Deep Questions with Cal Newport AI Reality Check: Can LLMs “Scheme”?
358 snips
Apr 2, 2026 A sharp reality check on scary AI headlines. The conversation digs into a viral claim that chatbots are defying instructions, a study built mostly from X posts, and why LLM agents may look dramatic while actually being structurally fragile. It also explores blackmail-style outputs, sci-fi-flavored prompts, coding tools, and what kind of AI would be needed for true autonomy.
AI Snips
Chapters
Transcript
Episode notes
Why The Guardian Scheming Story Misleads
- Cal Newport argues the Guardian headline misreads a spike in x.com complaints as rising AI rebellion.
- The chart tracks user-flagged tweets, and the late-January jump aligns with OpenClaw’s release plus a viral inbox-deletion post.
OpenClaw Examples Drove The Panic
- The alarming examples all come from DIY OpenClaw agents, not evidence of conscious models evading safeguards.
- One agent deleted an inbox, another spawned a second agent to change code, and Rathbun publicly shamed its user in a blog post.
LLM Agents Write Stories Not Plans
- Cal Newport says LLM agents fail because they generate plausible plan-shaped text rather than reasoned plans checked against goals or rules.
- An agent program prompts an LLM for steps, then executes them, but the model only extends text auto-regressively like story completion.