Screenshot of OpenAI CEO Sam Altman's speech at KLCII (Source: Photo by Titanium Media App Editor)
As an important academic conference held for five consecutive years, KLCII has been attended by more than 500 top experts such as Turing Award winners. This year, including Turing Award winner Geoffrey Hinton, "Turing Award" winner and one of the "Deep Learning Three" Yang Likun (Yann LeCun), OpenAI CEO Sam Altman, academician Zhang Cymbal of the Chinese Academy of Sciences and other heavyweights in the field of AI to discuss the frontier and hot topics of artificial intelligence.
On June 10, Altman gave his first speech in China, talking about AI risks and future general AI development models, bluntly saying that in order to understand general AI technology, OpenAI must promote AI research reform.
However, the AI academic community does not fully agree with Altman's statement. Stuart Russell, a professor at Berkeley, lowered his head on the spot to modify the PPT, did not look at Altman at all, and even attacked OpenAI-developed ChatGPT and GPT-4 in his speech for not "answering" questions, they do not understand the world, nor is it a general AI development step.
Yang Likun countered GPT, believing that this autoregressive model simply does not work, has no ability to plan and reason, and as the text data becomes larger and the error rate is higher, the next step in AI should be a world-class model. He predicts that in the next 5 years, people without normal brains will continue to use autoregressive LLM (large language modeling), and GPT systems will soon be abandoned.
As the big model has become the dominant force in the development of the AI industry, a debate between academia and OpenAI has begun.
Sam Altman: There could be very powerful AI systems in ten years, and we need to be ready immediately
ChatGPT is popular around the world, and Sam Altman, the father of ChatGPT, co-founder and CEO of OpenAI, has also attracted attention.
Recently, Altman was doing a "global tour", visiting nearly 20 countries on five continents, including the United Arab Emirates, India and South Korea, meeting with political and business leaders, frequently releasing important information such as OpenAI will not IPO, AI needs regulatory regulations like an atomic bomb, and investing in Korean startups, demonstrating OpenAI's sincerity in cooperating with the government.
At the KLCII AI Security and Alignment Forum, Altman gave an opening speech and participated in a dialogue lasting about 40 minutes. It was Altman's first public speaking in China since ChatGPT's popularity.
Altman said that the world tour, where he met with students, developers and heads of state, inspired him to already see that people around the world are already using AI technology and getting feedback from users, as well as the opportunity to have meaningful conversations with countries about AI safety regulation, which is important to ensure increasingly robust, safe and reliable AI system deployments.
He believes that most of the global attention is focused on today's AI problems, and "given the progress we have made, I am confident that we will get there." And he wants to talk about the future.
Specifically, Altman mentioned that we are now witnessing an AI revolution, so we need to bring responsible AI technology to the world. He stressed that within a decade, the world may have very powerful AI systems.
"Imagine that in the next decade around the world, general-purpose AI systems may have 100,000 lines of binary code, surpassing human expertise in almost every field. These systems may end up exceeding the collective productivity of our largest companies. The AI revolution will create shared wealth and make it possible to dramatically improve the standard of living for everyone, but to achieve this, we must manage risk together. Altman said.
He stressed that universal AI will fundamentally change the powerful forces of our civilization, meaningful international cooperation and coordination are necessary, and everyone will benefit from a cooperative governance approach. If we navigate this path safely and responsibly, AGI systems can create unparalleled economic prosperity for the global economy, address shared challenges such as climate change and global health security, and improve societal well-being.
For international cooperation to build global trust in the safe deployment of artificial intelligence (AI), Altman proposes three approaches: scalable oversight, explainability, generalized (generalized) AI.
Altman believes that humans are unlikely to find out if some malicious models are doing something nefarious. So he's investing in new, complementary research directions in hopes of achieving breakthroughs. While scalable supervision is an attempt to use AI systems to assist humans in finding other system flaws, and explanatory capabilities are to explain GPT-2 neurons with GPT-4, although there is still a long way to go, OpenAI believes that machine learning techniques can further improve AI interpretability.
Altman emphasized that the ultimate goal is to train AI systems to help align the research itself, as future models become smarter and more helpful, so as to better realize the goal of general AI while reducing AI risk security.
"We see great potential for researchers in the U.S., China, and around the world to collaborate on AI goals, all working to solve general AI technology challenges. I believe we will be able to leverage general-purpose AI to solve the world's most important problems and significantly improve the quality of human life. Altman said at the end of his speech.
Afterwards, KLCII Chairman Dr Zhang Hongjiang spoke with Altman.
Zhang Hongjiang threw out some key questions about AI security, and Altman revealed information about AI security, open source plans and whether there will be GPT-5 in the short term, and Titanium Media App sorted out several key contents of the answer:
- Altman believes that within 10 years the world will likely have a powerful AI system, and we need to prepare immediately.
