Yi ZENG (@yi_zeng) header
Yi ZENG (@yi_zeng) profile picture

Wu Yuzhang Chair Professor, Gaoling School of AI, Renmin University of China (RUC). Founding Dean, Beijing Institute of AI Safety & Governance (Beijing-AISI).

Beijing
Joined August 2010
Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)Photo by Yi ZENG (@yi_zeng)
Proud to release "ForesightSafety Bench" AI Safety Evaluation Framework, beginning with 7 major Fundamental Safety pillars and progressively extends to advanced Embodied AI Safety, AI4Science Safety, Social and Environmental AI risks, Catastrophic and Existential Risks, as well as 8 critical industrial safety domains, forming a total of 94 refined risk dimensions. 22 LLMs have been evaluated based on ForesightSafety Bench at this point. arxiv.org/abs/2602.14135v1
1
1
9
1,084
Will be very looking forward to the discussions with @tegmark and colleagues on AI existential risks & international cooperation, chaired by Senator @BernieSanders
Senator @BernieSanders has invited me and three other AI researchers to a public panel on AI existential risk & international cooperation at the U.S. Capitol 7pm Wednesday April 29th. RSVP here to join us for this important conversation: forms.office.com/Pages/Respo…
1
1
2
24
4,447
Yi ZENG retweeted
Senator @BernieSanders has invited me and three other AI researchers to a public panel on AI existential risk & international cooperation at the U.S. Capitol 7pm Wednesday April 29th. RSVP here to join us for this important conversation: forms.office.com/Pages/Respo…
54
37
20
411
233,598
Introducing ForesightSafety Bench: the Beijing-AISI has systematically developed the 'ForesightSafety Bench', a comprehensive, multi-dimensional AI Safety Evaluation Benchmark, with 3 layers of safety pillars, containing 94 dimentions at this point, and keep evolving. foresightsafety-bench.beijin…
2
1
14
1,160
Until now, we do NOT have scientific evidence and practical way to keep Superintelligence safe enough NOT to bring catastrophic risks for human. The World has NOT been ready to welcome Superintelligence NOT as a tool that is controlable at this point. I am showing my strong support for "prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in". superintelligence-statement.… @tegmark
11
4
4
58
2,649
My endorsment to red-lines.ai : AI Red Lines need unity of knowing and acting, both from human and from AI, and it takes the World to save ourselves and our next generations.
1
11
598
For countries, in general, the level of AI Governance are positive relevant to their GDP per capita. This is very risky, since that means the level of preparedness for AI risks for relatively lower countries are also very low and very easy to be negatively influenced. The development level of AI technology may vary, while the level of AI Governance need to be more resilient for all countries to avoid potential (catastrophic) risks. See AGILE index 2025 for details agile-index.ai/AGILE-Index-R…
2
6
602
Announcing the release of the AI Governance International Evaluation Index 2025 (AGILE Index 2025). agile-index.ai/AGILE-Index-R… "countries left behind need to raise the level of preparedness for AI Governance to be more resilient." (1) 40 evaluated countries can be categorized into three tiers based on their AGILE Index 2025 Scores, with AI Development Level (Pillar 1) and AI Governance Instruments (Pillar 3) showing relatively wide disparities across countries. (Page 14) (2) Among the 14 countries from AGILE 2024, the ranking changes for 2025 reveal a pattern of dynamic shifts among top-tier countries and relative stability among lower-ranked ones, with notable changes such as the position swap between the US and China. (Page 16) (3) There remains a generally positive correlation between the AGILE Index total score and GDP per capita across 40 evaluated countries, which is not good, since countries left behind need to raise the level of preparedness for AI Governance to be more resilient. (Page 16) (4) The performance of the 40 countries across the four pillars of AGILE Index reveals four distinct types of AI development and governance status: All-round Leaders, Governance Overachievers, Governance Shortfallers, and Foundation Seekers. (Page 17) (5) High-income countries group show a clear advantage over upper-middle and lower-middle-income countries group in both AI Development Level (Pillar 1) and AI Governance Instruments (Pillar 3), while the latter tend to perform slightly better with lower AI risk exposure and higher social acceptance of AI. (Page 19)
2
9
691
It is a wonderful experience talking to Daniel in his interview series. My thinking chases Daniel to think about even deeper questions that chases myself to think and react. On AI Safety, Catastrophic Risks from AI. More importantly, Positive Future, and our potential efforts (inspired by both eastern and western thinking), as human species and very possibly together with AI.
