At 4 a.m. Beijing time, Musk led the xAI founding team to complete the first voice conference on Twitter Spaces.
As an early investment institution that has always adhered to the investment philosophy, we studied the growth experience of the xAI founding team members, especially the 4 Chinese, in an attempt to restore the whole process of the formation of the xAI founding team. Hope to inspire you, and welcome to communicate with us in the comment area.
In the just-concluded first Twitter Spaces voice broadcast of xAI, Elon Musk said:
- xAI will explore the nature of the universe, using artificial intelligence to generate deeper questions such as dark matter, the Fermi paradox, aliens, and give answers
- Confirm xAI's competition with OpenAI and Google
- The "most curious" artificial intelligence will be trained with Twitter public data
- xAI will develop technology that "understands the physical world, not just the internet," and Tesla's driving data may help
Two days ago, Musk announced on Twitter that the artificial intelligence company xAI was officially established - in order to understand the true nature of the universe.
The xAI official website was launched simultaneously. Occupying the largest page of the official website is the founding team of 12 people, including Musk. Members come from DeepMind, OpenAI, Google, Microsoft, Tesla, and the University of Toronto. Many breakthroughs in AI have come from their work, such as AlphaStar, AlphaCode, Inception, Minerva, GPT-3.5, and GPT-4.
"xAI has an all-star founding team. The density of talent is impressive, I've read so many of their papers that I can't count them." Linxi Jim Fan, a senior researcher at NVIDIA and Ph.D. in deep learning at Stanford, commented on LinkedIn.
Founding team member Igor Babuschkin said at the launch that xAI's team will remain small.
"In all the projects I've worked on before, I've found that GPU per capita is a key indicator of a project's success. We want to keep a relatively small team where the best minds have the freedom and lots of resources to try and implement their ideas."
01
The horn sounded
In early March, foreign media discovered that Musk registered a company called xAI in Nevada. In the same month, the news of poaching came out.
Long before xAI registration, Musk's first olive branch was handed to Igor Babushkin of DeepMind.
Since 2017, Babushkin has worked as a research engineer at DeepMind. Nearly four years later, Babushkin chose to leave and join OpenAI as a technician. He then returned to DeepMind in 2022 as a senior research engineer.
While at DeepMind, he and his colleagues developed the well-known StarCraft AI model, AlphaStar. AlphaStar learned from 500,000 StarCraft 2 rounds and then played 120 million rounds to refine its skills. In the end, it reached the highest level of Grandmaster, surpassing 99.8% of players.
Another tech talent Musk recruited from DeepMind is Manuel Kroiss. Cross previously worked as a software engineer at DeepMind. In fact, Musk offered him the position at the time as Twitter's senior director of software engineering.
In the following April, according to LinkedIn information, Wu Yuhuai joined. In May, Zhang Guodong and Christian Szegedy joined. In June, Toby Pohlen joined.
In July, xAI went live.
Pollenbun studied computer science at RWTH Aachen University in Germany and joined DeepMind in 2017 at the same time as Babushkin to develop AlphaStar. Pollan has also led the Large Language Model Project. After 6 years and 3 months at DeepMind, he joined xAI.
Most of the 12 have worked at DeepMind and OpenAI. Together with Microsoft, Google, and research institutions, it constitutes the treasure trove of AI talent that Musk craves.
But when Musk tried to invite multiple OpenAI members to join, he received quite a few refusals. Because the team is still in its infancy, although the imagination is large, the appeal is still limited at this stage.
Ross Nordeen's experience is the special one of the 12. Since 2020, he has served as Tesla's technical program manager, responsible for supercomputing, machine learning, and deep learning infrastructure.
He graduated from Michigan Technological University in 2012 with a bachelor's degree in computer networks and systems management and has worked for Whamcloud, Palantir Technologies and Handshake.
Prior to joining xAI, Musk brought Notting from Tesla to Twitter last October to manage recruitment and access.
Currently, xAI is continuously recruiting engineers and researchers in the Bay Area.
However, getting into xAI is not easy. A unique recruitment criterion is that candidates must pass unanimously 12 people to join the company.
02
Chinese gather
Of the 12, one-third were Chinese: Zihang Dai, Guodong Zhang, Greg Yang and Yuhuai Tony Wu.
