: data volume matters less than efficiency of the scaling curve. "We found that different data sources have very different curves. And the glove physical data we collect ourselves is among the steepest." [00:53:40] The glove supplies the missing action grounding: "you can get a lot of cognition, but you don't know how to do... a giant in cognition and a giant in action are two different things." [01:01:38] His "small leverage" claim: "though it's very little, once it fills in this part your model capability jumps." [01:01:38] On cost: real-robot teleop costs "in the thousands of RMB per hour" while gloves are "an order of magnitude lower" [01:06:01]; the glove replaces about 75% of real-robot teleop with equal or better results [01:05:41].

Vertical integration as a multiplicative advantage

Liu describes the core founding hypothesis as: "your final model effect depends not only on training the model, but on your other parts too. Including how your data-collection equipment is designed, how your embodiment is built... If that holds, then vertical integration is meaningful." [00:54:24] He quantifies the stacking logic: "maybe each part contributes 30%, multiplied together it could be two times" [00:55:06]. He insists LiberAI is "hardware for the model, not model for hardware" [00:36:49]. Hardware choices like waist and spherical wrist are made to match human motion in training data, not to finish tasks: "It's actually for the model." [01:08:44]

Commercialization: pick scenarios by a strict filter, and expect competition from non-standard automation firms

Liu's scenario criteria: not too non-standard; replicable capability; small "angle" between the scenario and the model (little fallback or customization); real willingness to pay; and the scenario must build a moat via know-how or data. "At least one of these two must be satisfied, otherwise it's meaningless and just for revenue." [01:31:37] He names the real competitors: "our core opponents will still be companies doing non-standard automation." [01:38:11] He targets a benchmark application within "two to three years" [01:33:01] and says each scenario has a flaw: home (generalization, safety), industry (cycle-time demands, replicability).

The endgame: a company on the scale of Apple, and why OpenAI and carmakers won't take the dirty middle

Liu predicts: "I think a company as big as Apple will appear. I think this is inevitable. This will definitely be one of the greatest industrial products in human history." [01:48:36] On why OpenAI won't compete in delivery: "in actual delivery everyone's situation is the same: you still have to go chew through customers, one by one, the hard bones. I don't think OpenAI has that gene, and I don't think it has much motivation." [01:36:25] On carmakers and Huawei: "they still look at volume... if you grind away for a year and only make 100 or 200 units, how much patience would they have?" [01:38:37]. He says the full-stack hardware path "may be one of the highest win-rate plays." [00:00:28]

Founder capital discipline, cycle awareness and "making sure you stay at the table"

Liu is unusually risk-conscious for a 26-year-old. He raised heavily and early: "it's not that you raise when you need it, but that you raise when you can." [01:50:11] He talked about "capital discipline," meaning efficiency on scaling curves [01:50:54]. His timeline view: the current investment trough "is temporary," and he prepares for "a 3 to 5 year scale." [01:54:33] Roots: "if family does business... very important is that you can stay at the table, because there will inevitably be various risks." [01:52:05]

Organization design: taste, subtraction, collaboration, "entropy reduction"

Liu names the cultural keywords: taste (how to decompose and trade off), restraint ("subtraction is more important than patching"), and collaboration. He describes managers who reduce entropy as scarce: "people don't naturally collaborate... a good manager who can do entropy reduction and let people collaborate well, I think that is very valuable." [01:57:10] He frames hiring as product building: "every person who joins is adding a new feature to this product." [02:02:50]

Why now: shortening of lab-to-company time in physical AI, and the founder-age cohort

Liu explains the young-founder wave through field timing and LLM-assisted learning: "this field really started to develop relatively late... RT-1 in 2022" [00:02:20], plus the fact that LLMs let young people acquire expertise that long experience once provided [00:03:12]. He says lab-breakthrough-to-commercialization timelines are being compressed by AI [00:03:37].

2. Contrarian Perspectives

"Embodied is over" was the market consensus in H2 2025, precisely when he started

When Liu began, the capital market believed the embodied wave had ended. He recalls: "the whole market's feeling at that time was that embodied was already over. And I came out at this stage to do it, so maybe people would think it's a bit late." [00:28:35] Within months, driven by AI-market spillover, world-model narratives and the MiniMax/Zhipu IPOs, sentiment reversed. The edge came from reading technical signals (video modality, UMi data parity with teleop, dexterous hands maturing) rather than sentiment [00:06:20].

Being the "most hardware-heavy" is a feature when the AI model layer is commoditized

The 2026 world-model wave produced model-focused startups. Liu went the opposite way and took criticism for it: "some investors or peers might not understand, and would think you're not doing it very cleanly" [00:37:07]. His response: pure-model players are exposed to frontier labs, so depth in hardware and delivery is the moat [00:36:04].

