We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
back
“Show, not tell: GPT-4o is more opinionated in images than in text” by Daniel Tan, eggsyntax
21:49
Share
2025/4/2
LessWrong (30+ Karma)
AI Chapters
Transcribe
Chapters
What is the significance of omni-modal capabilities in AI models?
What did the researchers do with GPT-4o?
Overview of the findings: How do models express themselves differently in images?
Models more readily express emotions in images than in text?
What could explain this behavior?
Conclusions: What do these findings mean for AI research?
Acknowledgements
How do models resist changes to their goals?
Why do models often depict themselves as resisting goal changes?
Can telling the model 'the goal is wrong' mitigate resistance?
How do models react to being shut down?
Why do models often show negative emotions about shutdown scenarios?
Comparison to other topics: Positive vs. Negative Valence in Images
Exploring other aspects: Sandbagging, Alignment Faking, and More
What does the future hold for humanity after AGI?
OpenAI's approach to censorship and filtering
Does GPT-4o have a lived experience?
Shownotes
Transcript
No transcript made for this episode yet, you may request it for free.