"Slower than it looks; LLMs tend average; Not replacing humans; New, not substitutions; Cloud first, then AI; Must change your org; Just beginning... These seven predictions are not exhaustive. Could you share some insights that haven't been mentioned elsewhere?" I asked KK.

Kevin Kelly, affectionately known as "KK" among tech enthusiasts, paused for at least 30 seconds after removing his glasses. Then, after a few rounds of prolonged questions, he abruptly cut me off.

"My prediction is that, in 10 years, training data won't be important," Kelly declared.

Probably the nearest thing Silicon Valley has to a true sage, Kelly is a visionary figure in the tech world. Known for his bushy beard and graying hair, he has authored seminal works such as "Out of Control," "What Technology Wants," and "The Inevitable". Kelly has also long been celebrated for his prescient insights into future trends, having accurately forecasted developments like cloud computing, virtual reality, and the Internet of Things three decades ago. 

During his visit to Suzhou, China to participate in a Tech Lecture Series organized by Suzhou Institute of Science and Technology and the Shanghai Advanced Institute of Finance at Shanghai Jiao Tong University on June 16,2024, our conversation took place following his lecture, which stretched from the planned 20 minutes to nearly an hour.

In this conversation, Miss. Jia, founder and CEO of the prominent Chinese tech think-tank Jazzyear, discussed with Kelly of his recent observations, AI innovations, and the fundamental nature of humanity. 

Despite minor differences in specific viewpoints, both Kelly and Miss.Jia agree on a central theme: the transformative impact of AI on the world is just beginning.

Recent events: “It's my full time.”

“You have to have 1,000 hours. Maybe I’ve had 800 hours, not 1,000 hours.”

Miss. Jia: News comes and goes, and the attitude towards AI has changed a lot around the world, especially people's views on AI 2.0, AGI or big models. How much time have you devoted to tracking cutting-edge AI progress？

Kevin Kelly: It's my full time. All I do is just to read the AI papers.

Miss. Jia: Whose papers do you like most?

Kevin Kelly: Well, every day there's a new paper, a new article every hour maybe ,with some discovery about LLM. Just last week there was a paper, an article from Anthropic, on being able to illuminate the weights and try to nudge. it was this idea that there are black boxes. We don't know how they think. They're saying, well, actually, we could see into them a little bit. That was very interesting.

Miss. Jia: Would you try Midjourney or something yourself? 

Kevin Kelly: Give one a day for a year.

Miss. Jia: Now you are an AI native.

Kevin Kelly: Kind of halfway. I don't think I have 1,000 hours yet. You have to have 1,000 hours. Maybe I’ve had 800 hours, not 1,000 hours.

Miss. Jia: You are a philosopher in the field of technology. Has your philosophy of technology been interrupted or changed in the recent wave of AI?

Kevin Kelly: That's a good question. I don't think my philosophy about technology has changed. I would say it has confirmed my philosophy. I haven't seen anything that maybe changed my mind about the development, my theory of technology. So I have a theory of technology and that theory is that it's like evolution and nothing I've seen in AI has changed my theory of technology. 

Perspective: "I’m concerned about the weaponization of AI"

What are the best and worst decisions made by OpenAI?

Miss. Jia: There is a line at the top of your official website: OVER THE LONG TERM, THE FUTURE IS DECIDED BY OPTIMISTS. Are you worried about the recent wave of AI progress and rapid iterations, as well as the century-long human identity crisis you just mentioned?

Kevin Kelly: In general, I'm not worried too much. There are some things I would say I am concerned about, but I believe that we will solve those problems. Climate change. It's a problem, but we know what to do. 

There are other kind of problems like in AI, but we don't know how to solve it. Questions like the weaponization, or should we have a robot soldier? Should we have AI been able to kill? It's a really hard thing to this. So those are the kinds of things that I'm concerned about. The weaponization of AI. 

I think I'm also concerned about whether AI is open or closed, whether it's public or just owned by corporations. I think it should be public.

