Conflicting views
The Zhongguancun National Independent Innovation Demonstration Zone Exhibition Center was crowded, and many people even simply stood in the aisle next to the venue to listen to the lecture, so that it was hard to believe that this was a purely academic conference, but it continued to appear on June 9 and 10.
Now in its fifth year, KLCII's AI conference, hosted by KLCII, has been positioned professionally from the start, so you won't see any commercial publicity and gimmick forums for advertisers, even if every company involved in the field of AI today is a big name.
All guests came with insight and perspective – in other words, dry goods.
The guests attending the KLCII conference alone were 4 Turing Award winners, including Geoffrey Hinton, Yann LeCun, Joseph Sifakis and Yao Zhizhi, in addition to Zhang Cymbal, Zheng Nanning, Xie Xiaoliang, Zhang Hongjiang, Zhang Yaqin, Stuart Russell, Max Tegmark, etc., each of whom is thundering in the field of artificial intelligence. Sam Altman, co-founder and CEO of OpenAI, also spoke at the AI Security and Alignment Forum on the morning of the 10th.
"The top event for AI insiders", KLCII Conference is positioned as a true name. The exchange of views between the bigwigs made the academic and intellectual temperament of KLCII more prominent.
Sam Altman, who was clearly the most sought-after star, attracted countless people with his keynote speech and subsequent one-on-one conversation with KLCII Chairman Zhang Hongjiang, who later expressed his gratitude for the invitation to the KLCII Conference on Twitter.
Sam Altman's speech focused on the field of AI security, calling for international coordination to deal with the potential threats caused by the rapid development of AI, Sam Altman believes that within a decade humans will have a very powerful AI system.
ChatGPT's global popularity gives Sam Altman's speech weight, but that doesn't mean everyone agrees with OpenAI's development path. In the opening speech on the morning of the 9th, Turing Award winner and Meta chief AI scientist Yann LeCun directly expressed almost the opposite view: autoregressive models do not have the ability to plan and reason, in order to reach the general artificial intelligence AGI, it should not only imitate the human brain at the neural level, but also refer to humans in the cognitive module, Yann LeCun's answer is the so-called world model.
Another big man who recently attracted attention because of his departure from Google, Turing Award winner Geoffrey Hinton, known as the father of deep learning, also gave his own views on the current development of AI in his speech - artificial neural networks will soon surpass real neural networks, and artificial intelligence may even be more dangerous and urgent to the world than climate change.
Such a clash of views is not uncommon in nearly 100 reports and roundtable discussions in the past two days, all of which revolve around the most cutting-edge propositions about artificial intelligence, the hottest is of course the big model, in addition to open source and security is also the focus of heated discussions.
You can't find a second institution in China that can organise such a high-profile conference, as KLCII President Huang Tiejun said: KLCII is the first choice for international cooperation in AI research in China.
So what has KLCII done about these hot topics?
A series of KLCII models with multiple arrows
At the opening ceremony, President Huang Tiejun announced KLCII's achievements in the past year. The Wudao 3.0 model series has entered a new stage of comprehensive open source. Including the Aquila language large model series, FlagEval open source large model evaluation system and open platform, and Wudao · Vision Vision Large Model Series.
Among them, the Aquila language model supports commercial licensing agreements to meet domestic data compliance needs. In addition to the base model, it also includes the AquilaChat conversation model and the AquilaCode (text-code) generation model. According to KLCII's Vice-President and Chief Engineer, Lam Wing-wah, AquilaChat-7B has surpassed mainstream open source models at home and abroad at the same level by combining the results of a variety of objective and subjective evaluations in Chinese and English (22 evaluation sets and more than 20,000 evaluation questions randomly selected). Under the Chinese subjective evaluation that integrates information analysis, cross-language understanding, discriminant evaluation, knowledge application, revision and editing, style generation, code generation, creative generation, safety and values, etc., AquilaChat-7B currently achieves about 70% of GPT-4 capabilities.
The FlagEval large model evaluation system and open platform are built to facilitate the all-round evaluation of the performance of basic models and training algorithms, and the ultimate goal is to achieve full coverage of basic models, pre-training algorithms and fine-tuning algorithms in four aspects: natural language processing, computer vision, audio and multi-modality.
FlagEval has built a three-dimensional evaluation framework of "ability-task-index", and there are currently more than 600 evaluation dimensions, including 22 evaluation datasets, with a total of 84433 questions.
The FlagEval large model evaluation system and open platform are built to facilitate the all-round evaluation of the performance of basic models and training algorithms, and the ultimate goal is to achieve full coverage of basic models, pre-training algorithms and fine-tuning algorithms in five aspects: natural language, computer vision, speech, multimodal and cognitive ability.
Enlightenment · The Vision Vision Large Model series systematically solves a series of bottlenecks in the current computer vision field, including task unification, model scale and data efficiency, including: multimodal large model EMU that complements everything in a multimodal sequence, the strongest billion-level visual basic model EVA, a universal segmentation model that splits everything in one way, Painter, the first general vision model of contextual image learning technology path, and the strongest open source CLIP model EVA-CLIP, Simple prompt is the vid2vid-zero zero-sample video editing technology for video editing.
