Demis Hassabis says society is “not ready” for the AI future

Hassabis wants AI guardrails in place before machines improve themselves

Demis Hassabis built Google DeepMind around one bet, that AI is the fastest route to better science and medicine, and AlphaFold is his proof. That capability is improving itself in verifiable domains, and he thinks removing the human reviewer from the loop is the dangerous step toward AGI. His answer is more engineering and governance rather than less, with hardened cyber defences, third-party auditors, an industry standards body, and machines that stay tools.

Takeaways

• Hassabis's goal is AI for science and medicine. He says the whole point of his career in AI was to build the ultimate tool for advancing science, medicine and human health, through projects like AlphaFold and the drug discovery spin-out Isomorphic Labs.
• AlphaFold collapsed a fifty-year biology challenge. Protein folding resisted researchers for fifty years, and drug design needs a protein's three-dimensional surface to know where a molecule binds. AlphaFold finished in 2020 and four million researchers use it. Hassabis expects nearly every future drug to rely on its structures.
• Self-improvement works only where answers are verifiable. Recursive self-improvement works best where answers can be checked, in maths and coding, he says. AlphaZero went from random play to beating any chess entity within twelve hours after a million self-play games. With a human reviewing, systems make engineers three to five times as productive.
• Closed-loop self-improvement is the dangerous route to AGI. Removing the human reviewer from the loop carries the most risk and is a dangerous route to full AGI, he says. The idea was once called an intelligence explosion because control is hard to keep, and the result would likely be alien intelligence.
• Cyber defence has to outrun cyber attack. Hassabis says many recent cyber incidents could have been handled with more robust engineering. Labs need hardened sandboxes and monitoring so agents cannot run for weeks unnoticed, and an incident he calls an oil spill will happen unless far more resources go into defence.
• Labs should be audited by independent third parties. All leading labs agree on third-party auditors, Hassabis says, so an ecosystem of auditing firms can grow. His summer essay proposed an industry-run standards body with government oversight, and he wants labs to red team each other before release.
• He wants intelligent tools, not conscious machines. Intelligence and consciousness are double dissociable, and there is no agreed definition of consciousness. He sees a trade-off between self-volition and corrigibility, and says society is not ready for the debate. He would cross the first Rubicon before deciding on the second.