Dai Zihang Dai
In 2019, Carnegie Mellon University (CMU) and Google Brain released XLNet, a pre-trained language model, that comprehensively surpassed BERT on 20 tasks.
The common work of this famous paper, one is Yang Zhilin, the other is Dai Zihang.
As one of the top AI scholars in China, Yang Zhilin entered Tsinghua University two years later than Dai Zihang, and when he graduated and entered CMU Computer School for a Ph.D., Dai Zihang was also at CMU. In April this year, Yang Zhilin founded an AI large-model startup "The Dark Side of the Moon", whose main business is AGI, and ZhenFund is its angel round investor.
Yang Zhilin said that academic discussions with friends are a good way to generate ideas, "For example, I will often discuss ideas and write papers with my friend Dai Zihang, and the best current language model architecture Transformer-XL is the result of cooperation."
Previously, a domestic star model company invited Dai Self-Aviation to join, but he finally chose xAI.
Dai Zihang studied at Chongqing Nankai High School and entered Tsinghua School of Economics and Management in 2009, majoring in information management and information systems. When he first entered school, he wrote that he was "happy to find that he was not intimidated by the irresistible pressure, and at the same time was always reminded by the environment that he had to fight with all his strength."
The four-year program combines mathematical foundations, computers, management and economics. After consulting internships at Accenture and Roland Berger, he joined the internship in the same month as NetEase Capital was founded, responsible for early-stage Internet enterprise investment.
If the three internships at the university level still belong to the "typical route of Tsinghua economics and management", by the time he graduated from his undergraduate degree in 2013, Dai Zihang completed his transformation to the field of machine learning.
At that time, it was less than a year after Yu Kai represented Baidu in the auction on the shores of Lake Tahoe, California, for Geoff Hinton, the winner of the Turing Award and the "father of deep learning". Driven by the failure of the auction, in early 2013, Baidu announced the establishment of a deep learning research institute (IDL) to recruit top talents at full speed, with Robin Li as the dean and Yu Kai as the executive vice president.
That summer, he joined IDL as an engineer intern, focusing on Baidu's image recognition algorithm, covering content-based image retrieval CBIR, Fisher Vector, and measurement learning. He also works on deep structure RankNet, image semantic genome research, and sometimes product design.
A year later, he left Baidu and went to CMU to begin a six-year master's and doctoral career in computer science, studying under Yiming Yang. After the Mila Lab, founded by Turing Award winner Yoshua Bengio, the Google Brain team, after graduating with a PhD in 2020, Dai officially joined Google Brain as a research scientist, mainly focusing on natural language processing and pre-training.
Guodong Zhang
Zhang Guodong, another member who is an undergraduate student in China, once said that if he were an athlete, he would be a long-distance runner because he never stopped running.
In 2013, he entered the School of Information and Electronic Engineering, Zhejiang University, majoring in information engineering, and minored in the Advanced Honor Class of Engineering Education (ACEE) of Zhu Kezhen College. Among the 182 people in the whole major, it ranked first in the major for three consecutive years.
From wireless communication in freshman year, to computer vision in sophomore year, to the intersection of statistics and deep learning at the University of California, Los Angeles (UCLA) in the summer of his junior year, under the tutelage of Zhu Songchun and Wu Yingnian, and then to advanced visual deep learning in his senior year, Zhang Guodong has long clarified his scientific research ambitions.
After graduating from his bachelor's degree, Zhang Guodong went to the Department of Computer Science of the University of Toronto in Canada and obtained a doctorate degree in machine learning. During his Ph.D., he published top papers in the fields of multi-agent optimization and application, deep learning, and Bayesian deep learning.
The University of Toronto is where Jeff Hinton teaches. During his Ph.D., Zhang Guodong also worked as an intern on the Google Brain team under the guidance of Hinton, working on large-scale optimization and fast-weights linear attention.
In 2022, Dr. Zhang Guodong graduated and joined DeepMind full-time, as a core member of the Gemini program, a competitor to GPT-4, responsible for training and fine-tuning large language models.
In May 2023, Zhang Guodong joined xAI, becoming one in twelf.
Greg Yang
Before the age of 12, Young grew up in China. He was born in Hunan, and his mother would buy him math Olympiad books when he was in elementary school, participate in math competitions after coming to the United States, and enter the mathematics department of Harvard University, as if "I have been learning mathematics since I was a child, like a hamster stepping on a wheel."
