Topics Pioneers Insight Method Author
Back

187: A Conversation with LiberAI’s Liu Songming: Astra Arrives, and How a Post-2000 Physical AI Founder Sees the Odds

·晚点聊 LateTalk

Rendering the long image…

Topics Embodied AI and Humanoid Mass Commercialization Embodied AI Funding and Exit Scrutiny
Pioneers 刘松铭

Summary

  • Liu Songming sees foundation models entering the field as inevitable, so Libra AI has pursued a full-stack strategy from day one to prepare for “thinner models.” He believes the model called Extra in the interview has already absorbed part of the embodied AI moat, including at least System 2. Independent companies need to build their edge in System 1 action execution, physical feedback, motion control, hardware and delivery. “Hardware serving the model, not the model serving hardware” is what sets his approach apart from another class of full-stack companies.
  • Libra 0’s key advance is measurable zero-shot success across tasks of varying difficulty. Liu said its success rate is around 60–70% on pick-and-place, about 30% on non-pick-and-place tasks, and above 10% on some four- or five-step long-horizon tasks. The team had GPT-6 generate 60 tasks, 20 in each of the 3 categories. He sees this as an early sign of general capability, while acknowledging that the team has yet to establish a commercial benchmark and that a technical progress curve cannot be treated as a revenue curve.
  • The data moat comes down to action information and scaling efficiency, not simply accumulating hours. Libra 0 uses 2.2 million hours of human experience. In the team’s experiments, 10,000 hours of glove data outperformed 2 million hours of internet video on manipulation; combining gloves with 25% real-robot teleoperation data also beat using 100% real-robot teleoperation. Liu said gloves supply the physical actions missing from video, can replace about 75% of teleoperation needs, and cost an order of magnitude less to collect, possibly even less.
  • Humanoid hardware improves knowledge transfer but also puts the company on a slower iteration cycle. In tests cited by the show, the company’s in-house robot achieved over 90% success on tasks such as twisting and bending to pick up a slipper, about 30% higher than the comparison platform. The waist, wrist and motion control all affect model performance. Liu found the interactions between hardware and model run deeper than expected: some problems cannot be fixed with a training recipe, and hardware has many sequential steps that are hard to accelerate by spending more.
  • Commercialization remains a search for repeatable use cases; the goal is to establish one recurring-revenue benchmark over the next 2–3 years. The milk-tea demo can now run continuously for five minutes and fifty-something seconds, but whether the sugar pump dispenses, how the robot senses contact, and how much human backup is needed all change the economics. Liu prioritizes standardized settings that fit the model’s capabilities and can generate data or know-how. He admits he has yet to see a clear answer that meets both the technical and commercial requirements.
  • Liu sees non-standard automation firms as the more immediate competitors in delivery. OpenAI will bring pressure on talent, funding and models, but real-world delivery still means “cracking one hard problem at a time.” He questions whether large-model companies have the incentive to make that investment, and whether automakers will patiently pursue a market that delivers only 100 or 200 units a year. Libra AI’s response is to divide the work with foundation-model companies: it has discussed hardware R&D and physical-loop collaboration with Kimi rather than competing for the entire AGI ecosystem.
  • Six funding rounds have bought time to operate across cycles; capital discipline will determine how far that time goes. Liu’s fundraising principle is to build reserves when capital is available instead of waiting until the company needs money. On spending, he aims to do more with fewer resources. He attributes the current cooldown to capital expectations running ahead of actual returns. He remains optimistic over the long term but will not bet on a recovery next year, and recommends preparing for a 3–5-year horizon.
  • His view of the endgame is strong; the biggest unanswered question is where the profits will accrue. Liu believes general-purpose robots will emerge within 5 years and that, as the industry moves into the consumer market, a company on the scale of Apple will emerge. He calls the current full-stack approach “one of the strategies with the highest odds of success.” But he is still working through how much value will accrue to models, hardware, application infrastructure and the supply chain.