- After touring the world, Altman found that the risks and opportunities of general AI were taken very seriously. He bluntly said that global cooperation is always difficult, but this is an opportunity, AI has really brought the world together, and he can propose a framework and safety standard for the system. Altman emphasized that we must design AI systems that are equitable, representative, and inclusive.
- He specifically quoted the words of Lao Tzu's Tao Te Ching: A journey of a thousand miles begins with a single step. He believes that the most constructive step that can be taken now is the cooperation between the international scientific and technological community.
- Altman praised during the conversation that China has some of the world's best AI talent. Considering that solving the difficulty of aligning AI systems requires the best minds from around the world, he really hopes that AI researchers in China can make a huge contribution in terms of AI risk.
- As a next step, OpenAI's technical focus on training powerful AI large models is to build a global database that reflects global AI values and preferences, as well as develop tables to share AI safety research.
- When Zhang Hongjiang asked: Will OpenAI open source large models? Altman said we will have more open source in the future, but there is no specific model or timeline.
- When asked about general AI and big models, what's next for it? Will GPT-5 be seen soon? Altman said I'm also curious that we don't have an answer. We will not have GPT-5 anytime soon.
- As for why OpenAI was developed, Altman said that when we started doing OpenAI, we thought the chances of success were low. But to figure out how to build AGI, OpenAI is going to drive the AI revolution. Of course, as part of this, OpenAI must ensure security, but the benefits to society are enormous. "I really believe that this (AGI) is going to be the most transformative thing society has ever built."
Altman also tweeted after the meeting to thank him for the invited speech.
But interestingly, during Altman's speech, the next speaker, computer scientist Stuart Russell, a professor at Berkeley, looked down and modified the PPT, not looking at Altman at all.
He even bluntly said in his speech: General artificial intelligence (AGI) has not yet been reached, the big language model is just one of the puzzle pieces, we even what the puzzle will look like in the end, and what is missing, it is not certain. He said ChatGPT and GPT-4 were not "answering" questions, they didn't understand the world.
To a certain extent, this seems to indicate the degree of recognition of Altman's "playing tai chi" view in AI academia.
Yang Likun vs. Max Tegmark: Should we be optimistic about developing AI or strengthening control?
In addition to Altman, there was also a wonderful speech at the KLCII conference on June 9, and Yann LeCun, winner of the "Turing Award" in France, one of the "Big Three of Deep Learning", and Meta's chief artificial intelligence (AI) scientist, and Max Tegmark, founder of the Future Life Institute, a well-known physicist and AI scientist, talked about AI risks and challenges at 4 a.m.
Yang Likun believes that we should be optimistic about the development of artificial intelligence, and now is far from the time to be nervous. But TegMark bluntly said that it is necessary to strengthen the control of AI technology now.
In his speech, Yang Likun came up with the logic of countering GPT: autoregressive models simply do not work, because they do not have the ability to plan and reason. Generating autoregressive large language models based solely on probability cannot solve the problem of hallucinations and errors at all. As the input text increases, the probability of errors increases exponentially.
So where is the next step for artificial intelligence in general intelligence (AGI) if you want to move to general intelligence (AGI)?
The answer given by Yang Likun is the world model. A model that mimics the human brain not only at the neural level, but also a world model that perfectly fits the human brain divisions in cognitive modules. The biggest difference between it and the big language model is that it can have planning and forecasting capabilities (world model) and cost accounting capabilities (cost module).
Yang Likun believes that through the world model, it can truly understand the world and predict and plan for the future. With the costing module, combined with a simple requirement (must plan the future according to the logic of the most cost-effective action), it can eliminate all potential toxicity and unreliability. However, for how to learn the world model, Yang Likun only gave some planning ideas, such as using self-supervised models to train, we must establish a multi-level thinking model. He also admits that no deep learning training has done this before, and no one knows how to do it.
Tegmark directly said at the scene that we need organizations to develop a super-intelligent AI that can be effectively managed, which may face new risks and use AI safely, rather than just pursuing speed.
"Now, the question is, can we really control AI? I think the answer is yes, but we need to study it more and more comprehensively. Tegmark believes that if AI were closer to Life 3.0, we could do more than fight stupid battles on this little spinning ball in space.
When asked to evaluate the level of AI governance, Tegmark praised that China has done the most in regulating AI so far, with Europe in second place and the United States in third. He thinks that's a good thing. "I think we can do better."
Regarding human anxiety, Tegmark said the economy and job market are changing faster and faster, and if you're strong in the basics and very good at creative open-mindedness, you have the flexibility to go with the flow.