At the end of Younger's sophomore year, his obsession with electronic music led him to take a break from school to pursue his musical dream of becoming a DJ. While making music, I read artificial intelligence, quantum mechanics, physics, mathematics. But as he read, Younger found that he was spending more and more time on artificial intelligence and less and less time making music.
In a year and a half, Yange figured out several things:
One, he wanted to make AGI a reality. "Making something much smarter than you sounds good." It was still 2012.
Second, he loved mathematics.
Third, mathematics is the most essential language of all sciences, which will promote the progress of AI and more disciplines.
After returning to Harvard for a semester of classes, Young chose to take a second break for two years, quickly teaching himself all branches of mathematics by reading.
When he returned to Harvard, Yau became his academic mentor. Chengtong Yau took Yang Ge to attend the event, met doctoral students and mathematicians in various directions, and recommended him to apply for the highest honor that undergraduate students in mathematics can achieve: the Morgan Prize.
After earning a bachelor's degree in mathematics and a master's degree in computer science from Harvard, KLCII reported that Yau and Michael Freedman recommended Yanger to Shen Xiangyang, Microsoft's then-executive vice president.
A week later, Young received an offer to join Microsoft Raymond Research as a researcher, including Tensor Programs, neural networks and machine learning.
Shen Xiangyang said to him, "There are two Fields Medal winners who recommend you, it would be too silly for me to refuse again."
After Musk announced the official establishment of xAI, Young tweeted:
"The mathematics of deep learning is esoteric, beautiful and efficient. Developing a "theory of everything" for large neural networks is central to pushing AI to the next level.
At the same time, AI will enable everyone to understand our mathematical universe in ways that were previously unimaginable."
Yuhuai (Tony) Wu
Wu Yuhuai attended Fredericton High School in Canada, and studied mathematics at the University of New Brunswick from 2013 to 2015.
In 2015, Wu entered the University of Toronto to pursue a PhD in machine learning, where he studied under Roger Grosse and Jimmy Ba.
During his studies, Wu worked as a researcher at Mila, OpenAI, DeepMind, and Google. He was part of the Google N2Formal team under Christian Szegedy.
Wu Yuhuai wrote on his personal website, "My main research interest is to create a machine that is good at reasoning. I chose mathematics as a starting point for my research reasoning, with the goal of creating an automated mathematician."
In a paper co-authored by Wu, the researchers trained Minerva, an augmented large language model, in an attempt to teach Minerva to solve natural language math problems step by step. Minerva has a strong mathematical ability and in Poland's 2022 State Mathematics Exam, answered 65% of the questions correctly.
But according to Heart of the Machine, such models can only be imitated, and cannot be independently trained to improve mathematical level. Formal proof systems provide a training environment, but there is little data involved, so we need automatic formalization as a bridge to natural language mathematics.
Before joining xAI in April 2023, Wu did postdoctoral research at Stanford University.
Jimmy Ba was Wu's mentor, and he studied with Jeffrey Hinton during his PhD. After completing his master's and doctoral studies in electrical and computer engineering at the University of Toronto, he became an assistant professor in the Department of Computer Science at the University of Toronto, where he was one of the authors of the optimizer Adam. His research interests focus on efficient learning algorithms for deep neural networks, and the long-term goal is to answer the question of how to build a machine that can solve a general problem with the same efficiency and adaptability as a human?
Another member of the founding team, Christian Segdi, is the leader of Wu Yuhuai's Google N2Formal team and the oldest in the team. In 2005, Segedi received his PhD in Applied Mathematics from the University of Bonn. In 2010, Segedi joined Google. In the following 13 years, he devoted himself to deep learning, artificial intelligence, computer vision and other fields. On Google Scholar, Szegedi has nearly 220,000 citations.
03
The Answer
xAI is probably the company with the most confidence to understand the true nature of the universe as its goal, crazy, romantic, and born. There is money, and cards.
xAI is also indeed the company with the most GPUs per capita.
In April, Musk bought about 10,000 GPUs, saying they would be used to advance a new AI project on Twitter — perhaps the precursor to xAI.
But most importantly, someone. In this long journey to question the nature of the universe, talent is the key.