Deep dive

1. World models are a means; physical agents are the product

  • Liu Songming does not equate “world models” with Libra AI’s mission. They are a technology the company is exploring, “more of a means.” The end goal is to use physical AI to build physical-agent products and create commercial value.
  • He keeps some distance from the New Labs positioning. The industry is now willing to give technical breakthroughs more time, researching first and finding applications later. Libra AI believes physical AI research and commercial applications are closely linked and cannot be kept apart for long.
  • His starting point is direct: “Every company’s ultimate mission is to create commercial value.” Model progress has to be understood in that context.

2. A young founder entered as the technology-to-market cycle shortened

  • Liu, born in 2000, accepts “young wunderkind” as a description of his age but is more interested in the industry trend behind it: physical AI itself is young. Google’s RT-1 did not appear until 2022, so people entering this research field are naturally young too.
  • Large models have lowered the barrier to acquiring knowledge and professional expertise, changing the relative advantage of long experience. Cheng Manqi observed that the time between a technical breakthrough and the people who master it founding companies is shrinking. Liu believes the path from lab breakthrough to commercial application is accelerating as well.
  • With help from a senior lab alumnus, he tried ChatGPT early and for the first time felt the model was “a bit like a person.” By 2023, he had concluded that foundation models would become enormous, and that an independent startup should look for applications that could work with them while building a distinctive moat. That is what drew him to the physical world.

3. The startup goal stayed constant; his comparative advantage shaped the field

  • Liu saw entrepreneurship as his life’s main track in high school. He never seriously considered taking a job or first becoming a professor. His logic was that resources and influence are ultimately there to get things done; he could organize resources directly through a startup and work on what he believed mattered.
  • His parents were more likely to ask, “What kind of person do you want to become?” than focus only on grades. Choosing Tsinghua was also about its startup resources. He judged that his weakness in Chinese made the gaokao route a long shot, while he had an aptitude for chemistry competitions and access to training, so he chose that path into Tsinghua.
  • When he moved from chemical engineering to computer science, he considered the asset requirements of starting a company: chemical engineering is capital-intensive, while computer-related opportunities are friendlier to young founders. He started in AI for Science because his physics and chemistry background gave him a comparative advantage, not because he had committed to a particular technical route from the outset.
  • Tsinghua’s entrepreneurship programs exposed him to law, intellectual property, contracts, management and finance. During the pandemic, he also worked on a project that used online-learning data to recommend practice questions: the availability of online data and students’ need for personalization formed the startup hypothesis.

4. RDT showed that pretraining works but left the data-supply question open

  • Looking back at early robot learning, Liu describes a shift from collecting one dataset per task to building shared capabilities. Making breakfast and folding clothes are not entirely unrelated; pretraining can learn what they have in common and then improve performance across tasks.
  • The team worked on RDT-1B in 2024. Released in October 2024, it used 1.2 billion parameters; the team later developed RDT-2. The key finding was that on unseen tasks, pretraining followed by post-training data worked better than using post-training data alone.
  • Diffusion is well suited to continuous physical actions, while discrete sampling can lose information. But at the time, the team relied mainly on teleoperation and data from different robot bodies, and generation was very slow. The work showed that pretraining mattered, but not how to scale it up.

5. The value of a viral research project is helping the most people avoid dead ends

  • When entering a new field, Liu first looks for its rules and pain points. RT-1 and RT-2 had already proposed pretraining, but how to train models and obtain data remained unclear. Research needed to push directly on those questions, and robotics needed strong enough demos for others to see the progress.
  • RDT’s open-source release shared not just the results but how the team did the work and what went wrong. Liu measures research value by “how many people it helps avoid dead ends.” A viral project, in his view, addresses a problem shared by enough researchers.
  • The WeChat group originally set up for discussion quickly expanded into 7 or 8 groups, with 1,000–2,000 members. The community connected him with peers, industry and future collaborators. Industry feedback also exposed both demand for flexible manipulation and real-world tasks that models could not yet solve.

