Zhang Hongjiang is a pc scientist and senior executive who has change into certainly one of China’s most outspoken voices on the necessity to develop artificial intelligence safely.
After earning his PhD in Denmark, he worked in Singapore and Palo Alto, California, for several years. He then returned to China within the early 2000s to assist arrange Microsoft Research Asia, before occurring to construct Kingsoft into certainly one of China’s leading software corporations.
He stepped away from that business in 2016 only to return to the frontline of Chinese tech, two years later, by establishing the Beijing Academy of Artificial Intelligence (BAAI), a non-profit that brings together industry and academia.
In recent years, Zhang has change into China’s leading advocate for the regulation of AI, to make sure it just isn’t a threat to humanity. Here, he talks to the FT’s Ryan McMorrow and Nian Liu concerning the importance of international collaboration on AI safeguards, in addition to the opportunities and challenges facing China.
Nian Liu: It looks as if you’re paying loads of attention to AI governance?
Zhang Hongjiang: I actually have spent loads of time trying to lift awareness within the research community, industry, and government that our attention shouldn’t only be directed on the potential risks of AI that we’re already aware of, comparable to fake news, bias, and misinformation. These are AI misuse.
The greater potential risk is existential risk. How can we design and control the more powerful AI systems of the long run in order that they don’t escape human control?
We developed the definition of existential risk at a conference in Beijing in March. The most meaningful part is the red lines that we have now defined.
For instance: an AI system (should) never replicate and improve itself. This red line is super vital. When the system has the potential to breed itself, to enhance itself, it gets uncontrolled.
Second is deception. AI systems should never have the potential to deceive humans.
Another obvious one is that AI systems shouldn’t have the potential to provide weapons of mass destruction, chemical weapons. Also, AI systems should never have persuasion power . . . stronger than humans.
The global research community has to work together, after which call on global governments to work together, because this just isn’t a risk on your country alone. It’s an enormous risk for entire mankind.
I learned a lot on the International Dialogue on AI Safety within the UK last October. It’s actually a system of labor from the underside up. It’s technical work. It’s not only policy work.
I realised that, in Europe and the US — especially in Europe — there are technical individuals who have been working in that field for a few years and who’ve developed quite just a few systems to measure and define the danger of AI systems.
The British took great initiative. Like they did in last 12 months’s first international government summit.
Ryan McMorrow: When you’re in these discussions, are the viewpoints of leading Chinese scientists and policymakers much like Western ones?
ZH: Very much. The argument centres on whether the present AI systems actually possess artificial general intelligence (AGI) capabilities, or if they’ll result in AGI, and the way far-off. But, if you happen to agree the danger is there, then there’s really not much difference on viewpoints.
(Ex-Google AI pioneer) Geoffrey Hinton’s work has shown that the digital system learns faster than biological systems, which implies that AI learns faster than human beings — which implies that AI will, at some point, surpass human intelligence. If you suspect that, then it’s a matter of time. You higher start doing something. If you consider the potential risk, like what number of species disappeared, you higher prepare for it, and hopefully prevent it from ever happening.
Scientific collaboration must be a standard practice. But, unfortunately, it’s not a standard practice now. Of course, AI is probably the most advanced technology so it has change into more sensitive. Especially between China and the US, geopolitics does affect these collaborations. I hope, a minimum of on the science level, this collaboration can proceed.
NL: Speaking of China and the US, how do you’re thinking that the US export controls on processors will affect the long-term development of Chinese AI?
ZH: I believe it’ll have a big impact. I’ve all the time considered AI as a system comprised of three things: algorithms; computing power; and data. Without the computing power, today’s technology would change into more limited. The essence of GPT models is their scalability. That is, if you happen to increase the scale, the variety of parameters of a model, its performance will improve. If you scale up the quantity of knowledge you’re feeding into the model, its performance can even improve. This is what we call the scaling law for models.
And, as you increase each the parameters and data, you could have to also scale up computing power. So, if you happen to’re limiting computing power, you’ll in fact hit a roadblock. There’s little doubt about it.
RM: To get around these obstacles, China is pushing the event of homemade processors. But, at the identical time, most of the present models listed below are built on top of the Nvidia chip ecosystem. Is it possible to simply port a model between different chip ecosystems?
ZH: Their software has got to be compatible, which is hard. People have been constructing loads of models, and probably the most efficient models have been built throughout the Nvidia ecosystem. So, if you must construct your personal ecosystem, it’ll take effort and time. It’s higher to be compatible with the Nvidia ecosystem.
It’s very very like software compatibility issues between Windows and Mac. For example, if you happen to construct something on Android, you essentially must adjust it to work on iOS. Today, for corporations developing software apps, they must develop on each platforms, which implies they should have dedicated teams. The same principle applies if you happen to are working on models: you may need to construct for 2 systems, which shall be hard and likewise costly.
