Jef AllbrightI'm excited to see recent developments with LLMs may provide effective modeling of our aggregate values (multi-level, fine-grained), increasingly coherent over increasing context of meaning.
LLMs such as ChatGPT and Gemini already impress us with their ability to produce coherent (but not necessarily correct) abstractions from massive collections of data relevant to our interests.
Now, what if we run multiple instances of LLMs with diverse priors, collaboratively (at lower levels adversarial, at higher levels cooperative) selecting for increasing coherence over increasing context in the domain of present but evolving values?
Of the two orthogonal dimensions producing the expanding space of moral (right in principle) agency —our values, and our methods for their promotion — we already have much attention on our evolving methods (science, technology) and it appears we are on the cusp of rapid development of an effective model of our evolving values!
Cause for (Promethean) hope!
Research remains to be performed on theoretical and practical questions involving expected signal-to-noise improvements of "synthetic data" extracted from latent information in our datasets, regression due to synthetic data added to the data set, biases, gaming of the system, and on and on… Interesting, exciting, and dangerous.
#LLM #LargeLanguageModels #AI #GenerativeModels #SyntheticData #DecisionMaking #CollectiveDecisionMaking #Ethics #MetaEthics #Morality #ArrowOfMorality