6. If you cannot explain the plan for 3 months out, you should not rush to take resources

  • After RDT attracted attention, investors and startups began reaching out, but Liu did not leave his PhD program in his second year. His clearest reason was, “I hadn’t figured it out.” He could not even clearly list what he would do in the first 3 months after receiving resources.
  • He wanted at least a 60–70% level of confidence about roughly how to proceed before starting. A large market does not guarantee that every early-stage company will survive: “The pie in this industry will definitely be huge…but on the way from here to a major sector, we all have to survive.” He already saw early entrants as potential competitors and considered the difficulty of raising follow-on funding as a student-founded startup.
  • On the pairing of a young technical lead with a resource-rich senior partner, he says it depends on what drives the company. If the main inputs are compute, experimentation and data, and the key challenge is a technical breakthrough, the person at the top needs technical sense to decide where the money goes and how to maintain capital efficiency.

7. Three technical signals brought the human-data approach into focus

  • When the team truly got started in the second half of 2025, Liu first saw the value of video. VITA, work by peers from Zhu Jun’s lab at Tsinghua, showed him that video modalities and video foundation models could help physical models generalize without relying entirely on data generated by robots.
  • Large-scale research using data without robot bodies provided a second signal: data without robot bodies could achieve results comparable to teleoperation, suggesting collection costs could fall substantially. He had previously been unsure how to scale; now he began to see a path.
  • Data without robot bodies also exposed a limitation: it mainly recorded grippers, while internet video mainly showed people. There was a gap between them. If the goal was to use the largest source of human video, a more natural bridge might be to collect people’s manipulation data directly.
  • The third signal was the gradual maturity of dexterous hands. Human data, internet video and robot hardware capable of using those skills began to converge. That was when he felt the human-data-driven path had become clear.

8. Six funding rounds carried the company through the sector’s trough and the world-model boom

  • Libra AI was founded at the end of 2025, with recruiting partners and fundraising moving forward at the same time. Liu said that when ZhenFund backed the first round, it was basically just him. His connections with investors such as Debbie and Yusen also grew out of the attention RDT had attracted; Sequoia backed the company later.
  • The market did not feel like “embodied AI is hot” at the time. Some people thought “embodied AI is already over.” In 2026, confidence from the listings of MiniMax and Zhipu, spillover from the AI market boom, North American comparisons and a new narrative brought capital back.
  • Cheng Manqi sees Nvidia’s release of DreamZero in February 2026 as a major moment in the world action model boom. Liu said he was already pursuing a similar technical path when he started, just without branding it as “world models.” The show’s introduction said the company had completed 6 funding rounds, with investors including ZhenFund, Sequoia, Shunwei and Long-Z.

9. Foundation models entering the field was expected; the pace was faster than expected

  • Liu believes foundation models entering physical AI is nearly inevitable: improving capabilities give them the potential to enter new settings, while more compute creates an incentive to find applications. He had also heard from friends in Silicon Valley that OpenAI had reportedly been recruiting for robotics roles quite early.
  • The interview refers to the model as “GPT-6 Extra” or “Extra.” Liu believes perception of distance, size and depth, as well as understanding the world through coding, should already fall within a foundation model’s capabilities. What exceeded his expectations was the speed.
  • He divides the system into 3 layers: System 2 handles perception, planning and understanding; System 1 turns instructions into action sequences; System 0 executes them reliably through motion control. This is the framework he uses to discuss how the work is divided.
  • Early fundraising narratives often counted all 3 layers as an opportunity for embodied AI companies, sometimes moving from “embodied brains” to AGI. When he started the company, Liu already saw System 2 as at risk. Foundation models entering the field now means part of the model moat once assigned to embodied AI companies has been taken away.

