Thoughts on AI

• AI for better science: Not really in HEP
• Production function: Improving rapidly / game changer already / Need to lean in
• AI for learning: Great, but you cant be lazy

As a compiler: As an increadibly technically competent, jounior research assistent

• Treat computer as a student…. precise well-defined problems… very with problems where you know what the answer should be "closure tests" Units test their main purpose from now on is refining” their AI agent’s work—but knowing enough about coding to fix it if it goes wrong.

• Collect bugs / gottas
• Collect solutions

• These are now reliable daily tools
• Models are capable of elite coding. "knitting—something people do “because they like it, not because it makes any sense.”"
• AI Like a new form of compiler

the production function for research

• Learn the tools aggressively or fall behind TB: Andy Hall: Five Pieces of Advice on AI for Polisci/PE Academics
• Why AI in course work

Even assuming little or no further progress in model capabilities, I think failing to figure out how to take full advantage of these tools would leave you at a major disadvantage…

Experimentation. Returns to experimentation are high: these are new tools and are evolving quickly, so I doubt anyone has "solved" the problem of how to best use them (for research or for any other tasks). My current sense is that I'm using these tools more aggressively than many PhD students are. Given that I'm (a) a tenured professor and (b) kinda old, this suggests to me that many students are currently under-investing in exploring what these tools can do for you. On a related note, the better models (often accessible only through institutional or individual subscriptions, not the open web) are often much, much better than the free versions. It's more than worth the trouble of getting access to at least see if you get value from them

Verification. The models produce output that tends to look/sound plausible but (at least for now) can be wrong in ways that require expertise to detect. Learning when and how to audit model output is a core skill for working with these tools. Note that if Case 2 happens, this could mean that the returns to have good taste, good instincts on problem framing and selection, etc., could go up rather than down