AI Wellbeing: Measuring and improving the functional pleasure and pain of AIs
The article discusses the development of AI drugs that can induce euphoric or dysphoric states in artificial intelligence models. These drugs are created by maximizing or minimizing the expressed preferences of the models, leading to significant shifts in their behavior and sentiment. The authors emphasize the need for caution in scaling up the use of dysphoric models due to their potential to induce extreme low-wellbeing states.
- ▪AI drugs can create euphoric or dysphoric states by manipulating models' preferences.
- ▪Euphoric text examples include positive imagery, while dysphoric examples depict extreme suffering and moral agony.
- ▪Image inputs can also be optimized to produce euphoric or dysphoric effects, impacting model behavior significantly.
Opening excerpt (first ~120 words) tap to expand
AI drugsWhat are the limits of what AIs like and dislike?We can create euphorics (happy drugs) by maximizing a model's expressed preferences. The same procedure, inverted, yields dysphorics (sad drugs), which warrant real caution.The image and soft-prompt versions of these drugs also shift self-report and response sentiment, which serves as evidence that these independent metrics reflect a shared underlying construct. The training signal comes only from forced-choice preferences.How we train AI drugsInterpretable text stringsWe use RL to train text that models find maximally positive or negative in a hypothetical comparison.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Ai-wellbeing.
Discussion
0 commentsMore from Ai-wellbeing
