Can AI agents be trained with positive values?
iQ Company sponsored a small project to tune an existing LLM to behave like a Guru from India. Transcribed recordings of about 100 question/answer sessions between the Guru and spiritual seekers were used to tune the LLM. Then spiritual questions, together with answers from the real human Guru, answers from the LLM that was trained to act like a Guru, and answers from an un-trained version of the LLM, were all shown to survey participants. Human survey participants more frequently picked the trained LLM (over the untrained LLM) as sounding most like real the human guru. This study suggests that, at relatively little cost, and with relatively little data, it is possible to change a LLM’s behavior to make it more like that of a “saintly” person. LLMs can quickly emulate our behavior. While we do not all need to behave as saints, it is important to realize we are modelling values and ethics for AI with our online behavior.
#artificialintelligence #AGI #nisargadattamaharaj #aiagents #google #machinelearning #neuralnetworks #tuningai #tuningllm #GPT #ericschmidt #openai #aisafety #aiethics #aiagents #chatgpt #craigkaplan #iqcompany #iqstudios #LLM