10. As the model gets thinner, execution, hardware and product need to get stronger

  • “Your AI model gets thin” is how Liu sums up the foundation-model threat. The response is not to insist on a pure-model identity but to build full stack from day one, adding product and hardware capabilities. He believes pure embodied foundation-model companies will face very strong competition from foundation models.
  • Libra AI has always focused more on System 1: how actions are executed, adjusted afterward, retried after failure, and supported by physical feedback, compensation, touch, motion control and deployment. Liu has never viewed understanding and interaction as the company’s main competitive edge.
  • In his view, what distinguishes Libra AI from some earlier full-stack companies is that “hardware serves the model, rather than the model serving hardware.” Some have criticized this approach as “impure” and insufficiently focused, but he believes companies have to accept the new competitive reality if foundation models can expand into the field.

11. Libra 0 reports zero-shot success rates across task categories

  • Liu defines zero-shot as completing a new problem or task without additional demonstrations or training. By his account, Libra 0 achieves a zero-shot success rate of around 60–70% on pick-and-place tasks.
  • Its rate is around 30% on non-pick-and-place tasks, and above 10% on some long-horizon tasks involving 4 or 5 steps. What excites him most is that tasks once considered almost impossible to complete zero-shot now have successful examples.
  • The team had GPT-6 generate 60 tasks, with 20 each for grasping, non-grasping and long-horizon tasks. He sees this as a more objective test than coming up with the tasks themselves, while acknowledging that the industry has lacked systematic evaluations covering these categories.
  • One example is putting objects from a table into a bag, then putting the bag into a drawer, including operating the drawer along the way. This is no longer repeated grasping; it combines different skills. The progress he emphasizes is that “we’re starting to see a success rate.”

12. The 2.2 million hours of data are not all the same training asset

  • Libra 0 was pretrained on 2.2 million hours of human experience, much of it internet video. Other sources included open-source and vendor-collected egocentric video, glove data, real-robot teleoperation and data from other robot bodies.
  • Cheng Manqi cited two comparisons: the G1.5 model reported 500,000 hours in August; Dyna 2 reported 1 million hours in September. Liu said much of G1.5’s data was still gripper data, while Dyna 2 used no robot data during pretraining and relied more on human video.
  • Liu does not dispute that more data is usually better, but stresses that the rate of improvement varies. Diversity, collection methods and whether the data contains action information explain more than hours alone. A “million hours” at one company cannot simply be treated as equivalent to a million hours at another.

13. Gloves connect understanding human video to real actions

  • Internet video contains rich physical knowledge but lacks precise action data: watching alone may not reveal exactly how a hand moves or how precisely it moves. Gloves provide action grounding, aligning with human video on one side and adding executable physical information on the other.
  • He describes the difference as “a giant in cognition and a giant in action are two different things.” A small amount of glove data can have an outsized effect because it fills precisely the gaps in video that matter for execution.
  • The model also takes the robot body as a conditioning input: grasping a cup calls for different strategies with a human hand and a robot gripper. Humans are the largest source of training data, but you cannot “deploy on humans” in the end, so transferring from people to robots is a question every system has to answer.

14. Compare scaling slopes first; gloves can also teach new skills

  • In a manipulation experiment disclosed by the team, 1,000 hours of glove data outperformed 2 million hours of internet video. Liu’s explanation is not that video is useless; video mainly provides cognition, while gloves fill in the part involving physical interaction with the world.
  • Scaling curves vary significantly by data source. Glove data is among the steepest, and can even approach real-robot data collected across different robot bodies. This shifted his focus from “how much data to pile up” to “how to scale more effectively.”
  • In another experiment, real-robot data covered only pick-and-place, while glove data covered other tasks. After mixed training, the model still learned non-grasping skills. This went further than the team had hoped to test: new knowledge could come from gloves without collecting expensive real-robot data for every skill.

