Guardian article on Song-Chun Zhu

[Rich_Morin]

‘I have to do it’: Why one of the world’s most brilliant AI scientists left the US for China

At his lavishly funded Beijing Institute for General Artificial Intelligence, Zhu is one of a handful of individuals who the Chinese government has entrusted to push the AI frontier. His ideas are now shaping undergraduate curriculums and informing policymakers. But his philosophy is strikingly different from the prevailing paradigm in the US. American companies such as OpenAI, Meta and Anthropic have collectively invested billions of dollars on the premise that, equipped with enough data and computing power, models built from neural networks – mathematical systems loosely based on neurons in the brain – could lead humanity to the holy grail of artificial general intelligence (AGI). Broadly speaking, AGI refers to a system that can perform not just narrow tasks, but any task, at a level comparable or superior to the smartest humans. Some people in tech also see AGI as a turning point, when machines become capable of runaway self-improvement. They believe large language models, powered by neural networks, may be five to 10 years away from “takeoff”.

Zhu insists that these ideas are built on sand. A sign of true intelligence, he argues, is the ability to reason towards a goal with minimal inputs – what he calls a “small data, big task” approach, compared with the “big data, small task” approach employed by large language models like ChatGPT. AGI, Zhu’s team has recently said, is characterised by qualities such as resourcefulness in novel situations, social and physical intuition, and an understanding of cause and effect. Large language models, Zhu believes, will never achieve this. Some AI experts in the US have similarly questioned the prevailing orthodoxy in Silicon Valley, and their views have grown louder this year as AI progress has slowed and new releases, like GPT-5, have disappointed. A different path is needed, and that is what Zhu is working on in Beijing.

[AgentRev]

So, I went down the rabbit hole on Song-Chun Zhu when this was published about 6 months ago, before I even knew about TBT: https://www.science.org/content/article/ai-gets-mind-its-own

Long story short, Zhu spent most of his career on computer vision, focusing on probabilistic and symbolic approaches. He dabbled a bit in deep learning when it caught on in the 2010’s, but he was always critical of it, even before LLMs.

However, when you look at the papers he co-authored over the last few years, many do involve LLMs, especially since ChatGPT took off. The papers from his lab are essentially a big melting pot. There are some interesting nuggets that might be a good fit in Monty’s toolbox later on, such as this algorithm for solving abstract reasoning tests.

Regarding his TongTong AI, I can’t seem to find any scientific paper on it. At first I thought it was a robot, but it doesn’t seem much more than an NPC from The Sims… https://www.youtube.com/watch?v=QnURXpEbPi4

Most of the info on TongTong comes from press releases, so very little details are available. Pretty much the only technical info I can find is this:

TongTong is guided by two cognitive systems - the U system (capability) and the V system (value). This allows her to approach tasks in a unique manner, depending on her current state, which is evaluated across five dimensions: hunger, boredom, thirst, fatigue, and sleepiness.

He also says he wants a “framework that emulates the top-down mechanisms in the brain.” To me, this just screams Good-Old Fashioned AI with a new coat of paint. I think he’s trying to hardcode a cognitive architecture; the very opposite of TBP! How can general intelligence emerge if you hardcode the whole damn machine!? His approach doesn’t seem to involve any biological plausibility either.

Also, he puts a lot of emphasis on “Chinese values” and helping “China dominate AI”, which is kinda shallow to be honest. A real G would acknowledge that the ultimate AGI must transcend borders, accept all of humanity under its wing, and choose to care.

[markellingsen]

I would be interested in Numenta’s response to Song-Chun Zhu’s thesis, why you do not think his vision is the correct way to proceed.

[AgentRev]

I’m not part of the team, but if I may chime in again, I think that Zhu’s vision of smarter ways to AGI than opaque deep learning is perfectly aligned with TBP’s approach. My critique above was not about his vision or thesis, but how he plans to enact it, based on the few details he’s revealed so far. So, I presume the team’s likes on my post are not meant to imply that Zhu is wrong.

Many believe that deep learning is both inefficient and reaching its limits, but few have proposed potentially viable alternatives. There’s a lot of hype, ego, and controversy out there! It’s important to slice through the noise and challenge each other to walk the walk with rigor, for the public’s sake.

[markellingsen]

Interesting thought AgentRev. My thinking about Zhu is that he would cut down on AI structures, while the human brain and its neocortex are much more complicated. But perhaps I have that wrong, so it will be great to learn what the Numenta insiders say.

Thanks for this dialogue,
Mark

[vclay]

@AgentRev thanks a lot for your more detailed replies here already!

@markellingsen as @AgentRev already pointed out, there isn’t a lot of information available on how Song-Chun Zhu’s approach actually works, so I can’t rely talk to whether it is consistent with our approach and how the brain implements intelligence. As Jeff mentioned in this thread, the way a system works internally is crucial to understanding whether it implements human-like intelligence. The general sentiment of moving away from the non-biological, power- and data-hungry deep learning/LLM solutions towards something that works more like how our brains work and how children learn sounds great.

[Side note: the Thousand Brains Project is a non-profit, independent of Numenta https://thousandbrains.org/about/\]

[markellingsen]

Thanks Dr. Clary and to AgentRev again. I think the issue is getting clearer for this amateur. Is your collective point that the “minimal inputs” Zhu seeks is a critique of structures/systems which are power-hungry and data-hungry? I have wondered whether the brain is not more complex than Silicon Valley AI models, but if Zhu’s critique is not of complexity, but of reliance on complex data, then I see your points about how Numenta’s project may be more in the spirit of Zhu’s commitments.

[AgentRev]

Zhu’s lab did publish a paper that does appear to give more details about his broader “CUV framework” for AGI robotics: A Mathematical Formulation of AGI in the (C, U, V) Framework - ScienceDirect

I’m not gonna do a deep dive this time, I’ll just leave its figures below to let you judge for yourself…

---

On a side-note, I fed the paper to Gemini, went back and forth for half an hour to try and make sense of it, and this is what Gemini concluded:

Zhu’s CUV framework defines morality as a manageable set of mathematical constraints (V), intentionally sacrificing the full complexity of human values for the sake of safety, explainability, and guaranteed AI alignment.