Gaoyang: A Thinking Reed, A Machine That Acts | Beipo Initiative

Over the past two months, Spirit AI has closed two consecutive funding rounds, totaling 3 billion yuan.

Embodied intelligence is rapidly becoming the consensus among capital and markets.

As new hot topics and narratives keep emerging, Gao Yang's attention always returns to the same question: If we're going to build a general-purpose foundation model for embodied intelligence, what problems lie ahead? How do we actually solve them and push scientific progress forward?

For the first episode of "North Face Project," we invited Gao Yang, co-founder of Spirit AI and assistant professor at Tsinghua University, to discuss embodied intelligence's approaching "GPT-3 moment" — breaking down the technical premises and future outlook behind this prediction, from data and models to system capabilities. We also turned our attention to the individual. Gao Yang says: man is but a thinking reed. He moves at his own pace, not trying to become any "standard answer," returning to himself between external expectations and personal fulfillment, maintaining his internal rhythm, and unfolding his own vitality.

Scan the QR code to listen to the full podcast.

If you have any thoughts you'd like to share, feel free to leave a comment below.

The original podcast transcript is approximately 25,000 words; this article has been edited and condensed to roughly 4,500 words.

Enjoy

"ChatGPT Reshaped My AI Values"

Oasis Capital: What did you see back then that made you think VLA (vision-language-action model) might be the right path?

Gao Yang: Historically speaking, ChatGPT's emergence was a huge shock to me and many researchers at Berkeley. We used to not know where intelligence lay, or how it could emerge. Around 2018–2019, almost every professor and student at Berkeley didn't believe OpenAI's claim that scale could produce miracles — we thought they were bluffing, saying things they couldn't themselves verify.

Until ChatGPT 3.5 came out, my AI values were reshaped. Many previous assumptions, expectations, and premises suddenly collapsed like a quantum wave function — it confirmed some things and falsified others.

Oasis Capital: It collapsed.

Gao Yang: Yes. Simply put, a statement that may be common sense today: massive data produces intelligence.

Actions, language, and images — to a model, they're fundamentally not that different. At that point in 2023, the conclusion was easy to extrapolate to the action space. Once the first path worked, it immediately meant the second and third paths could definitely work on similar problems too, just requiring some adjustments to details — tuning hyperparameters, for instance — and the model could certainly be built.

At this point in time, robotic manipulation capabilities remain very weak, and there's plenty of noise questioning whether this can really be done. But for me two years ago, this was definitely achievable.

Oasis Capital: You were certain two years ago?

Gao Yang: Yes, definitely achievable.

Oasis Capital: And you were certain because of ChatGPT?

Gao Yang: Because GPT worked, and fundamentally, the Universal Approximation theorem from 40 years ago provided theoretical explanation — neural networks with hidden layers can fit any continuous function in the world. This has been proven countless times in experiments. So if large language models solved it this way, it necessarily meant video and action could be solved similarly.

At that moment, we were already quite certain that embodied models would arrive.

VLA: A paradigm proposed and gradually developed in 2020, holding that a unified model can directly map from perception to action in an end-to-end manner, enabling systems to complete tasks in real environments — a critical path toward general embodied intelligence.

Universal Approximation Theorem: A theory proposed in the late 1980s stating that neural networks, given sufficient width, can approximate any continuous function to arbitrary precision, proving models have adequate expressive power and serving as the theoretical starting point for deep learning.

Next Year, Embodied Intelligence Will Enter Its GPT-3 Moment

Oasis Capital: At that time, how long did you estimate it would take for embodied intelligence to reach something like ChatGPT 3?

Gao Yang: In 2023, I thought it might take 5–8 years to reach GPT-3 or 3.5 level.

Oasis Capital: Now you've been in industry for two years, and you're probably the most advanced team globally in this direction, and the leading team in China — has your view changed?

Gao Yang: It has, and it's moved up significantly. My current prediction is around early or mid-2027.

Oasis Capital: Have we passed GPT-2 yet? What milestone confirmed this for you?

Gao Yang: We've just passed GPT-2. The standard is that you can see some generalization, though it may not be that strong yet.

Oasis Capital: Many people define the GPT moment from different angles. From your perspective, what do GPT-2 and GPT-3 mean respectively?

Gao Yang: There are several dimensions. On the effects dimension, reaching the GPT-3 moment means that in any scenario, when you say anything to the robot, it gives some seemingly intelligent response.

Oasis Capital: Regardless of whether it does it well or has hallucinations — first, generalization is definitely achieved. At worst, it might tear your clothes.

