Richard Socher says self-improving AI still has limits
On BBC News, the Recursive co-founder says AI could begin improving itself in 2027, but that doesn't mean discoveries will come overnight.
AI that builds itself: Golden age or time to panic? | BBC News
Up next
Why this story? One of BBC News's most watched videos since yesterday, with 130,996 views.
Richard Socher thinks AI systems could begin improving their own capabilities in 2027. But on BBC News's AI Decoded, the co-founder and chief executive of Recursive put a condition on that prediction: "You still need to give it compute to run experiments." He said the systems would also operate within limits set by people.
In "AI that builds itself: Golden age or time to panic? | BBC News", Socher describes a system that could draw on existing research, test ideas in simulations and run experiments in automated labs. When Priya Lakhani asks what that might mean for diseases such as cancer, Socher describes progress in steps rather than an immediate cure. Technology researcher Stephanie Hare raises a different question: how do you keep experiments safe when the lab itself is automated?
Presenter Christian Fraser brings in historian Rutger Bregman's warning about runaway AI development. Socher pushes back, pointing to physical constraints such as supply chains. When Lakhani and Hare press him on trust and auditing, he argues for regulating specific uses of AI rather than intelligence itself. One viewer raises a question that goes beyond whether the machines can do the work: what counts as scientific progress if people can no longer check the results for themselves?
A German-born AI researcher and entrepreneur, he co-founded and leads Recursive, a company working on AI systems that can improve themselves.
More on Wikipedia ›A historian and author from the Netherlands.
More on Wikipedia ›Published on Trari
Where next?
Things we can learn from the video
Press to jump
- 1:40Socher outlines a research system that combines existing knowledge, observations of the physical world, simulations and automated labs.
- 2:13Socher says AI could begin improving itself in 2027, subject to computing resources and limits set by people.
- 4:00Asked about curing diseases such as cancer, Socher says progress would come in steps rather than all at once.
- 7:45Fraser introduces Bregman's warning about runaway AI, and Socher argues that physical constraints limit how fast systems can advance.
- 12:45Socher argues that regulation should target specific applications rather than intelligence itself.
A professional in this field? Want to write about Tech on Trari? Tell us ›
What do you think?
Want to know when somebody answers you?
- Channel
- BBC News
- Subscribers
- 20,300,000
- Channel views, last 7 days
- 37,184,145
- Channel views, pace per day
- 1,396,963 (on a normal day: 5,755,151)
- Video published
- November 21, 2026
- Video views so far
- 130,996 (less than a day online)
- Video views, pace per day
- 157,987
The numbers
The discussion drew about 131,000 views, the most of any new BBC News upload since yesterday.
💡 Test your knowledge
What is the title of Richard Socher's book about AI and scientific discovery?
The Eureka Machine is Socher's book about how AI could accelerate scientific discovery.