KLCII 3.0 is open source, and the big model evaluation system is online
As the organizer of Altman and a number of Turing Award winners, Beijing KLCII Artificial Intelligence Research Institute (hereinafter referred to as "KLCII", BAAI) also showcased the latest strength and progress of China's AI models at this year's KLCII Conference.
At the opening ceremony on June 9, KLCII announced the latest open source "Wudao 3.0" series of large models, including the release of the first "Aquila" language large model series that supports bilingual knowledge in Chinese and English and open source commercial license, the "FlagEval" large model evaluation system that provides a three-dimensional evaluation framework of "ability-task-index", the FlagOpen large model open source technology system that strives to build a "new Linux" ecology in the era of large models, and six "FlagEval" large model evaluation systems. Horizon" visual large model research results.
After the meeting, Professor Huang Tiejun, President of KLCII, told Titanium Media App that the main constraint of the current large model is that the domestic training is too small and the industry is too hot, and now the tens of billions of models have just begun to emerge. Although there are also some technical capabilities in the middle, there is definitely a gap in the level of intelligence. Repetitive force, the more it emits, the divergent resources.
Huang Tiejun believes that with the development of the domestic large model industry, the reasonable number of large models that can survive in the future is about 3. Today's big models are an intermediate product of technological iteration.
Huang Tiejun emphasized that large models have become the leading force in the development of the AI industry, and "Wudao 3.0" has moved towards a new stage of comprehensive open source. As an institution that has been working hard in the direction of general AI, KLCII hopes to do more supporting work in the future and contribute its unique strength.
"Our KLCII scholars have started the exploration of large models, taken the lead in forming a research team for large models, and become the main force of China's big model learning research today. We were the first to foresee the era of AI big models. Huang Tiejun said in his speech.
In June 2021, KLCII officially released the ultra-large-scale, trillion-level intelligent model "Wudao 2.0", and China's first original virtual student based on "Wudao 2.0", Hua Zhibing, was officially unveiled in Beijing. It is reported that the scale of Wudao 2.0 parameters reached 1.75 trillion, which is 10 times that of the AI model GPT-3, breaking the previous record of 1.6 trillion parameters created by the Google Switch Transformer pre-training model.
Although the magnitude of Wudao 3.0 parameters was not disclosed at the scene, Huang Tiejun told Titanium Media App that the amount of Wudao 3.0 parameters was lower than that of Wudao 2.0 1.75 trillion, and he stressed that the focus is not on the amount of parameters, but more in video and voice applications and system construction.
Specifically, according to Huang Tiejun, the Aquila base model provides two versions of 7 billion and 33 billion parameters, as well as the dialogue model AquilaChat based on these two versions. Huang Tiejun said that through architecture optimization and high-quality Chinese-English bilingual pre-training dataset, the model achieves better model performance with less data and shorter training time, and the training efficiency of the model is 24% higher than that of Meta AI's open-source LLaMA model.
Also released is the text generation code model AquilaCode-7B, which also supports Chinese and English bilingualism, with 7 billion training parameters. AquilaChat, on the other hand, has strong dialogue capabilities and extensible special instruction specifications, allowing users to achieve multi-tasking and tool embedding, such as AltDiffusion developed by KLCII and Stable Diffusion to achieve image generation in multiple languages.
"What we are doing now is something similar to Linux, Linux is not an open source ecology of an enterprise, Linux is everyone's," Huang Tiejun told Titanium Media App and others, now the large model is still in the early stage of technological development, too early attention to commercialization will restrict the development of technology, he called on the industry to invest in the ecological construction of the large model infrastructure. He emphasized that large models are a universal capability.
Lam Wing-wah, Vice President and Chief Engineer of KLCII, told Titanium Media App that KLCII is closer to learning and research, and then they hope to use evaluation, research and suggestions to build an iterable, looping and automated process "big model production line" to efficiently handle data cleaning, model training and automatic evaluation. She believes that large models are key to determining whether the technology can fall into the industry, especially the formation of high-quality large models.
"In addition, from my feelings, it is necessary to have the spirit of craftsmanship to create large models." Lin Yonghua said.
Huang Tiejun said at another sub-forum on June 9 that it is difficult for large-model industries to form monopolies, and it is necessary to build a closed loop of joint construction of industries. In the long run, the big model is just a label, not a product or a company's tool, so the entire big model open source ecology is inevitable.
Huang Tiejun stressed that in the future, the language model will continue to develop and become more and more capable. But more importantly, in the next 1-3 years, visual large models will also have huge new technological breakthroughs in the fields of automatic driving and robotics. The longer term is to achieve truly universal AI technology. (This article was first published on Titanium Media App, author | Lin Zhijia)