Gao Yang: Exactly. From a purely technical perspective, in terms of data volume, the growth from GPT-2 to GPT-3 was roughly 100x.

Oasis Capital: Any other angles? For instance, with language models, GPT-2 might have proven the existence of Scaling Law at a theoretical or laboratory level, but didn't engineer it; the GPT-3 milestone was that it could actually be implemented in engineering. From the embodied intelligence perspective, do you agree with this view?

Gao Yang: In embodied intelligence, Scaling Law actually appeared before GPT-3. Because in language models, the first ones to walk this path had to feel their way across the river. When they first started experimenting, people quickly discovered that throwing a large amount of data at the model made it significantly better — but they didn't know if they were doing it right, or if some things could be done better.

But each experiment took two months, and people couldn't afford to wait. So researchers studied how to use small-scale experiments to extrapolate to large-scale scenarios — that was the more rational choice at the time.

Today's rational judgment is somewhat different. We know Scaling Law likely exists. What we want to verify is whether our current methodology carries risks when scaling. Because unlike large language models, embodied intelligence can't freely obtain 100x data — you need to put in extra effort to collect and clean it, requiring upfront investment to create data and do scaling.

Oasis Capital: This development is quite interesting, because large language models as the "south face" climbers are further ahead, and their development in turn gives more confidence to those on the "north face" of embodied intelligence. Because at 4,000 meters they encountered snow or oxygen deprivation, so at 2,000 meters you prepare oxygen in advance. Data issues that used to only be considered at the GPT-3 moment are now being thought about earlier, so you'll actually move faster — because they've helped blaze some trails for you.

Gao Yang: Yes. I often say what I call History Echo — it's not repetition, but echo. From the perspective of historical unit models, each echo teaches us a lot. Including autonomous driving development — it's like seeing where they'll suffer oxygen deprivation at certain altitudes, so we avoid those traps from the start.

VLA and World Models

Oasis Capital: At this moment, who do you think has the most advanced embodied intelligence model globally?

Gao Yang: Physical Intelligence. Much information here is non-transparent, but judging from behaviors demonstrated by models like Generalist, PI's model is better in terms of generalization.

Oasis Capital: There's another direction that's getting increasing attention — world models (World Model), especially with recent large funding rounds for World Lab and AMI Lab. What's your view on this path?

Gao Yang: Let me first clarify the concepts.

What VLA does is essentially provide a description of the world — it can be images, it can be tactile sensation. I want to express action A, and what machine learning captures is the mapping from language to action A. What world models do is assume the current description and the action A I want to take — this A is also an assumed fact — and the robot outputs the next state S' based on this. They're fundamentally solving different problems, but these two concepts can mutually reinforce each other. Both are model-based reinforcement learning, things thoroughly studied in Model-based RL. They have a kind of conjugate relationship: VLA is the action policy, the robot moves in the world; the world model provides that world for the robot to act in, and tells it what the next state S' looks like.

So returning to how world models will affect embodied intelligence development — in the short term, I think they're one way to augment data, because you can perform actions in world models and generate new data. But at least today, the generation quality isn't sufficient for models to freely interact in the world. They function more as an auxiliary loss that makes VLA training more stable, giving the model constraints that make the mapping easier to learn.

Oasis Capital: Is it possible that once world models mature, generated simulation data could completely replace real data for training?

Gao Yang: Model development depends on real-world data collection, which is then put into world models for training, so machines can interact with the physical world more realistically. Models can't imagine data they haven't seen.

But I don't rule out the possibility of joint training with world models plus real-world collected data, making training effects so good that world models themselves can generate more data. I think that's one future direction.

Loss Function: In deep learning models, a function introduced to enhance main task performance by optimizing secondary objectives related to the main task, helping the model learn better representations — particularly when data is complex or model depth is substantial.

Similarities and Differences Between Spirit AI and PI in Technical Choices

Oasis Capital: Since Sergey is one of PI's co-founders, you had interactions in the Berkeley lab, and later did your postdoc under the same advisor — so you're essentially academic siblings.

In terms of exploration in robotics, what are your similarities and differences? And how did these eventually lead to two companies, Spirit AI and PI — what are the technical path similarities and differences between the companies?

Gao Yang: Sergey worked on robotics many years before me, but there were 4–5 years when nobody knew what was correct. During that time Sergey also tried many things, they just weren't as effective as today.

Our similarities and differences may stem from several aspects. First, we're both doing end-to-end VLA model training. Our thinking is quite aligned on robot memory mechanisms, current network architectures, and data primarily coming from the real world.

Where we differ: first, on wearable device data collection. I don't know if Sergey has changed his view, but he was previously opposed, while we're quite embracing. Even some compromised robot data, we still consider very valuable.

