Highlights: Accidentally teaching AI models to deceive us | Ajeya Cotra (2023)
Highlights from our episode with AI researcher Ajeya Cotra about the AI deception pitfalls we might be walking into as we train today's models (episode #151 of The 80,000 Hours Podcast, released May 2023).
There's much more to say on this topic, so if you enjoy this, definitely check out the full episode: • Accidentally teaching AI models to deceive...
Highlights include:
• Rob’s intro (00:00:00)
• How ML models might develop situational awareness (00:00:20)
• Why situational awareness makes safety tests less informative (00:04:21)
• What misalignment doesn't mean (00:08:31)
• Why it's critical to avoid training bigger systems (00:10:31)
• Why it's hard to negatively reinforce deception in ML systems (00:12:58)
• Can we require AI to explain its reasons for its actions? (00:18:44)
• Ways AI is like and unlike the economy (00:22:25)
Learn more and find the full transcript on the 80,000 Hours website:
https://80000hours.org/podcast/episod...