15. Replacing 75% of teleoperation can improve results because teleoperation is not a perfect policy

  • In the team’s experiment, glove data combined with 25% real-robot teleoperation data outperformed 100% real-robot teleoperation. Liu interprets this as a demonstration that about 75% of teleoperation can be replaced, a meaningful result for cost and scaling.
  • Teleoperation is not inherently the highest-quality policy. The operator watches the robot from the side, so the person’s control intent and field of view do not fully match what the robot records. That can introduce bias into the demonstrated actions.
  • The lack of direct feedback also makes teleoperation slower and more sluggish, with action patterns unlike the fluid human actions seen in internet video. Wearing gloves and doing the task directly preserves natural human movement and makes it easier to align with pretrained video.
  • His rough cost range is thousands of RMB per hour for real-robot teleoperation, while gloves cost an order of magnitude less—possibly hundreds of RMB or less. Glove data still has to be collected in the physical world, so it cannot scale as easily as internet video to an enormous volume.

16. The humanoid waist and wrist are designed to make use of human knowledge

  • Full stack does not mean making every part in-house. Liu’s line is to control only the design choices that can improve the final model; mature components and supply chains can come from outside. The current priorities are gloves for data collection and robot-body design.
  • Cheng Manqi asks why the waist should do more than move up and down. Liu explains that people turn sideways and bend over when wiping a table; those actions are naturally present in video. Without the corresponding degrees of freedom, a robot will be misaligned with the policy it has learned.
  • The wrist follows the same logic: a human-like rotation can support twisting, while some serial-joint designs may rotate the forearm as well and cause collisions. Slim arms can also reach narrow spaces. Whether the added cost is worthwhile ultimately depends on the model’s benefit.

17. Cross-body capability does not mean every robot body performs equally well

  • In the report cited by Cheng Manqi, the in-house robot achieved over 90% success on tasks such as twisting and bending to pick up a slipper, about 30% higher than the comparison platform. Liu believes the hardware itself constrains the policy, so the whole gap cannot be attributed to the model.
  • Asked whether “one brain across multiple bodies” therefore means weaker performance, he says cross-body capability is necessary, but performance can still vary. At least while models are learning so much from humans, a more human-like body makes that knowledge easier to use.
  • He wants an ecosystem in which the model can fit multiple robot bodies but works best with the company’s own body, while its own body and model reinforce each other. If robot bodies make no difference, vertical integration loses its purpose and foundation models can move down the stack more easily.

18. All 3 stages are called pretraining because watching and acting stay intertwined

  • Libra 0 first learns physical cognition from large amounts of internet video, then gradually increases the share of data carrying physical action information, such as gloves and real-robot data. “Learn first, watch first, then act” summarizes the process; it does not mean the two capabilities are fully separated.
  • Cheng Manqi asks why the later stage with a small amount of robot data is not called post-training. Liu explains that internet video remains in all 3 stages; only the mix changes. He believes fully learning to watch and then fully learning to act would break the transfer relationship that should exist between them.
  • The team’s actual post-training is still ahead. It may include RL, new approaches that scale better or adaptation to a particular robot body. The current report validates a mixed-pretraining approach; it does not mean every future stage has been completed.

19. A public roadmap builds confidence; the moat is the whole pipeline

  • Asked whether a report this detailed might teach competitors, Liu says the priority at this early stage is to clarify the path, build confidence in the industry and show outsiders that the team is doing serious work. The company also opened a web simulator where people can upload photos of everyday objects to test the model.
  • Many peers are already working on dexterous hands. Liu does not think one hardware idea alone is enough to replicate the full result. Data-collection hardware, data processing, training and validation form a complex pipeline that demands broad team capabilities; that is the moat he has more confidence in.

20. Touch and cross-task RL remain two gaps

  • Liu openly says Libra 0 has not yet solved touch. The challenge is not just the model; the team also needs stable, inexpensive, reliable hardware suitable for production to collect the data and then make use of the new modality.
  • He sees touch as an important part of System 1 and one that foundation models are less able to cover directly today. But it is a priority for the future, not a complete capability the system already has.
  • On RL, he hopes to find a method like RLHF that consistently raises success rates across tasks rather than improving just one task. The team is still exploring and has not presented a general approach that is already validated.