A second difference — though this is gradually converging — is the use of human video data. PI previously didn't use human video data at all, but we used a lot for pre-training. Today PI is also gradually starting to use human video data, so consensus is slowly converging on this point.

"The Ultimate Form Will Be a Single Model"

Oasis Capital: Some say VLA will subsume motion control, others say motion control will subsume VLA; still others say you're both wrong, and large language models will subsume everything. What's your view?

Gao Yang: I don't think any will subsume the others.

Oasis Capital: For instance, as embodied intelligence develops, assuming we pass the GPT-3 moment, where do you think the boundary will be drawn between embodied intelligence large models and today's large language models? Or is it possible they'll just be one model?

Gao Yang: I lean toward the ultimate form being a single model.

Oasis Capital: Is it an either/or relationship or an integrated one?

Gao Yang: An integrated relationship.

Oasis Capital: Since the future involves integration of many models, is it possible everyone today is just one piece?

Gao Yang: Today's discussions of end-to-end have different starting and ending points in the hierarchy. Some may start from VLA to the motion control layer, others from different points. I think first, VLA and motion control will become end-to-end models — each with their own parameters, relatively independent, but their mutual interaction will be end-to-end. At the top layer, like current OpenClaw or what we call the reasoning layer, it's a human-scale workflow of several hours. This layer's development trend won't be fully end-to-end, but rather models doing some research themselves, storing content to something like a hard drive, then reading data to understand and continue deeper research.

So I think in future systems, the more bottom-layer parts will be more tightly integrated in end-to-end fashion; the more top-layer parts will be more like current Agent forms, doing macro-level orchestration.

Oasis Capital: Then there must be some dimension that integrates everything together. What dimension do you think this is?

Gao Yang: It ultimately becomes a system across multiple time scales. For instance, the fastest system is motion control and some reflex control — touching something hot and pulling your hand back, or adjusting when you lose balance. This is the bottom layer. Above that is the VLA layer, requiring the robot to have some understanding of tasks.

Motion control frequency is roughly 50–100 Hz, needing fine-tuning every ~10 milliseconds; if you want the robot to be smoother, this may go even higher. VLA frequency is lower, currently around 10–20 Hz. Above that, language is actually non-thinking mode, roughly 1 Hz, outputting 10–20 tokens per second — humans have this capability. Further up is the reasoning layer, perhaps only 0.1, 0.01 Hz. They're a temporally hierarchical system.

In the human brain there are brain waves at different wavelengths that serve as integrators, combining different sensory information into the same event. For instance, when I hold a water cup, it's a bit cold, and meanwhile my hand feels weight and pressure — these are different events integrated so you feel it's not three separate things but one thing. I think ultimately there will be something similar in models, integrating across wavelengths.

Non-thinking Mode: Refers to a model's operation mode where it directly generates results without explicit reasoning or multi-step thinking. This contrasts with Thinking Mode, which performs multi-step reasoning or chain-of-thought.

Man Is a Thinking Reed, No Different from Machines

Oasis Capital: I'm curious — what is your life's reward function? What value are you pursuing?

Gao Yang: That's quite a difficult question. Let me think.

I think it's probably the realization of self-worth. You see something, find it interesting, want to pursue it, and in the process — you feel happy, you've realized your value.

Oasis Capital: If there's something that makes you happy but has no value, and another thing with great value but doesn't make you happy, which would you choose?

Gao Yang: I'd probably choose the "I feel happy" thing.

Oasis Capital: I'd choose the same, haha. Because the first principle of Buddhism is "liberation from suffering and attainment of joy" — everything is for happiness. Punishment for a person is suffering; incentive is joy. So happiness is the only standard.

Gao Yang: I feel flattered, haha. My wife says I haven't studied Buddhism, but she thinks I have quite the cultivation or something.

Oasis Capital: You've done robotics research for so many years — I believe in the process you've certainly gained deeper understanding of yourself and life. As a living being, what insights has research brought you?

Gao Yang: Many aspects. One of them is, I think humans are actually just machines. What we discussed earlier is also something I think about in research — I've increasingly found that humans and machines are very similar in many underlying mechanisms.

This sounds very "non-humanist," because it says humans are machines. But from another angle, in moments when I feel value realization, I feel as if I am a thinking reed.

I say humans are like machines because, corresponding to "reed" in this phrase, machines can be mass-produced and seemingly lack specific value; reeds too — tens of thousands grow in a pond at once. But humans can think, like the lobster we discussed at lunch — a being with stories, preferences, likes and dislikes.

Whether reeds or machines, both are actually quite special.