Better teleop-free human glove data beats expensive "gold standard" real-robot teleop

Conventional wisdom treats real-robot teleoperation as the highest-quality data. Liu's data show gloves with 25% teleop beat 100% teleop. His explanation: the teleop executor's visual perspective and the robot's recorded view mismatch, there is no force feedback so operators move slowly, and the motion pattern differs from the human video the model pretrains on [01:03:43]. "The glove is the human, so it's very aligned" [01:05:20].

Frontier labs probably won't win the physical-AI delivery game because the money is worse than software

Liu: "OpenAI today, if he invested the same effort in coding, could have 20x growth, but I think this field wouldn't... He's already earning software SaaS money very comfortably; would he still want to earn this very hard money?" [01:37:36] He cites OpenAI shutting down Sora as evidence that even better-margin businesses get cut [01:37:50].

Closed-loop retry and intuition, not pure reasoning, is the missing piece for robots

Liu disagrees with the view that GPT-7 will "break through" embodiment (Wang Xingxing's prediction). He says reasoning models cannot eat the direct-reflex layer: "the computational overhead of an intuition-type model and a thinking-type model is theoretically different... you'd be taking the long way round" [01:45:26]. Distilling a reasoning model into a small one may break chain-of-thought, a risk he flags [01:44:43].

3. Companies Identified

LiberAI (缰绳科技)

Embodied/physical-AI startup founded in late 2025 by Liu Songming. Raised six rounds from Zhenfund, Sequoia China, Shunwei, Longzhu and others. Released Liber-0, a world-action model pretrained on 2.2M hours, with a self-designed humanoid platform "Leo." Why mentioned: it is the subject; Liu calls it a full-stack company "hardware for the model." Quote: "Our core mission is to use technology, especially physical AI technology, to realize the physical-agent product and generate commercial value." [00:04:34]

OpenAI

Frontier lab. Mentioned as the main long-term threat and potential collaborator. Quote: "OpenAI had already leaked recruiting robotics people... this is certain, only a question of early or late." [00:32:15]. Also: "OpenAI... recently closed down businesses like Sora." [01:37:38]

Anthropic

Frontier lab. Mentioned in the interviewer's question about rumored acquisition of Physical Intelligence (π). Quote from interviewer: "there were rumors that Anthropic was going to acquire π, that is Physical Intelligence." [01:35:30]

Physical Intelligence (π)

Embodied-AI foundation-model startup. Mentioned as an acquisition target in rumors and as a pure-model counterpoint to LiberAI's full-stack thesis. [01:35:30]

GPT-6 Astra

OpenAI's frontier model, characterized here as having strong spatial and embodied perception and reasoning. Liu's tests: strong at semantic understanding and decomposition, weak on contact-heavy dexterous manipulation. Quote: "its length of strength is that its understanding is especially good... and when it fails it has a clear way back on track." [01:43:10]

NVIDIA GEAR Lab (DreamZero)

Released the DreamZero world-action model in February 2026, which triggered the world-model wave in China. Quote: "DreamZero itself is an embodied policy." [00:29:30]

Google (RT-1 / RT-2)

Origin of the robot-learning pretraining paradigm. Quote: "RT-1 work... in 2022 we only started to see such work come out." [00:02:20]; "the RT-1, RT-2 series... the pain point is this field is exploring pretraining, but how to do it and how to get data is unclear." [00:17:30]

Genesis AI ("GEN-1.5")

Mentioned (as transcribed "GEN-1.5") for 500K hours of mostly UMi/gripper data. Quote: "GEN maybe more UMi data... the UMi problem is that it's mainly gripper data, and the gripper and the human aren't aligned." [00:58:33]

Dyna-2 ("Dailer-2" in the transcript)

Mentioned for 1M hours of training data, mostly egocentric/human video with zero robot data in pretraining. Quote: "Dailer's choice is that in pretraining it uses zero robot data, all of it is human video." [01:00:05]

Kimi (Moonshot AI)

Chinese frontier-lab partner. Liu describes a collaboration model: "we are cooperating with Kimi now... first, we can build on its foundation model to do a lot of AI-for-infra work to shorten hardware cycles... second, for loops that touch the physical world, they solve the virtual part and we solve the physical." [01:40:18]

Tesla (Optimus)

Humanoid maker. Liu notes Optimus Gen 2 has a single waist degree of freedom versus LiberAI's multi-axis design, and says Tesla/carmakers care about volume [01:07:45]. Quote: "they still look at volume." [01:38:37]

Huawei

Mentioned as an advanced-manufacturing/tech company that might enter embodied AI. Liu does not expect them to enter in this phase. [01:38:11]

Zhenfund (真格)

First-round investor, backing Liu when he was solo. Quote: "they believed in me a lot, believed I could do this well." [00:28:02]

Sequoia China (红杉)

Second-round investor. Quote: "the second round was Sequoia supporting us." [00:29:35]

1X

Home-humanoid company. Liu cites it as an example of developed-market labor-shortage scenarios (care, cleaning). Quote: "That might be what companies like 1X are doing." [01:35:05]