The launch of these large models highlights the evolution and iteration pipeline of large models that KLCII is focused on, which allows large models to grow with the help of more data and more capabilities, continue to iterate, and upgrade rapidly, and ultimately both technical research teams and industrial development teams will benefit from the evolution of this iterative pipeline.
In addition to a series of new models released this time, KLCII also upgraded the FlagOpen large model technology open source system launched at the beginning of this year, from models to parallel acceleration technology, inference technology, to hardware evaluation and model evaluation, and finally data analysis, cleaning and annotation tools, KLCII FlagOpen platform aims to create an open source algorithm system and one-stop basic software platform that fully supports the development of large model technology.
In particular, KLCII has open-sourced COIG, the first large-scale, commercially available Chinese instruction dataset, with a total of 191,000 instruction data in the first phase, and the largest continuously updated Chinese multi-task instruction dataset in the second phase, integrating more than 1,800 massive open source datasets.
Observing the series of achievements launched by KLCII at this conference, there are not only various models, but also various tools related to the core ecology of the large model and the upstream and downstream of the industrial chain.
Although KLCII's large model has reached a fairly advanced level in terms of performance indicators, in Lam's eyes, this is not KLCII's most important mission.
"The pursuit of whom's model is bigger and stronger than everyone is not KLCII's mission. Our mission is more bottom-level, data processing technology, data aggregation, algorithm evaluation, model capability evaluation, and of course, open source. This kind of more basic work is what KLCII is doing and should be the only thing we are doing at the moment. Only what we are doing, we should stick to. Lin Yonghua said.
This sounds almost "idealistic", but KLCII has been synonymous with idealism since its inception.
Climbing the peak of artificial intelligence has always been a long-term war, and in the path planned by KLCII, it includes not only information intelligence represented by large models, but also embodied intelligence based on reinforcement learning and physical body, and brain-like intelligence modeled on the human brain according to neurobiology. For the first time, breakthroughs in large models show a possible path to AGI's general AI, but embodied and brain-like intelligence are equally noteworthy – who can guarantee that the next breakthrough won't come from them?
Compared with this supreme goal, which is full of idealism and long-termism, the pros and cons of a large model score seem small. KLCII has a bigger vision.
AI idealism
It is not enough to train a large model with a high score, but more importantly, the algorithm and technology used to train the large model.
Rapid iteration of technology will eventually make every model obsolete, but if a rich and deep scientific and technological soil is finally established, so that the flywheel of technology can rotate rapidly in it, and advanced algorithms are continuously introduced, linking the upstream and downstream of the entire industry chain to effectively reduce the cost of model training, and achieve a high degree of understanding and controllability of AI safety, then the prospects of artificial intelligence are bound to become more solid and powerful, compared with this grand vision, how many large models are launched. Or the question of the score of a specific large model is small.
This grand vision is exactly what KLCII pursues.
In March 2021, KLCII used the term "big model" for the first time, opening a new chapter in the development of AI. In a short period of time, the Great Model of Wudao was iterated to the third version.
As a platform-based, not-for-profit research institution, KLCII seeks to create an AI innovation ecosystem – big models are resource-intensive systems engineering, and if a technology is researched, validated and ultimately open sourced by KLCII, the entire industry will benefit.
KLCII does exactly that. Build a large model foundation to promote the scientific research and innovation of the entire large model through open source, and accelerate the industrial landing of large models. For this vision, KLCII has even done many things that seem to be "thankless" to the outside world, and even reluctant to seek the near and far.
For example, KLCII's large model adopts a commercial licensing agreement, and since KLCII has spent huge resources from algorithms to data to achieve full compliance, enterprises can safely adopt KLCII's model for commercialization.
As we all know, there are many controversies in the copyright field at present, one of the reasons is that the data used for training comes from mixed sources, and KLCII's underlying compliance capabilities have helped companies avoid this biggest risk point, and the positive impact is self-evident.
For example, many open source projects released by KLCII, from the AquilaCode-7B generation model to the FlagEval open evaluation platform, support NVIDIA and domestic Cambrian, Kunlun and other chip architectures, which means that developers do not even need to adjust their own hardware devices to use them directly, KLCII promotes the development and innovation of the chip field through open source of a variety of architecture codes and models.
The new Linux ecosystem that builds big models for the next decade of AI is KLCII's positioning in the field of big models, and open source is a very important and courageous step in the field, and it is also a step that disagrees with many in the industry. However, KLCII's confidence is quite firm, open source is not only an inevitable choice for the construction of artificial intelligence ecosystem, but also the only way to promote technological innovation and comprehensive industrial upgrading.
KLCII has been on this path for five years, and this romantic vision of artificial intelligence attracts equally idealistic talent. As the top AI research institute in China, KLCII has a line-up of nearly 100 top AI experts, and the KLCII community brings together more than 120,000 AI professionals. , which has made KLCII's research prowess recognised worldwide.
In the past four years, more than 500 top AI experts represented by Turing Award winners have spoken and participated in discussions at the conference, and tens of thousands of professionals from more than 30 countries have registered to participate.
KLCII will release the latest achievements in related fields at the conference, and the exchange of views on artificial intelligence has also spread from Beijing to the world, and the platform built around KLCII has also taken shape, which has provided the strongest support for China's participation in the big model war, and Chicii is one of the perhaps unknown but extremely important "hidden pillars" of China's rapid development of the big model business.
This is not the best reward for idealism.