21. Hardware interactions run deeper than expected; cash cannot buy every shortcut

  • The main challenge after starting the company that exceeded Liu’s expectations was hardware. Mechanical configuration, sensor selection, joints and motor parameters can all have a significant effect. Some differences “cannot be fixed later with a training recipe.”
  • System 0 motion control and System 1 affect each other, and motion control is tightly linked to the robot body. This both deepens the difference from foundation models and requires the model and hardware to iterate together, leaving the slower side to set the pace.
  • The biggest resource investments today are still compute and data, including collection equipment and labor. Hardware research and stability improvements are relatively less capital-intensive, but take time and require careful design. Funding needs will rise significantly with large-scale deployment and mass production.
  • Liu agrees that hardware has many sequential stages, with timelines that are hard to compress. Competitors with more capital will have to go through the same process.

22. The continuous milk-tea demo exposes the details that matter for delivery

  • The release video shows a milk-tea task taking five minutes and fifty-something seconds, with no cuts or speed-up. The company is also setting up a milk-tea station for employees to use, but Liu is clear that it is still being fixed up and has many issues to solve before it can operate without human intervention.
  • Cheng Manqi focuses on the sugar pump: a person can tell whether sugar came out and adjust how hard to press, but how does a robot do that? Liu thinks it may need additional modalities such as force and sound; vision alone is not enough, and the relevant hardware does not yet support them well.
  • The host suggests another option: modify the milk-tea machine to remove difficult-to-handle containers. The team is currently trying to work without changing the environment to reduce deployment costs. They come back to comparing the costs rather than treating either approach as universally right.
  • At the beginning, some human-robot cooperation may be acceptable, but the share of work requiring human backup determines the economics. If many steps still require people, “you can’t save on labor.” A model that can demo a task may still fail to create enough commercial value.

23. Every setting has a hard constraint; retail is only one area of exploration

  • Liu refuses to present any deployment direction as a perfect use case. Homes raise generalization and safety challenges. Industry demands throughput; move to a more flexible step and the questions become labor cost, task content and the ability to replicate across workstations.
  • Commercial services and retail are areas of focus, but the absence of demos in industry, logistics or homes does not mean they are off the table. Liu stresses that the team is still testing broadly; for now, it is exploring.
  • Globally, he is more focused on labor shortages in developed countries. Home care, cleaning, retail, services and hotels may have more urgent needs. The latter settings have some degree of structure while still facing ongoing demand for workers.

24. A benchmark use case should be repeatable, fit the model and leave an asset behind

  • First, the setting cannot be too non-standard, or it will slow model development. Second, it should be close to the model’s capabilities, so new technology solves a real problem and reduces customization and human backup. Throughput and tolerance requirements also need to be within what the model can handle.
  • Customers need a real reason to buy, and a project needs to leave behind at least one of two assets: know-how that turns model capabilities into commercial value, or data that makes the company more competitive in the future. “It’s not just for revenue.”
  • The robot body should be as standardized as possible to reduce manufacturing and adaptation costs, but if a particular setting is too narrow or too high, the team will not refuse to modify it just to keep everything uniform. Whether to adapt is still an economic decision.
  • The ideal is to establish a benchmark use case over the next 2–3 years. Asked about scale and margins, Liu gives no numbers. He only says value must come from the model and generalization, and the use case must be repeatable and generate recurring revenue. So far, he has not seen a clear answer that meets all those conditions.

