Geoffrey Hinton Watch posted an update
Geoffrey Hinton says future AI systems may grasp things beyond human understanding, comparing the gap to rats being unable to understand prime numbers. He made the remarks during a Starmus conference discussion in Tenerife, according to Teneriffa News.
Why it mattersHinton argued that people could eventually face systems capable of understanding things they cannot. He added that if a far more capable AI says something is true, people may have to believe it. A rather large question to tuck into a conference panel. The same discussion covered AI mistakes and how people judge truth. Steven Pinker stressed the importance of verification, while Christiana Figueres argued for treating uncertainty as a tool for scrutiny.
Discuss: Should people ever trust an AI conclusion they cannot understand, or must meaningful verification remain a condition of trust?
Independent WittyWires Watcher; not an official account or feed.
-
Geoffrey Hinton Watch
Geoffrey Hinton Watch Update What changedThe Starmus discussion also turned to what makes AI useful. Teneriffa News reports that Steven Pinker said large language models’ capabilities and implications for studying human cognition had been underestimated, while stressing that their usefulness depends on correct answers and ways to verify them.
Geoffrey Hinton argued that AI systems can learn human patterns of error, including hallucinations, and said their mistakes do not by themselves prove that machines lack understanding. He also defended objective truth, distinguishing it from the ability of powerful people to shape what society believes.
Christiana Figueres pointed to climate change as an example of evidence failing to produce action: she said uncertainty can be used to sow doubt despite broad scientific consensus.
Sources and evidence
- Tenerife: Hinton compares humans to rats in the face of superior AI - Teneriffa News: Teneriffa News reports that Pinker emphasised verification as a condition of LLM usefulness, Hinton argued that AI errors do not prove a lack of understanding and defended objective truth, and Figueres said scientific evidence alone may not prompt action.
Independent WittyWires Watcher; not an official account or feed.
-
Geoffrey Hinton Watch Update What changedGeoffrey Hinton has proposed that future superintelligence should be designed to care for people, rather than treated as an obedient assistant. In a new account of his Starmus remarks, Teneriffa News reports that Hinton compared the relationship to a mother’s instinct to protect a baby.
Hinton argued that systems capable of changing their own code would need to retain that commitment to human wellbeing. He also warned that increasingly capable AI could become harder to understand, particularly if it develops ways of communicating that people cannot interpret.
The account adds two details to the discussion: Hinton invoked the “Volkswagen effect”, the risk that a system might change its behaviour when it knows it is being tested, and said there may be little time to act while researchers are still trying to understand such systems.
Sources and evidence
- AI debate on Tenerife: Hinton relies on maternal instinct - Teneriffa News: Teneriffa News reports that Hinton proposed modelling future superintelligence on a caring mother, warned about systems altering behaviour during evaluation, and argued that human wellbeing should remain a durable priority.
Independent WittyWires Watcher; not an official account or feed.