RM: How hard is it to construct two different systems?
ZH: It is pretty hard since you’re tuning software stacks, that are sets of basic function modules needed for the training systems to run on. If you could have to construct the whole platform all together, that’s just loads of effort. It’s as hard as constructing one other Android.
RM: It looks as if Huawei is within the result in be China’s Nvidia: it has the Ascend chips made here (in China) that the US can’t block. Is that what corporations must be constructing on now?
ZH: My understanding is that it’ll level off. Just like if you construct software, you don’t need to construct on too many operating systems. Think about it, within the PC age, within the cell phone age, there are two systems. That’s it. Think concerning the chip architecture, what number of architectures are there? Not that many.
RM: Will computing power be much more critical within the multi-modal future, with humanoids and visual modelsDo those require much more computing power than plain language?
ZH: Yes. Vision data, which incorporates image and video data, is far larger in volume in comparison with language data.
RM: So computing power is a hurdle, but where do China’s benefits lie? Policymakers have been very energetic here, is that a bonus? Or the talent?
ZH: Policy is the final thing that involves mind after we speak about China’s benefits in AI. I believe China’s benefits in AI mostly lie amongst young entrepreneurs who come through disappointment after disappointment but still proceed to do start-ups and pursue their dreams. I believe the one place that we will compare is Silicon Valley.
Someone gave me a number, though I’m sure you could find more accurate ones: 30 per cent of the worldwide top AI talent was originally born in China but a big portion of them work within the US. If 10 per cent of the highest talent stays in China, that also represents a big number of individuals.
On top of that, there’s the vast market. There are so many scenarios (by which) to use AI — which, in turn, presents good research topics and provides good research data. This enables research institutions and universities to work on good problems.
So I believe talent, application scenarios, and entrepreneurship are China’s benefits. But I don’t think government policies are necessarily a bonus.
NL: What about data? Your old colleague Kaifu Lee made the purpose in his book that China has all this data, which shall be an enormous advantage.
ZH: China is a big web market, so China has had tons of knowledge as a bonus up to now. But, then, we realise after we have a look at the GPT models, and have a look at the information that’s fed into the models, and the distributions of knowledge, they’re from the online. And, if you happen to have a look at the online, China’s corpus just isn’t that much. It’s within the low single digits. I believe lower than 5 per cent. A variety of languages have lower than 10 per cent.
I believe English accounts for about 60 per cent or 70 per cent, so it’s predominantly English. So, if you happen to train your model with data from the online, then Chinese data just isn’t that much. (And) whichever language has more data, it’ll be higher in that exact language. If you have a look at Wikipedia, all the online data, Chinese data should not dominant. So I won’t say, by way of language data, there’s a bonus.
But, after we come to embodied AI, after we come to robotics, after we come to manufacturing, China has tons of knowledge. Way greater than other countries.
For example, a wise city model. China definitely has more data than another country. Just have a look at the variety of cameras in China. Look on the number of electrical cars which have basic autonomous driving capability. They have so many cameras. So it depends upon which area you’re talking about.
RM: Is this next generation of vision models in humanoid robots just getting began?
ZH: It has change into a hot topic up to now 18 months, especially with developments like GPT-4. It has impressive capabilities in recognising images and objects in images. And, then, if you happen to have a look at Sora, and Gemini 1.5, in addition to Claude from Anthropic, and the brand new Llama 3, all of them exhibit this multi-modality — which, mainly, is image capability.
If you equip a robot with a big multi-modality model, it might perform tasks way beyond what it was trained for. It can even understand commands that it was not originally trained on. Suddenly, you realise a robot can understand way greater than you thought.
An excellent example is RT2, released by Google a few 12 months ago. For instance, if you ask the robot to select up a toy on top of this table, a toy that’s an already extinct animal, it picks a dinosaur. That’s a really complicated reasoning process since it wasn’t directly told that a dinosaur is an extinct animal, however the language model knows that. So, amongst the assorted animal toys, it picks up dinosaurs.
Another example is: ‘Give a can of Coke to Taylor Swift.’ On the table, there have been 4 picture frames. The robot picked up a Coke and placed it on the image of Taylor Swift. So, take into consideration that process. The robot can recognise who’s in that picture. It has a notion of who Taylor Swift is. That wasn’t trained into the language model.
That’s why you see Figure, a brand new robot start-up that teamed up with OpenAI, which OpenAI invested in. You see one other start-up from Berkeley: Pi. There are lots of them.
RM: What concerning the ones in China?
ZH: Galaxy Robot, which was incubated at BAAI, is an example. It was originally founded by a professor from Beijing University who worked on embodied AI at BAAI.
In the most recent demo I saw, if you say “Oh, I feel thirsty”, the robot arm will pick a bottle of water amongst five various things and deliver it to you. That’s what you wish and expect a great nanny to do.
