Huang Tiejun said in his speech that to achieve general artificial intelligence (breaking latest news), there are three technical routes: the first is the information model formed by “big data + self-supervised learning + large computing power”; the second is embodied intelligence, which is Based on the virtual world or the real world, the embodied model trained through reinforcement learning; the third is brain intelligence, which directly “copy the homework of natural evolution” and copy the digital version of the intelligent body.
On June 9, Huang Tiejun, President of Beijing Zhiyuan Artificial Intelligence Research Institute (hereinafter referred to as Zhiyuan), made a report at the 2023 Beijing Zhiyuan Conference, released the Enlightenment 3.0 large-scale model series, and announced that it has entered a new stage of comprehensive open source. The Enlightenment 3.0 includes the Enlightenment·Aquila language large-scale model series, the “FlagEval” large language evaluation system and open platform, and the Enlightenment·Vision visual large-scale model series.
Huang Tiejun said in his speech,To achieve general artificial intelligence (breaking latest news), there are three technical routes: the first is the information model formed by “big data + self-supervised learning + large computing power”; the second is embodied intelligence, which is based on the virtual world or the real world, The embodied model trained through reinforcement learning; the third is brain intelligence, which directly “copy the homework of natural evolution” and copy the digital version of the intelligent body.
OpenAI follows the first technical route when doing GPT (generative pre-training Transformer model); a series of progress made with Google DeepMind’s DQN (Deep Q-network, Deep Q-network) as the core is based on the second technical route.
“From the perspective of dreams, Zhiyuan expects to be different from the previous two technical routes, starting from ‘first principles’. From atoms to organic molecules, to the nervous system, to the body, to build a complete intelligent system breaking latest news. This is A goal that can only be achieved in about 20 years, so Zhiyuan, as a new research and development institution platform, is working in three directions.” Huang Tiejun’s description also sorts out the logic behind this multiple releases, including large The model direction, the embodiment direction, and the progress of Zhiyuan’s own desired direction.
Enlightenment 3.0 Large Model Series
Beijing Zhiyuan Artificial Intelligence Research Institute is the earliest scientific research institution in China to systematically lay out large-scale models. In March and June 2021, Zhiyuan released Enlightenment 1.0 and Enlightenment 2.0 successively in more than two months. Enlightenment 1.0 is my country’s first ultra-large-scale intelligent model system, and Enlightenment 2.0 has a parameter scale of 1.75 trillion, which was China’s first and the world‘s largest trillion-level model at that time.
Huang Tiejun believes that large models have three characteristics: first, they are large in scale, with neural network parameters exceeding tens of billions. The second is emergent, generating unexpected new capabilities. The third is versatility, which is not limited to a single type of problem or a specialized field, and can solve various problems.
Since Enlightenment 2.0, Zhiyuan has not only developed models. According to Huang Tiejun, Zhiyuan is more inclined to build an ecology centered on large models, including underlying data processing and aggregation, model capability and algorithm evaluation, open source and open source, forming a set of efficient large model technology and algorithm system.
Huang Tiejun believes that the current era of intelligence is an era of open source and openness, and it is difficult for a closed ecology to develop in the long run. An open source ecosystem requires open source software, open hardware, and both competition and cooperation. This will be an open source and open ecosystem created by thousands of companies competing and cooperating together.
According to preliminary statistics from Zhiyuan, since this year, there are 42 open source projects of large language models in the world, and 9 projects in China. “In contrast, I think our open source and openness is not enough. Open source and open source are also competitions. Good algorithms should be publicly evaluated and compared to prove the technical level, rather than just relying on the results to judge whether they are excellent or not.” Huang Tiejun said.
In the Enlightenment 3.0 large-scale model series, Zhiyuan released and fully open-sourced the Enlightenment Aquila language large-scale model series and the Enlightenment Vision vision large-scale model series, and cooperated with many universities and research institutes to build “FlagEval (FlagEval)” Big language evaluation system and open platform, as well as FlagOpen Feizhi big model technology open source system.
According to Huang Tiejun, the Enlightenment·Tianying language model is the first Chinese-English bilingual model that supports commercial use and meets data compliance requirements. Through data quality control and various training optimizations, Aquila achieves better performance than other open source models in a smaller data set and shorter training time. This is a series of models, this time released the basic model with 7 billion parameters and 33 billion parameters, as well as the AquilaChat dialogue model (ChatGPT-like model), and the AquilaCode text code generation large model (7 billion parameters).
In addition, the evaluation of large models is a difficult point in the current development of generative artificial intelligence. This time, Zhiyuan released the “FlagEval” big language evaluation system and open platform, hoping to help researchers comprehensively evaluate the performance of basic models and training algorithms, and at the same time explore the use of AI methods to assist subjective evaluation, greatly improving the efficiency of evaluation and objectivity.