Is it possible to build an AI Confucius? One of China's greatest AGI thinkers joins me on The Trajectory. 18 months ago I asked Jaan Tallinn who I should talk to about AGI, and he mentioned Chinese researcher and thinker @yi_zeng. Today, at last, my episode with Yi (recorded recently in Paris) is finally live. Yi Zeng runs the equivalent of Beijing's AI Safety Institute in addition to being a Professor at the Chinese Academy of Sciences (at the "Institute of Automation", which also just sounds really cool). I became VERY aware over the course of One (of many) interesting take-away from my conversation with Yi: 1. He discusses the different cultural "frames" through which humans tend to see AGI, which have different prevalences in the east and west: A) Tool - AGI as a glorified smartphone, inert without the prompting and direction of a human. B) Partner - AGI as a part of society and helper. Yi mentions that this view is already prevalent in Japan. C) Enemy - Scifi films frame AGI as Yi believes that long-term we AGI won't be likely to stay a tool, and he's optimistic (admittedly more than I am) that AGI can become a partner to humans - even as its values expand and grow. Yi is very frank about AGI being a "becoming" more than just a "being", he suspects that its powers and even its values will evolve continuously - and ideally he sees humans benefitting from this growth. He mentions that when he's speaking with his Chinese-born students, they tend to more naturally see "Partner" as the frame to view AGI, while his international students don't tend to see "Partner" as AGI's likely role. ... Time stamps: 00:02:40 -- The difference between programming AGI with rules, and building "moral reasoning" 00:25:02 -- Yi explains his Tool, Partner, Enemy framing for AGI (worth listening to) 00:37:30 -- I ask Yi is AGI might be a Successor rather than a Partner (I express my skepticism about AGI eternally serving humanity), and he lays out why he suspects that human brain augmentation might prevent any kind of complete takeover from purely non-biological intelligence. ... Throughout our conversation he references eastern and western philosophers and thinkers - but all in all I walked away from this conversation thinking a LOT about how many of my own moral suppositions were somewhat distinctly "Western." I had tons of fun diving into Yi's very cosmic idea of the future and I hope you like this one. Track 2 dialogue with great Chinese thinkers is, IMO, even more important given all the arms race tensions we have going on at present. I hope to find some of Yi's best peers to speak with soon too. (I'll have links in the tweet below)
2
5
489
The AI Safety Institute (AISI)'s counterpart in China is currently the China AI Development and Safety Network (China AISI Network, ai-development-and-safety-ne…) which officially announced on September 25th, 2024 at the High-level Meeting on International Cooperation on Capacity-building of Artificial Intelligence (co-hosted by China and Zambia), at the headquarters of UN in New York, during the 2024 UN General Assembly. The recording from UN Web TV is available at (around 42:24) webtv.un.org/en/asset/k18/k1…
3
13
1
72
6,293
Announcing the release of Global Index on AI Safety. Evaluate 40 countries on readiness for AI Safety at the national level through six pillars: (1) Governance Environment for AI Safety, (2) National Institutions Targeting AI Safety, (3) Governance Instruments for AI Safety, (4) Research Status on AI Safety, (5) International Participation on AI Safety, and Existential Safety Preemption. Download the report here agile-index.ai/global-index-…
4
17
6
78
18,100
The efforts of AI Safety Summit, Scientific Reports on AI Safety, and various network of AISIs need to be incorporated and transformed into the AI Safety Process, to keep the continuity, effectiveness, and inclusivity. The AI Safety Process should not be influenced by the change of governments, hosts, and should be for good and for all. ai-ethics-and-governance.ins…
3
2
36
2,985
We do not need AI everywhere. Use AI proportionally and avoid wild adoptions. Thanks for the interview from Science and Technology Daily.
How can humans coexist harmoniously with #AI? Zeng Yi, a professor at the Institute of Automation, Chinese Academy of Sciences, and an expert of the United Nations' high-level advisory body on AI, shared his insights in the program #TechWorld.
1
1
17
1,232
I am wondering what AI can learn from these human generated advertisements. May be fire those who made and agree to use this advertisement? AI is still an information processing tool, which still make mistakes human will never make. AI cannot and should not replace human.
6
462
It is not only about a racial bias remark and inclusivity. For any reason, if you do not find credible citations, as a scientist? even ask AI service within seconds (as attached) will help you get initial but much better understanding of what you do not know. By the way, in China, it starts at least after 1981 with "Morality" in the course name, and now the course is titled "Morality and the Rule of Law", starting from year 1 as primary school student (covering all 12 semester, 1 new book every semester) to the whole middle school education. I am also attaching two pictures of the book covers from the Web, one from 1980s, and one from recent versions @RosalindPicard @NeurIPSConf @drjingjing2026
1/3 Today, an anecdote shared by an invited speaker at #NeurIPS2024 left many Chinese scholars, myself included, feeling uncomfortable. As a community, I believe we should take a moment to reflect on why such remarks in public discourse can be offensive and harmful.
1
687
There are many ways AI could go wrong to create (catastrophic) risks for human, let along misuse and abuse. Good memory for my first participation in Paris Peace Forum, and thank you for the interview. @ParisPeaceForum
"We have to be very cautious in the way we use AI to assist decison making - AI should never ever be used to control nuclear weapons, AI should not be used for lethal autonomous weapons." @yi_zeng, Founding Director of Beijing Institute of AI Safety and Governance #PPF2024
1
3
21
968
Glad to be invited by Paris Peace Forum @ParisPeaceForum on a Spotlight event "AI: An anthropological revolution?" With such an important topic and limited time, the discussion is only a beginning. Watch it here youtube.com/watch?v=7naUe4dZ… – at Paris, France
1
3
669
UK AI Safety Institute (UK AISI)'s visit to the Beijing Institute of AI Safety and Governance (Beijing-AISI) locally at the West Beijing AI Valley. Beijing-AISI's first group of face to face international visitors since its establishment last month. Beijing-AISI introduced several new research we did this month. It takes the world to keep AI safe, at our best, for all.
1
11
7
92
9,103