The leaker who was just fired from OpenAI defected to Musk at the speed of light.
party Pavel Izmailov(hereinafter referred to as Little P), is one of Ilya's allies and worked in the super alignment team led by Ilya.
Advertisement
Half a month ago, Xiao P was fired for allegedly leaking Q*-related secrets. Although it is not clear what he leaked, it caused quite a stir at the time.
It was too late, but then it was too late. Now, on his Twitter profile, it has been written with great fanfare:
Researcher@xai
Not to mention that Lao Ma is quick to recruit people. In addition to Xiao P, many outstanding talents have been recruited by Musk recently.
Netizens who watched the fun exploded. Many people praised him and said he did a great job:
Advertisement
Some people are disgusted and feel that hiring someone who leaks confidential information is tantamount to picking up garbage.
And the recent performance of xai, including the release of Grok 1.5V, has greatly enhanced its sense of presence, which can’t help but make people sigh:
xAI will become a major player in the game, competing with OpenAI and Anthropic.
Hiring the leaker OpenAI fired
The thing is like this, a blogger who is extremely concerned about new things in the field of large models made a big discovery:
How many new employees are there at Musk’s xAI? ? ?
Moreover, the research direction of several of them has something to do with OpenAI's most mysterious Q* algorithm. It seems that Musk is the true believer of Q*.
Specifically, who have just chosen xAI to live in?
The most eye-catching thing is the little P we mentioned at the beginning. He is also a member of the CILVR team at New York University. He revealed that he will join Tandon CSE and Courant CS at New York University as an assistant professor in the fall of 2025.
Half a month ago, his personal page still read, “Engaging in large model inference work at OpenAI.”
Half a month later, things have changed.
But Xiao P’s tweet at the top has not changed. It is the first paper of the Super Alignment team, and Xiao P is the author of this paper.
The Super Alignment Team was formed in July last year. It is one of the three major security teams established by OpenAI to deal with security issues that may arise from large models on different time scales.
The Super Alignment Team is responsible for the distant future, laying the foundation for superhuman superintelligence security, by Ilya Sutskever and Jan Leike lead.
Speaking of which, although OpenAI seems to take security seriously, it is no secret that there are great differences within it regarding the security development of AI.
This disagreement is even considered to be the main reason for the drama on the OpenAI board of directors last November.
It is said on the Internet that Ilya Sutskever became the leader of the “coup” because he saw something that made him feel uneasy.
As for the super-aligned team led by Ilya, many members also stood on Ilya's side. In the subsequent Solitaire event supporting Ultraman, the members of the super-aligned team basically remained silent.
However, after the palace fight subsided, Ilya seemed to have evaporated from OpenAI, causing rumors to spread. However, he never appeared in public again, or even clarified or refuted the rumors online.
Therefore, we don’t know what the current situation of the super-aligned team is.
As a member of the Super Alignment team and a subordinate of Ilya, Little P graduated from OpenAI half a month ago, and was speculated by netizens to be Ultraman's “settlement of accounts”.
What a talent, he defected to Musk overnight
Although the outside world does not yet know the full picture of Q*, various signs indicate that it is committed to combining large models with reinforcement learning, search algorithms, etc. to enhance AI reasoning capabilities.
In addition to the most gossipy little P, the research directions of the following new talents who joined xAI are more or less related to it.
Qian Huangis currently a doctoral student at Stanford University.
Qian Huang has been working at Google DeepMind since last summer. He is currently identified as @xai on Twitter. It is not yet known what position he holds.
But as you can see from her GitHub personal homepage, her research direction isCombining machine reasoning with human reasoningespecially the rationality, interpretability and scalability of new knowledge.
Eric Zelikmana PhD student at Stanford, wrote “study why @xai” on Twitter.
Before that, he spent time at Google Research and Microsoft Research.
He said on his personal homepage: “I am fascinated by how (and whether) algorithms can learn meaningful representation and reasoning, and I am studying this at xAI.”
In March this year, his team launched the Quiet-Star algorithm, which is indeed Q*, allowing large models to learn to think independently.
Aman Madaana doctoral candidate at the Language Technology Institute at Carnegie Mellon University.
His research areas include large language models, feedback-driven generation, and the intersection of code generation and natural language reasoning. The primary topic of research is using inference-time compute to enhance reasoning capabilities (Reasoning).
During his PhD, Aman served as a student researcher and collaborator at Google Brain and the Allen Institute for Artificial Intelligence; earlier, he was a principal technical staff member at Oracle.
With the addition of several new employees, including Pavel Izmailov, Musk's technical talent pool has since expanded to 34 people (excluding Musk himself), which is about 2 times the original founding team of 12 people.
Among the newly joined members, there are a total of 7 Chinese people, plus 5 people from the founding team, there are 12 people in total since then.
Xiao Sun, who previously worked at Meta and IBM, graduated with a PhD from Yale and is an alumnus of Peking University.
Ting Chen, who previously worked at Google DeepMind and Google Brain, graduated from Beijing University of Posts and Telecommunications with a bachelor's degree.
Juntang Zhuang, who previously worked at OpenAI, is a core contributor of DALL-3 and GPT-4. He graduated from Tsinghua University with a bachelor's degree and Yale with a master's and doctoral degree.
Xuechen Li, who graduated with a Ph.D. from Stanford this year, is a core contributor to the Alpaca series of large models.
Lianmin Zheng, PhD in computer science at UC Berkeley, founder of Vicuna and Chatbot Arena.
Qian Huang is a doctoral candidate at Stanford and graduated from Tianjin Nankai Middle School.
Haotian Liu, University of Wisconsin-Madison, graduated from Zhejiang University with a bachelor's degree and is the first author of LLaVA.
From the perspective of institutional distribution, it is mainly concentrated in Google, Stanford, Meta, OpenAI, Microsoft and other institutions. They all have rich experience in large model training, such as the GPT series, the Big and Little Alpaca series, and Google Meta-related large models.
Judging from the time of joining, most of them were concentrated in February and March this year, with an average of one new member joining every 5 days, totaling 13 people. Only five people joined between August and October last year.
And if combined with the corresponding progress of Grok, we can also see Musk’s xAI recruitment plan for each stage.
For example, on March 29 this year, Musk suddenly released Grok-1.5, and the context length soared from the original 8192 to 128k, which is the same as GPT-4.
Pushing the time back one month (February this year), Juntang Zhuang, a former employee of OpenAI, joined xAI. He invented the GPT-4 Turbo algorithm at OpenAI that supports 128k long context capabilities.
And like April 15 this year,Grok-1.5V multi-modal model releasedit can handle various visual information including charts, screenshots, and photos in addition to text information.
In March this year, Haotian Liu, the first creator of LLaVA, just joined. LLaVA is an end-to-end training multi-modal large model that demonstrates capabilities similar to GPT-4V. The new version of LLaVA-1.5 achieves SoTA on 11 benchmarks.
So now you might as well boldly imagine what kind of new upgrades Grok might have with the introduction of a new wave of talents?
Netizen: No matter, where is Grok-1.5 (it’s not open source yet).
But no matter what, based on the talent standards previously mentioned by Musk, this netizen told the “truth”:
Everyone says that Lao Ma's big model company is full of talents. In fact, Musk doesn't care whether you are talented or not.
People say that as long as you can work 80 hours a week without breaking down, you can join them.
80 hours? ! If the qubits do some calculations, wouldn’t that mean working 11.5 hours a day, all week long…
Let alone IQ, I can't do this job physically.