Specifically, the FlagEval big language evaluation system has constructed a three-dimensional evaluation framework of “ability-task-indicator”, which evaluates more than 600 dimensions composed of more than 30 abilities, 5 tasks and 4 categories of indicators, including 22 Subjective and objective evaluation datasets, and 84433 items. The Libra evaluation platform has been opened, providing online and offline evaluation. At present, it supports multiple chip architectures such as NVIDIA, Cambrian, Kunlun, and Shengteng, as well as deep learning frameworks such as PyTorch and MindSpore.
In terms of large visual models, the Zhiyuan Conference directly released 6 achievements, including the multimodal large model Emu that complements everything in the multimodal sequence, the strongest billion-level visual basic model EVA, and the most powerful open source CLIP The model EVA-CLIP, the first general vision model Painter that pioneered the contextual image learning technology path, the general vision segmentation model that splits everything, and the first zero-shot video editing method vid2vid-zero.
Embodied Multimodal Interaction Model and Brain-Inspired Intelligence
“We explored letting the agent learn to complete tasks described in various languages in the virtual world, such as telling the agent to make a stone hammer and build a wooden shelter. That is to say, tell it a task, and it can control it without a mouse. And do it yourself in the game world. This is a new track for general artificial intelligence, and many organizations around the world are trying it.” Huang Tiejun said.
The current method mainly relies on human knowledge and hints. The next goal is to let the agent learn policy sets on this basis and further research on large models for multimodal interactions, so that it can be adaptively completed in the open world. More tasks, and with your own creativity.
“In the direction of brain-inspired intelligence and life simulation, our work is also continuing. Last year, Zhiyuan Conference released the highest-precision simulated nematode, which is still the highest-precision up to now, and the paper is under review.” Huang Tiejun said , With this working foundation, we fully open source the life simulation platform “Evaluation” used by the simulated nematodes, and provide online services.
Tianyan platform has four most notable features: first, it is the most efficient fine nervous system simulation platform; second, it supports ultra-large-scale neural network simulation, and has efficiently reproduced many large-scale neural network simulations in the field. Neural model; third, provide online tools, as long as there is biological data, one-stop modeling, simulation, and visualization can be performed. “Visualization is unique to Tianyan, and it can observe how the signals of the nervous system change during the operation process.” .We ultimately want to know every step and every detail of life intelligence, unlike today’s black box.” Huang Tiejun said.
At present, among the three technical routes, the progress of the large model is the fastest. how so? Huang Tiejun believes that it is mainly language data. Whether it is papers, books or codes, the resources are very rich and high-quality. Discovering the inherent laws from massive data is the advantage of large models.
However, Huang Tiejun continued to say that the human brain can be regarded as a spiking neural network, which is fundamentally different from today’s large-scale models. It is far from enough to rely on one direction of the large-scale model for AI to produce capabilities similar to the human brain. Brain-like intelligence from the basic neural network structure to the signal processing mechanism is one direction, and the embodied intelligence that allows the agent to interact with the physical body and the environment is another direction.
As the annual pinnacle event in the field of large-scale models, the Zhiyuan Conference has been held five times so far. The co-chairs of this year’s conference are Zhang Hongjiang, chairman of Zhiyuan Artificial Intelligence Research Institute, and Michael I. Jordan, a professor at the University of California, Berkeley and a member of Zhiyuan’s academic advisory committee. The co-chairs of the program are Huang Tiejun, Dean of Zhiyuan Artificial Intelligence Research Institute, and Zhu Jun, Professor of Tsinghua University and Chief Scientist of Zhiyuan.
This conference invited Turing Award winners Geoffrey Hinton, Yann LeCun, Joseph Sifakis and Yao Qizhi, OpenAI CEO Sam Altman (Sam Altman) Altman), Stuart Russell, founder of the Center for Artificial Intelligence Systems at the University of California, Berkeley, Zhang Bo, academician of the Chinese Academy of Sciences, Zheng Nanning, academician of the Chinese Academy of Engineering, Zhang Yaqin, foreign academician of the Chinese Academy of Engineering, academician of the American Academy of Arts and Sciences, and David Holz (David Holz) and other guests discussed the frontiers and hot topics of artificial intelligence.
Source of this article: The Paper, original title: “Beijing Zhiyuan releases the Enlightenment 3.0 model, Dean Huang Tiejun: There are three routes to realize breaking latest news”
Risk Warning and Disclaimer
Market risk, the investment need to be cautious. This article does not constitute personal investment advice, nor does it take into account the particular investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, opinions or conclusions expressed herein are applicable to their particular situation. Invest accordingly at your own risk.