chenhaotan.bsky.social
Associate professor at the University of Chicago. Visiting scientist at Abridge. Working on human-centered AI, NLP, CSS. https://chenhaot.com, https://substack.com/@cichicago
181 posts
3,866 followers
265 following
Regular Contributor
Active Commenter
comment in response to
post
Huan says this is an example of how our society is not ready. Personal opinion, the person who creates the content with AI should be the owner.
comment in response to
post
Audience question: ownership of data
comment in response to
post
Not slowing down to the panel’s knowledge.
comment in response to
post
Audience question: postdoc hiring status.
comment in response to
post
Tuomas talks about time scale. Fundamentality comes from something that takes a longer time frame to solve and more difficult questions.
comment in response to
post
Huan draws the analogy with computer architecture. The best model may be built by industry, but many things can be done in industry that do not require building models from scratch.
comment in response to
post
Haifeng returns to the challenge of academia vs industry. What questions are good for academia?
comment in response to
post
Both are good questions, the third thing is science without insights. For example, hardness about combinatorial auctioning says that there is no easy takeaway about auctioning multiple items. What is the future of science if the theory is not comprehensible to humans?
comment in response to
post
Audience question, insights from AI or insights for AI, which one is more important right now vs in 2050?
comment in response to
post
Tuomas still has faith about academia about saying what is true or what is not true.
comment in response to
post
Audience question, a lot of hype driven by industry and the funding might go away after the hype. What can the academia do? Academia are relatively silent and are not heard in the media recently.
comment in response to
post
Huan adds that the first couple of years might be challenging, but being actively learning new things can be quite helpful.
comment in response to
post
Audience question on transitioning to machine learning, from image processing to AI.
comment in response to
post
Tuomas reflects on changes to the AI field, statistics moving into AI, game theory, optimization. What I like about AI is how AI is pushing the boundaries in other disciplines and has fast publication cycle. a little worried that people associate AI with GenAI
comment in response to
post
Huan adds that Fundamentals have not really changed, but one needs to learn more about integration with AI.
comment in response to
post
Audience question: what is the fundamental knowledge for the AI-based future?
comment in response to
post
Chaowei spent two years on cybersecurity and then switched to AI safety. And started from the question of why is it lp bounded space and then proposed the first non-lp bounded adversarial examples.
comment in response to
post
Huan adds that for senior PhD students, it would be helpful to have one focus and do that very well. For junior students reading papers and proceedings is a good starting point. He shares three factors with students: impact, novelty, depth.
comment in response to
post
Tuomas says there are many ways to be successful. If we look at people successful for a long time, they usually have an application. A good way to do is to have an application. The application will drive new questions.
comment in response to
post
Haifeng says the current pace is fast and changes the paradigm, for example, how CoT becomes influential is different from his experience. Any recommendations for junior researchers?
comment in response to
post
Chaowei finds it hard to even predict what his group will work on in 2-3 years. He hopes that AI will get great at software engineering tasks, physical embodiment, and scientific domains. There are many security and safety problems to regulate in these domains. It not only requires tech expertise.
comment in response to
post
Huan agrees LLMs may not be final solution. We will likely figure out better solutions in 25 years. Our society is not ready yet.
comment in response to
post
We need to use the right AI for the right problems. There should be limits.
comment in response to
post
Tuomas has had the pleasure to work with superhuman AI, but not AGI, for two decades. Humans have a lot of overconfidence in human decisions and AI can help in many ways. But AGI is not possible. Examples include recommendation systems and strategic games.
comment in response to
post
Haifeng asks about predictions on what AI can do in 2050 and what we need to do to make sure AI is beneficial.
comment in response to
post
Chaowei Xiao from Wisconsin spent his entire career in Midwest and works on AI safety from the security perspective.
comment in response to
post
Huan Zhang from UIUC introduced him starting PhD in computer architecture and how he transitioned to machine learning in the fourth of his PhD. He works on trustworthy machine learning both from small and big perspectives, robustness of foundational models, formal guarantees of small models
comment in response to
post
Tuomas Sandholm introduced himself and his active effort in solving imperfect games, kidney exchange, and a new initiative on heart transplant, auction, automated algorithm design.
comment in response to
post
Haifeng started by introducing the challenges in the current climate and fast-moving AI industry.
comment in response to
post
comment in response to
post
Considerations for whether to deploy Glaze or not
comment in response to
post
Creators are required to think about regulation.
comment in response to
post
Congratulations!