- What remains fundamentally hard in AI and computer science, and what are the emerging bottlenecks?
- What computing architectures will AI require beyond CPUs and GPUs?
- Are language and foundation models (LLMs, VLMs, VLAs, …) becoming a new computational substrate, or are they components within larger systems? Is language the new medium of general or AI-focused computing?
- How should software systems (languages, databases, operating systems, networking) change to support AI-based computing? Should they?
- What skill set is necessary for various career paths in this modern era of computer science? How should we teach it? Is it possible?
- How should humans and AI divide cognitive labor? As AI becomes increasingly capable, what should remain under human control?
- Does embodied AI represent a fundamental shift in computing, or an evolution of robotics, planning, control, and learning? And should computer science care beyond optimizing for energy use?
- Trustworthy, aligned, collaborative AI: What should remain under human control? What should be the division of cognitive labor between intelligent agents and their human users? Are intelligent agents peers, subordinates, or something else?
- How should we adjust the scientific process and its communication: Capable reviewers are already overtaxed, and journals and conferences are facing growing heaps of AI-generated slop. Evaluation benchmarks are rapidly saturating, and are not sufficiently distinguished in what we gain?
- What would Computer Science and Artificial Intelligence look like in 20 years? Would AI still be a part of computer science, or do the new undergraduate programs in AI prove to be the beginning of a historical split?
Additional topics listed below (list will expand as the program develops)
Physical, Emobdied AI
What do physical bodies bring to the table? How do they challenge Artificial Intelligence and Computer Science?
Classic (and still developing) technologies have produced functioning robots that operate in the real world: vacuum cleaners, hospital delivery robots, automated warehouses. These use motion planning, simultaneous localization and mapping (SLAM), and multi-agent sequential-decision making technologies.
Operating systems and compilers now consider power usage as part of their optimizations, scheduling, and device management. Job scheduling and allocations in clusters benefit from reinforcement learning and machine learning.
So what’s really new? What is missing? What is redundant?
- Where do we really need Vision Language Action models (VLAs) and Vision Language Models (VLMs)?
- Other than optimizing for energy usage, should general computer science care?
Splitting or Reformulating?
Traditionally, AI has been viewed as an area within computer science, giving rise to the old adage: “AI is everything we don’t understand yet in computer science”. This is still the common view. However, the recent rapid emergence of new undergraduate programs specializing in artificial intelligence within computer science departments is challenging this view, and has revived a deep and old discussion about the relationship between artificial intelligence and computer science.
This discussion is nicely captured by revisiting the Hebrew name for Computer Science, which is מדעי המחשב (an exact translation). The root for the word can be interpreted in two ways: מדעי החישוב (the science of Computing) or מדעי המחשבה (the sciecne of Thought).
Both elements have co-existed in the CS scientific community. Will they now split? Or perhaps AI should be reformulated differently within computer science?
Teaching Computer Science in the Age of AI
The undergraduate education in computer science has always been challenging. It has been challenging to filter and identify good computer science undergraduate students when they register. It has been challenging teaching them how to think like a computer scientist—in theory and in practice. The breadth of theoretical and empirical skills is difficult to manage and teach. In many cases, the ever-present (so far) market demand for computer science graduates has made things even more challenging, pressuring towards programming (rather than architecting), towards implementing (rather than inventing), and towards rapid delivery rather than optimized performance.
And then ChatGPT arrives. Now what?