Miss. Jia: You think it should be public？

Kevin Kelly: Yes, public in the sense of open source.So that's another concern.

Miss. Jia: Are you still the optimist?

Kevin Kelly: Yes, I'm very optimistic. I think we will solve this,but I don't know how, and I don't think they're impossible to solve. So I'm very optimistic.

I think there are some things that people worry about that I don't worry about. Like unemployment. I'm not worried about that. Also, I'm not worried about the AIs gonna kill us. I'm not worried about that either.

Miss. Jia: You have fans all over the world, and you must know many great scientist friends. Do they agree with your views on AI, or are there more opponents?

Kevin Kelly: I Think that's interesting right now, there's a big disagree. There's two camps about the Super AI, those who are worried, and those are not. There are a lot of really good scientists who are worried. There's a lot of good scientists who are not worried. It's interesting. There's two different beliefs about the power of a super intelligence AI. I'm in the camp that is not worried about it.

Miss. Jia: I have a more specific question now. What are the best and worst decisions made by OpenAI so far?

Kevin Kelly:  OpenAI should have made their models open. They didn't. OpenAI is not open. They should make it open. That was the worst decision. I think another worst decision was firing Sam.

Their best decision is their choices to keep speeding, making it faster, to innovate faster and faster, to rehiring Sam and then committing to not be so careful and really try and grow fast. 

Boundary: "AI is good at hill climbing, not hill making"

"You can get a Midjourney or Dall-E to make a famous astronaut riding a horse, but you can't get it to have a horse riding an astronaut."

Miss. Jia: You mentioned two types of creativity, type 1 and type 2, and you drew a very interesting picture saying that AI is good at hill climbing, not hill making. What is the difference?

Kevin Kelly:  The kind of creativity that the LLMs that they operate within the boundaries of what's known. They're kind of exploring everything within that space of what I know. They're not inventing whole new territory. So breakthroughs basically are making up new territories rather than exploiting of finding solutions within an existing term. And mostly what they're doing right now is finding answers within what we, imagining pictures within the boundaries of what we know. That's why it's really hard. You can get a Midjourney or Dall-E to make a famous astronaut riding a horse, but you can't get it to have a horse riding an astronaut cuz it's outside of.

Miss. Jia: Are you a fan of the Scaling law?

Kevin Kelly: Yes, there is some. So to explain to the audience, the scaling law is saying that there is a mathematical proportion train how big the model gets, and the loss factor and how close it can become to maximizing the results. 

So far, it seems like the scaling law works. When they add compute at the training size, it gets the performance gets better. What we don't know is whether that is indefinite. Can I continue forever? Does it plateau? I think the evidence suggests that it will plateau. That's not the Internet.

Miss. Jia: There is an opinion that has been very popular in the AI industry recently - It's all about dataset, not so much with the algorithm or some other method, only the dataset is important.

Kevin Kelly: There was a paper, an article saying that the quality of data had more influence than the algorithms. And I believe that's very possible. 

I would predict that we're going to see AI companies advertise the AI is based on the training data. So you will have people say, we didn't train our algos or this stuff. We trained only on the best. We trained it on high quality books and other high quality materials. We didn't train on Reddit.

It's like education. If you have a child, how you gonna educate them? You're gonna have them read Twitter or you're gonna have them read the classics? There'll be some people say our AI only read the classics. They read the highest quality books and scientific journals. They weren't reading Reddit or Twitter, or Weibo. They were reading good stuff. They were trained on the best. And there will be some people who will sell their AI on this idea of curating the training data. Just yesterday, Getty was gonna have an AI image generator that was only trained on the Getty library.

Prediction: "In 10 years now, training data won't be important"

Doing reasoning without needing to have millions of examples.

Miss. Jia: Your fame comes largely from your identity as a prophet, but you just flashed big words on the screen: No predictions. But you mentioned seven judgments:

Slower than it looks

LLMs tend average

Not replacing humans

New, not substitutions

Cloud first, then AI

Must change your org

Just beginning

These seven predictions are not exhaustive. Could you share some insights that haven't been mentioned elsewhere?