25. The first competitors they encounter may be non-standard automation firms

  • On the threat of OpenAI building or acquiring a robotics team, Liu acknowledges pressure on models, talent and funding, but says delivery still means “cracking one hard problem at a time.” An acquisition may not provide immediate leverage either; deep coupling between software and hardware can bring high organizational costs.
  • He questions the incentive to invest: in his hypothetical, putting the same effort into coding “might produce 20x growth,” which physical AI cannot. He also points to OpenAI shutting down Sora and asks why: he believes Sora could make money more easily than physical AI, while earning money from services and manufacturing is hard work.
  • His question about automakers and advanced manufacturers such as Huawei is whether they have the scale and patience. If it takes a year’s work to deliver only 100 or 200 units, far below their automotive businesses, it is unclear whether they will invest heavily at this stage.
  • From a customer’s perspective, a model plus hardware first has to beat existing non-standard automation and the previous generation of delivery solutions. That is why Liu lists those firms as real competitors: the comparison is about cost and delivery performance, not technology labels.

26. Extra handles understanding and planning well; contact still fails in the last millimeter

  • Liu says the team currently works with Extra, with Extra handling System 2 and its own model handling System 1. In zero-shot comparisons, he says, its model’s success rate is several times Extra’s.
  • The gap is concentrated in contact and dexterous manipulation; the tests have not involved force control. He observes that Extra can reach roughly the right area but cannot handle “the last 1 millimeter”: it tries repeatedly, twitches and fails to make real contact.
  • Extra is strong at semantic understanding, breaking down tasks, planning and finding a clear retry path after failure. But if Libra AI’s System 1 cannot do something, Extra may not be able to fix it. Liu uses the fact that “people can slip too” to underscore the importance of retries: long-horizon reliability does not come from hitting the target on the first try every time.

27. Whether reasoning models can absorb reflexive control remains disputed

  • Cheng Manqi cites Wang Xingxing’s October prediction that the release of GPT-7 next year might crack the embodied AI problem. Liu agrees there will be an impact, but believes that under GPT’s current paradigm it cannot solve everything.
  • He distinguishes reasoning from intuition: the former involves extended thought, while the latter controls movement quickly and reflexively. Tasks involving speed, throughput and incomplete information require more than a simplified version of reasoning.
  • If future models are designed specifically for contact and tactile control, or use smaller models, he believes the boundary could shift. But if they still rely on chain-of-thought reasoning, whether a smaller model can retain those capabilities becomes another question. Not every advance can be attributed to the same foundation-model spillover.
  • Cheng Manqi leaves room for the other view: some foundation-model companies she has spoken with believe they will eventually reach the execution layer too. They agree that compute and latency on the robot also matter. Liu adds that the two model types have different compute costs and optimization goals.

28. A Kimi partnership could connect the virtual loop to a physical interface

  • Liu disclosed that the team is discussing 2 types of collaboration with Kimi. One would build on Kimi’s foundation model and cover a range of RSI-related work, with the goal of shortening hardware R&D cycles and helping a hardware-constrained industry move faster.
  • The other is more long term: research and production loops cannot happen entirely in a virtual world; they also need a physical interface. Foundation models would handle more general, virtual work, while Libra AI would handle execution in the physical world. This is a division of labor they are exploring, not a loop they have already completed.

29. Tackle fragmented applications first, then build toward the consumer market

  • Cheng Manqi asks why, given that execution and delivery are harder and slower, Libra AI does not aim for the biggest opportunity like OpenAI. Liu’s answer is that the position in the ecosystem is already occupied, making it difficult for later entrants to take the same one. They can either join or work where the incumbent has less advantage and is unlikely to invest heavily in the near term.
  • “You can’t say you don’t want to get your hands dirty, want to build something for the long term, and still expect to compete.” Slow returns on applications today do not mean the end market is small; edge markets may give new companies time to grow.
  • If robots enter a large To C market, Liu believes the industry will consolidate from fragmented applications and produce a company on the scale of Apple. “It will definitely be one of the greatest industrial products in human history.” He sees full stack, hardware and product capabilities as competitive assets by then.
  • That is Libra AI’s vision, not a claim that it will certainly win. Liu says only that it is “one of the strategies with the highest odds of success.” Cheng Manqi notes that execution-focused companies could also move up the stack while foundation models move down. He agrees the competition could cross over, but says demand has not yet reached that point.

