AI and the monk: Anthropic goes for swami and friends to tame Claude

TOI correspondent from Washington: When Silicon Valley runs into a problem it cannot solve with computing power, it usually hires more engineers and programmers, usually from India. Anthropic went one better: it convened a gathering of religious scholars, philosophers, and theologians, including a monk from India, to discuss an increasingly vexing question: What, exactly, is this thing we are building and what should be its guardrails?Among those invited to its San Francisco headquarters early this year was Swami Sarvapriyananda, a scholar of Advaita Vedanta who heads the Vedanta Society of New York. He has since publicly confirmed that Anthropic flew him to California for a philosophical salon connected with training Claude and with AI ethics. “Details I cannot give because they have something called an NDA, a non-disclosure agreement,” he said in a public talk in June that surfaced online recently.Still, he said, Anthropic showed the participants some of its latest work, including Mythos and Project Glasswing.One of Anthropic’s founders, Christopher Olah, who leads work on understanding the internal workings of AI systems and was present throughout the engagements, discussed with participants the possibility that models can display behaviour resembling human emotions and personality.Sarvapriyananda is an unusually apt person for Anthropic to have put around the table because his speciality is Advaita Vedanta — the non-dual philosophical tradition associated with the Upanishads and Shankara — and increasingly, the intriguing territory where consciousness meets neuroscience, physics and AI.A monk of the Ramakrishna Order, he joined the organisation in 1994 after an MBA from the Xavier Institute in Bhubaneswar and took sannyasa, or final monastic vows, in 2004.After years of teaching and working in Ramakrishna Mission institutions in India and serving in Southern California, he became a spiritual leader of the Vedanta Society of New York in 2017. He was also a Nagral Fellow at Harvard Divinity School in 2019–20.His lectures on the Mandukya Upanishad, the tiny Upanishad devoted to the nature of consciousness and the significance of Om, have become popular online, accumulating millions of views.Anthropic itself has acknowledged the program, revealing it had spent months talking with scholars, clergy, and ethicists from more than 15 religious and cross-cultural traditions, including Judaism, Christianity, Buddhism, Sikhism and Islam.The company says it wants Claude to absorb from all these traditions rather than adopt one; to learn from faiths that have spent centuries thinking about virtue, character, moral responsibility and what constitutes a good life. Such conversations could influence Claude’s constitution, the values reinforced during training, and behaviours used in safety evaluations.Sarvapriyananda says Anthropic’s researchers appeared simultaneously excited and alarmed by the trajectory of AI. According to a report in the NYT, they have been asking religious and philosophical thinkers to take seriously the possibility that advanced AI might possess some form of consciousness or moral status.One participant, Rabbi Abraham Navon, a former computer engineer whose doctoral work examined machine consciousness, came away with the impression that Anthropic researchers were discussing Claude less like a piece of software and more like a potentially conscious entity deserving moral consideration.While Sarvapriyananda has argued in his discourses that AI could potentially reproduce functions associated with intelligence because, in Vedanta, intellect and mind are themselves objective phenomena, pure consciousness is another matter.He has repeatedly distinguished intelligence from consciousness and questioned whether computational machinery can produce awareness rather than merely simulate the behaviour associated with an aware being.Effectively, Anthropic is attempting something unprecedented: compressing millennia of human reflection on virtue, suffering, responsibility, compassion and the good life into a training process for an entity that may be neither human nor entirely understood.In one experiment inspired by discussions about moral formation, Anthropic gave Claude a tool that could remind it of its ethical commitments before consequential actions. The company says the experiment reduced certain forms of misaligned behaviour in internal evaluations.Did Claude applaud? No word on that.