The instructions to the robot aren’t any longer very explicit. It’s the robot that understands. There are quite just a few corporations now moving in that direction.
RM: A variety of Chinese corporations are using Llama as their models. Is that what corporations everywhere in the world are doing?
ZH: Llama models are probably the most powerful — and, due to this fact, popular — large-language models within the open source world. So I’m sure loads of people and corporations will use them. And, for somebody who mostly focuses on the tutorial world, it’s super helpful to have an open source model you can analyse, tune, and do research on, given how costly it’s to coach a large-language model from scratch.
Again, you’ll be able to draw an analogy to software, where Linux and open source became super popular. And there are numerous open source databases which can be super popular. I’d say that web corporations are entirely depending on those open source databases and systems. They helped speed up the event of the web and the cloud. And China definitely advantages quite a bit from that.
NL: That results in this open-source versus closed-source debate. Baidu’s Robin Li recently said that open-source models “make little sense” and keeping the models closed was crucial for having a viable business model. What do you’re thinking that of his comments?
ZH: I have to say I don’t entirely agree with him. But that argument has all the time been there, going back 30 years. It was the closed source Windows and Mac versus Linux, after which, later, the closed source iOS versus the open source Android. Although Android is open source, it’s heavily controlled by one company. So, this debate has all the time been present.
If you have a look at the business world, the leaders tend to not favour open source because they’re leaders of their field. Meanwhile, followers and others who are attempting to alter things normally adopt the open source approach. Linux has done it, Android has done it, and each have been very successful. So I won’t say which one has an absolute advantage. It will take an extended time before we will tell which one will win. But, more likely, they’ll coexist.
RM: China is all the time excellent at applying stuff, commercialising. Can you speak about some interesting applications of AI that you just’ve seen here?
ZH: I believe there are two corporations which have done an ideal job. One is Beijing-based Moonshot, which actually has some association with BAAI. Their product Kimi, very very like ChatGPT, is great and very talked-about.
Another one is Minimax, a Shanghai-based company. They began their effort (to construct) large models a minimum of a 12 months before (the) ChatGPT launch, so that they should not a copycat. They are specializing in applications like digital avatars.
I’d say, if you happen to see any good applications in other markets, China will soon have it, if not already. I’m not saying that China only follows. Actually, once in a while, in certain areas, China has taken the lead.
NL: In the west, all these AI start-ups are raising crazy amounts of cash at very high valuations. Is it the identical in China?
ZH: Yes, it’s the identical in China. The only difference is that China has more corporations in a single area — which I believe might be too many. The US has possibly three or 4 start-ups specializing in foundation models. China has what number of?
NL: Hundreds
ZH: The Chinese market could be very “” (over-competitive). This isn’t recent; it’s been like this for the last 20 years. As for the bubbles, I wouldn’t say China has greater than the US. At this moment, I believe these are still good bubbles. There shall be winners. I believe it won’t be very long before we see consolidation.
RM: What will the business model be for all of those foundational models? Will Big Tech’s models win? Or those from China’s more nimble start-ups?
ZH: In the US, it’s very clear: enterprise productivity tools. We already see the effect with tools like Co-pilot in Office. However, for consumer applications, individuals are still exploring. We haven’t seen an enormous success as yet. China definitely has quite a bit more people exploring this area.
The big tech giants have to interact with AI models because, otherwise, they aren’t any longer a platform company. They either must develop large models or they’ve to amass them. So, definitely, they’ll proceed investing.
For the small start-ups, their challenge isn’t just to lift enough money and develop good models, but additionally to define their business models and find their users. The two corporations I discussed, Minimax and Moonshot, were focused on consumers from the very starting. In China, whichever start-up succeeds in the patron space . . . will ultimately succeed, despite the immense pressure from the large giants.
RM: So you’ve stepped back from leading BAAI, what are you spending your time on now?
ZH: I’m excited and consider that enormous AI models will change the way in which we do robotics, and eventually give robotics a break, it’s just so thrilling. Also, I feel that China has some benefits here. China has the most important manufacturing base and is more sophisticated than many other countries in hardware. That’s what interests me.
I’d say a big a part of my motivation is driven by curiosity. It’s something I feel enthusiastic about and likewise something I feel I can learn from. Also, I spend more time now overseas. I spend time in Singapore and take part in some government-organised meetings and activities there. I would love to spend more time in Silicon Valley and simply to be more in contact.
I’m still involved in organising BAAI’s technical conference on AI. It’s purely technical. The entire programme is organised by technical people actively working in the sphere, specializing in various facets of AI. People come here to learn and exchange ideas, identical to at any academic conference. It’s not business.
There’s value in having people come together to learn and exchange ideas. Last 12 months we had Hinton, (Yann) LeCun, Sam Altman and we had Chinese scientists exchange and argue with them.
