Key highlights from BAAI Conference, the Super Bowl for the AI industry
Founded in 2018, BAAI (stands for Beijing Institute of Artificial Intelligence) is a non-profit organization supported by the Ministry of Science and Technology, the Beijing Municipal Government, technology companies such as Xiaomi, Baidu, Megvii, and Meituan, as well as leading universities.
If you’re really into generative artificial intelligence (AI), you should know BAAI, the institution that Brad Smith, vice chairman and president of Microsoft, considers to be one of the leading players in this groundbreaking technological revolution.
In an interview with Nikkei Asia in April, Brad identified three entities at the forefront of the innovation, namely, Microsoft-backed OpenAI, Google-supported DeepMind, and BAAI, which stands for Beijing Institute of Artificial Intelligence.
Founded in 2018, BAAI is a non-profit organization supported by the Ministry of Science and Technology, the Beijing Municipal Government, technology companies such as Xiaomi, Baidu, Megvii, and Meituan, as well as leading universities.
While it’s kept a relatively low profile compared to its peers in terms of launching large language models (LLMs), you shouldn’t miss its annual conference, which brings together AI moguls from around the world to explore where new technologies are headed.
On June 9 and June 10, the annual conference was held, with Geoffrey Hinton, Yann LeCun, Joseph Sifakis, and Andrew Chi-Chih Yao — a squad of Turing Prize winners, as well as Sam Altman, the founder of Open AI, and David Holz, the founder of the popular text-to-image startup Midjourney, all in attendance. The lineup is pretty dope, by the way, much like the Super Bowl in AI.
We listed key takeaways from the big names for you:
Sam Altman attended the “AI Safety and Alignment” panel virtually on June 10 and had a Q&A session with Zhang Hongjiang, the Chairman of BAAI.
It’s the first time that Altman has given a speech in China. During his opening keynote for this session, the entrepreneur highlighted the urgency of AI regulation and the importance of international cooperation.
Altman acknowledges the impact of geopolitical conflicts on global AI cooperation, noting that despite differences, great powers have historically found ways to collaborate on key issues. “We see great potential for researchers in the US, China, and around the world to work together to achieve the shared goal,” he said.
He used an ancient Chinese proverb, “A journey of a thousand miles begins with a single step,” in describing how the world should start on the difficult but necessary cooperation.
Altman also disclosed that OpenAI is discussing opening source more models in the future, but does not have a specific model or timetable.
Yann LeCun, chief AI scientist at Meta AI, and one of the “godfathers” of AI and deep learning, delivered a keynote speech titled “Towards Machines that Can Learn, Reason, and Plan.”
He takes a slightly more critical stance in relation to the autoregressive type of language model represented by OpenAI’s GPT, pointing to the technology’s penchant for hallucination. LeCun also believes LLMs are far from being able to “reason and plan” on the order of human beings and other animals.
Getting machines to behave like humans and animals has been the quest of LeCun’s life, thus, he proposed that the “world model” could be an alternative approach in pushing AI-based systems to go beyond natural language processing capabilities to “perceive, reason, predict and plan”.
LeCun believes that the current concern that “AI will destroy humanity” is unnecessary because super-intelligent AI is not out yet.
Geoffrey Hinton has expressed his concerns about superintelligence. Yes, as always.
Like LeCun, Hinton is one of the godfathers of AI, but he is more pessimistic about the future of this technology. He quit Google in May and warned that the latest advances in AI posed “profound risks” to society and humanity.
Hinton’s talk was titled “Two Paths to Intelligence”, in which he discussed the concept of immortal computation represented by digital computers, and mortal computation represented by the human brain.
He argued that digital neural networks are not energy efficient, while biological brains consume much less energy. Therefore, the idea of a more power-efficient “brain” based on analog hardware was brought out.
Digital computation is immortal because it separates the software from the hardware, which makes it possible to run software on one computer that can be exactly replicated on any other digital computer. This is immortality: if the hardware dies, the algorithm can survive on another computer. On the other hand, mortal computation is a new way to achieve AI because it eliminates the separation of software and hardware. But the side effect is that when the particular hardware dies, the algorithm also disappears.
Toward the end of the speech, he emphasized his concerns about the threat posed by superintelligence and hopes that young talents will find ways to make superintelligence help humans live better, rather than letting them fall under the control of technology.
Max Tegmark, an AI ethics advocate who founded the Future of Life Institute, calls for “keeping AI under control”
Tegmark suggested that two key issues should be addressed in keeping AI under control, the first being alignment AI: getting an individual AI to actively do what its owner wants it to do, and the second being multiscale alignment: aligning all organizations, companies, and people around the world so that their motivation is to use AI for good.
When it comes to AI governance, he said China is doing the most to regulate AI, Europe is second, and the US ranks third.
Tegmark’s nonprofit called for a halt to research on “all AI systems more powerful than GPT-4, and in April organized an open letter titled “Pause Giant AI Experiments” and received more than 27,565 signatures (as of May 8).
In addition to bringing the AI industry leaders to the conference, BAAI also announced the latest progress on its super-scale AI modeling system Wudao:
Wudao Aquila, the first large-scale open-source language model that supports Chinese-English knowledge, commercial licensing, and complies with Chinese data regulations, was released.
According to the institution, Wudao Aquila, which has 7 billion and 33 billion parameters models, inherits the architectural design advantages of GPT-3 and LLaMA, replaced a batch of more efficient underlying operator implementations, and redesigned the tokenizer to support bilingual Chinese and English. (Note: LLaMA, or Large Language Model Meta AI, is an LLM released by Meta in February 2023. It has trained a variety of model sizes, with parameters ranging from 7 billion to 65 billion.)
Through data quality control and various training optimization methods, Aquila achieves better performance than other open-source models with smaller datasets and shorter training time.
Building on its in-house LLM, AquilaChat, a ChatGPT-style product, has come out. According to the agency’s internal review, AquilaChat is able to achieve a performance level of about 70% of GPT-4, and it even surpassed GPT-4 on some individual tasks.
It also introduced FlagEval, an evaluation toolkit for AI large foundation models. Also open-sourced, FlagEval supports multidimensional evaluation of base models in/across different modalities, such as NLP, audio, CV, and multimodal in the future. Overall, evaluation tools play a vital role in understanding LLM performance, addressing biases, and improving their reliability and effectiveness.
The recently launched Wudao Vision Visual Model Series revolutionizes computer vision, tackling tasks like unification, model scaling, and data efficiency. Its portfolio includes Emu, a large multimodal model that completes everything in multimodal sequences; EVA, a high-performing billion-scale visual representation model; EVA-CLIP, the top-performing open-source CLIP model; Painter, a universal vision model that pioneered the in-context visual learning; and vid2vid-zero, a zero-shot video editing technique that allows for simple prompt-based video editing.
All of these solutions are accessible under FlagOpen, an open-source ecosystem initiated by BAAI that aims to be the “Linux” in the LLMs domain.
Speaking about the current frenzy among Chinese companies to create their own LLMs, Huang Tiejun, President of BAAI, stressed the importance of prioritizing the development of an LLMs ecosystem in China rather than overwhelming the field with numerous teams solely focused on LLM research and development.
BAAI is poised to engage in intense competition with its international counterparts, leveraging its open-sourced, superscale AI modeling system. Echoing the sentiments by Smith, “Who’s ahead and who’s behind can change a bit from one part of the year to another, but one thing has been absolutely constant: the gap is almost always measured in months, not years. And so there’s this enormously competitive race to innovate.”
References:
https://hub.baai.ac.cn/view/27407
https://hub.baai.ac.cn/view/27403