DJI

Used by the interviewer as the archetype of a founder with extreme hobbyist passion (Frank Wang). [00:39:25]

MiniMax and Zhipu

Chinese AI companies whose January 2026 IPOs boosted sentiment. Quote (interviewer): "in January MiniMax and Zhipu listed, which also strengthened confidence." [00:30:19]

BYD

Used as an analogy for backward-integrating into a key component (battery) early. [02:06:44]

4. People Identified

Liu Songming (刘松铭)

Founder and CEO of LiberAI, born 2000, Tsinghua (Zhu Jun's lab), RDT-1B author. Why mentioned: the guest. Quote: "I've actually wanted to do the entrepreneurship itself." [00:40:36] Track record: entered Tsinghua through chemistry olympiad, won the Tsinghua Special Award (top 10 undergrads per year) and gave the department's graduation speech, per the host [00:15:36].

Cheng Manqi (程曼祺)

Host of LateTalk. Quote capturing the episode framing: "Tsinghua, world model, post-00s genius: this is the hottest investment theme in H1 2026." [00:00:34]

Zhu Jun (朱军)

Liu's Tsinghua advisor, whose lab produced RDT, VIDAR and AI-for-Science work. [00:06:57]

Sam Altman

Mentioned for publicly recruiting robotics people in May and emphasizing full-stack hires. Quote (host): "Sam Altman posted on Twitter publicly recruiting robotics people." [00:32:28]

Wang Xingxing (王兴兴)

Unitree founder, cited for the view that GPT-7 next year could "break through" embodiment. Quote (host): "he said when GPT-7 is released next year it may crack the embodied problem." [01:39:41] Liu disagrees in part.

Ilya Sutskever

Cited by the host for explaining why OpenAI abandoned robotics early: scaling cost, "building 150 robots is a huge cost." [01:03:32]

Frank Wang (汪滔)

DJI founder, used as the archetype of extreme-passion founder. [00:39:25]

Demis Hassabis

Used by the host as an example of a model-field founder with extreme passion. [00:40:16]

Dai Yusheng (戴雨森)

Zhenfund partner who backed Liu; transcribed as 戴汨/雨生. Quote: "Yusheng was also very bullish on me, so supported us coming out to do this." [00:28:11]

Yao Song (姚颂)

Tsinghua-linked founder the host interviewed earlier, as contrast on Tsinghua science association versus entrepreneurship association. [00:44:01]

5. Operating Insights

Treat recruiting as product development

Liu spends most of his time hiring and frames each hire as a feature. "It's as if recruiting is a product. Every person who comes in is adding a new feature to this product." [02:02:50] He personally recruited every senior leader and key technical staff, and sells on two points: the market is big enough to justify their time, and a credible founder conviction.

Publish the hypotheses you validated, not just results

Liu releases detailed technical reports with data mixes and open simulators because "the more detailed you disclose, the more solid what you do, it helps your brand because people believe you're serious" [01:15:13]. He argues the validated hypotheses outlast any benchmark: "the height achieved isn't that important... but verifying your hypotheses will have a big impact on your team's fundamental development." [02:05:20]

Guard against premature decisions when you are most familiar with a domain

Liu's self-diagnosis: bad decisions come from too little information, not bad choices, and happen more in familiar domains due to untested assumptions. "I try to let my organization resist this inertia, so different voices can appear." [02:01:06] Practical: he debates broadly but keeps final decision authority ("rights and responsibilities are equal").

Select a first commercial scenario that does not slow research

His rule: avoid very non-standard scenarios, minimize fallback engineering, and require that the pilot yields either know-how or data moat. "I don't want to do commercialization and end up unable to push the model forward." [01:30:45]

Validate first-principles hypotheses by experiment before committing capital

Liu structured Liber-0 as a set of testable bets (zero-shot, data-curve efficiency, vertical integration, learning unseen skills from gloves). Each is tied to a spending decision (glove design, body design). Quote: "I deliberately explained what hypotheses it verified." [02:01:36]

6. Overlooked Insights

Liu's field-switching method is a portable playbook

His one-line approach to entering any new domain: "first thing is to clarify the rules of this field... then figure out which of your competitive advantages you have, push what you have, and fill what you lack." [00:16:26] He applied it three times (chemistry to CS, AI-for-Science to embodied, research to company). Investors should note the pattern as a predictor of founder adaptability, and the specific research recipe: hit the field's pain point directly, and prepare a strong demo, with open-source detail so the field "avoids detours" [00:18:16].

The most valuable asset from RDT was the community, not the paper

An easily skimmed passage: after RDT-1B's open-source release, WeChat groups filled to "groups 7 and 8, with over a thousand people" [00:21:49], and those contacts became his industry network, early investors and co-founders: "some who collaborated well have become my comrades-in-arms today." [00:22:12] The takeaway is that open-sourcing a field-defining model built the founder's fundraising and hiring pipeline before the company existed.