DataFramed #358 How AI Agents Will Work While You Sleep | Ruslan Salakhutdinov, Professor at Carnegie Mellon
51 snips
May 4, 2026 
Ruslan Salakhutdinov, UPMC Professor at Carnegie Mellon and former AI research lead at Apple and Meta, joins to explore agentic AI. He discusses promising use cases like coding and web agents. They cover long‑horizon tasks, credit assignment and rubric‑based training. Conversation touches on multi‑agent designs, human‑in‑the‑loop workflows, safety guardrails, and lessons from self‑driving cars.
AI Snips
Chapters
Transcript
Episode notes
Agent Fixes Overnight Experiments
- CMU students run experiments overnight and invoke agents to patch failed runs, restart jobs, and save compute time.
- Ruslan uses this example to show a simple agent fixing segmentation faults so results are ready by morning.
Agents Can Automate Complex Job Searches
- Web-based 'computer use' agents can outperform humans on laborious information-gathering tasks like academic job searches.
- Ruslan imagines agents returning verified spreadsheets of suitable faculty openings with deadlines and fit notes.
Keep Humans In The Loop For Uncertainty
- Let agents work autonomously until uncertain, then pause and ask clarifying questions.
- Design interfaces where agents report what they found, what they're unsure about, and options for the user to verify.