Kevin Kelly: (long pause) Generally, if I have an idea, I tell people. But let's have a conversation, and then I will try to come up with one.

(continued pause) So about AI, tell me about AI in China. I don't know very much about AI in China. You obviously are reading papers. What would you say is happening in China right now in terms of AI?

Miss. Jia: I think the similarities between China and the United States are much greater than people think.

Kevin Kelly: The similarities？What happened？

Miss. Jia: For example, talent. There are many young talents in China, they are students or in startups. They are very similar to the young talents I met in the United States or some other countries, because AI is so sharp and so new.

My major is mathematics. When we compare AI and mathematics, the length of history is different. Many friends around me think that artificial intelligence is too complex and difficult to understand. But the history of AI is only more than half a century. If you just want to understand the overview, history, and subject classification, reading two or three books is enough for you to build a basic overview of AI. China may not have big names like Elon Musk or Sam Altman, but when you look around at young talents, the grounds are quite similar. 

The second dimension is data. Maybe China have some advantages.

Kevin Kelly: Who has access to data? A young startup, would they have access to that data?

Miss. Jia: I think we are just getting started. The government is trying to build the infrastructure for people to get the data they want in a good way. 

Kevin Kelly: What does that mean, good way? 

Miss. Jia: Data market. You know, data has been written into fundamental policies and has become a factor, just like capital, labor, technology, and land, which are called "factors of production" in China.

Kevin Kelly: Your entrepreneurs will not have any difficulty getting data?

Miss. Jia: They can get data as easy as they can get data in other countries, maybe easier. But I think that there are a lot of problems or challenges they have to face. The biggest challenge is not policy, not authority, but datasets. Datasets for various languages are different.

Kevin Kelly: So here's a prediction. My prediction is that in 10 years now, training data won't be important.

All the large language models require scaling huge methods of data. But there are other kinds of cognition intelligence. A human toddler can tell the difference between a cat and dog after seeing 12 examples. They don't need 12 million. So I think in 10 years, we'll have abilities of doing reasoning without needing to have millions of examples. And that was a huge advantage to startups at all because they don't need to have all that data. 

Essence: "Consciousness made us special, then we give consciousness to AIs."

Co-discovering what it means to be human with AI.

Miss. Jia: Can you give me some insights into your essence of human beings and artificial intelligence?

Kevin Kelly: That's the thing is. We don't know what the essence of human is. The way we're gonna find out is by making AIs. We will make an AI and we will make it creative and we'll realize, oh, we thought creativity was what made us special, but now we changed our mind because AIs can be created too. And we'll say, well, we thought consciousness was what made us special. Then we'll give consciousness to AIs.

Miss. Jia: Where will this "giving" process end?

Kevin Kelly: We'll constantly redefining ourselves because of technology and AI. And I think the more important question, it's not like who are we, but who we want to become. What would we like humans to be? That's a more powerful question, because we get to kind of choose that a little bit, deciding what it is. That's what's exciting about, for me, that's the ultimate excitement about AI, that it's illuminating who we are, as well as inspiring us to who we could become.

Miss. Jia: Accelerated computing is touching the no-man's land of science. What is the limit of this path?

Kevin Kelly: We don't have the theory of intelligence, we don't have the theory of humans. We can't prodict where AI is going cuz we have no theory about AI. If you do this, your theories have predictions. If you do this, then this will happen. If you add all this compute, then you'll get this. We don't have any theory, which is unusual. 

In physics, we have theories. If you make a big enough collider, you'll find this particle. We don't have that for intelligence right now. But I think that's what's exciting is that we're, with AI co-discovering what it means to be human. 

Miss.Jia: Thank you. I do like your answers.

Kevin Kelly: I like your questions.

Kevin Kelly (right) and Miss. Jia（left）

*Zhang Yijia is the founder and CEO of Jazzyear, and "Miss Jia" is Zhang Yijia's pen name.

(Photo: Jazzyear)