LessWrong Sequence Posts Discussed in this Episode:
on hiatus, returning someday
The Bayesian Conspiracy 259 – Chill Out on AI in 2028?
Apr 1, 2026
Casual debate about whether we should 'chill out' on AI alarmism in 2028, exploring tone, persuasion, and communication pitfalls. They weigh capabilities trajectories, singleton versus multipolar risk views, and likely real-world warning shots. Conversations cover human-guided AI creativity, an UnSlop fiction contest, fundraising dynamics, sustainable motivation for safety work, and plans to revive a rationalist learning community.
AI Snips
Chapters
Transcript
Episode notes
Polish AI Creations With Heavy Human Iteration
- Invest human taste and iterative critique to turn AI-generated art or stories into high-quality work rather than settling for quick 'slop'.
- Use many rounds of criticism and refinement, spending compute and hours to polish rhythm, details, and avoid defensive over-explanation.
Gwern's UnSlop AI Short Story Competition
- Gwern launched an "UnSlop" competition funding AI-only short stories with judges like Alexander Wales and a $10,000 prize.
- The contest recommends at least $100 in compute and encourages many hours of human-led refinement to remove 'slop'.
Run Diverse Critics Over AI Outputs
- Use multiple critic filters and repeated improvement passes when generating AI fiction to catch new regressions and maintain readability.
- Stephen Zuber notes failure modes like becoming overly defensive and unreadable after iterative strengthening.