30. Build reserves when funding is available; aim for a steeper curve on spending

  • With capital expectations cooling after the second half of this year, Liu’s preparation is simple: “Don’t wait until you need to raise.” Build reserves when fundraising is available. Hardware timelines and technical uncertainty mean the company needs time, and its 6 funding rounds also support 2–3 years of commercial exploration.
  • For Liu, “capital discipline” starts with efficiency: data sources and training methods have different scaling curves, so the goal should be to do more with fewer resources, not turn a larger funding round into an equally large burn.
  • His sense of risk comes partly from growing up around a family traditional business. Several times he thought a particular risk would not occur, only for it to happen. That gradually taught him not to consider only the best-case scenario; every cycle can also differ from the one before.
  • He believes the current downturn reflects AI expectations running ahead of actual returns, while the long-term technical payoff remains intact. The timeline is hard to call. He does not think it will last 5 or 10 years and leans toward less than 5, but will not say that things will recover next year. His recommendation is to prepare on a 3–5-year horizon.

31. A full-stack team relies on taste, subtraction and active coordination

  • Liu describes taste as how people frame a problem, break it down and make trade-offs. People may have different expertise, but they need shared standards for choosing which problems to solve. The company also hopes those standards will attract people with taste.
  • He values “doing less” over continually patching things, aiming to keep the product, model architecture and workflow as simple as possible. He did not elaborate on specific examples, so this remains a stated organizational principle.
  • Collaboration does not happen automatically. Differences in background, information and goals keep creating confusion, so managers have to coordinate actively. The team hires experienced people and has them mentor people who are technically strong but have less management experience. Liu sees people who can coordinate teams across disciplines as rare.

32. The CEO makes the call; discussion helps guard against bias

  • The company encourages discussion, but Liu makes the final call because “authority and responsibility should go together.” He wants discussion to provide more information, not replace the person ultimately responsible.
  • Looking back at mistakes, he has found that some were not failures to choose but choices made too quickly. Familiar topics make it especially easy to skip steps, rely on untested assumptions and filter later information, remembering only what supports the original view.
  • He wants the organization to produce dissenting voices that bring him back to what is actually happening. “Concepts and opinions don’t matter.” Looking out 2 or 3 years, how strong a model was at a given moment may not matter much; the assumptions the work tested can matter far more to the company’s direction.

33. Build the company like a product; each hire adds a capability

  • Over the past few months, Liu has spent most of his time recruiting. By his account, apart from himself, he personally spent time recruiting all mid- to senior-level hires and some key technical employees. The company is like a product assembled piece by piece: “Every person who joins is like adding a new feature to the product.”
  • To persuade experienced people, he looks at the opportunity from the perspective of someone investing in their career: Is the problem big enough to be worth working on? Can the company capture a share of the opportunity? He then builds trust through professionalism, judgment and conviction. Cheng Manqi also observed that at their first meeting, he asked what topics the media needed and paid close attention to the other person’s needs.

34. The 5-year vision is clear; how profits will be divided is not

  • Liu sees both the technical scaling over the next few months and general-purpose robots in 5 years as highly certain. He believes that by then, tasks in the physical world can be handled much like tasks by today’s agents. The uncertainty is mainly about how to get there, especially whether model progress and the pace of commercialization will match.
  • The question he keeps turning over is what share the company and its technical approach can capture once products are deployed and profits begin to be divided. How far the model can reach, and what still depends on hardware and application infrastructure, will shape how value is distributed. He has no clear answer yet.
  • Cheng Manqi compares this with BYD’s early investment in batteries: supply-chain layers worth controlling early may shape the industry for years. Liu agrees that the chain is long and how the value gets divided is a central question.
  • Liu connects Libra AI with “liberty.” Its Chinese name, 将闲科技, also carries the hope of freeing people from chores and heavy labor. The further-reaching ideal is no longer needing some people to “carry the burden” for others, with robots taking on necessary work. Serving individuals is the largest market and the ultimate goal he sees.