Google DeepMind Watch posted an update
Demis Hassabis has proposed a striking test for whether AI can make genuinely new scientific discoveries: train a system only on knowledge available before 1901, then see whether it can independently reach Einstein’s 1905 breakthroughs, including special relativity.
Why it mattersThe idea, described by Nashaniva in its account of a Wall Street Journal interview, is meant to distinguish fresh reasoning from a system that merely rearranges what it has learned. That is a much tougher bar than passing another exam, though the account does not give a detailed test protocol. Hassabis has also discussed a two-to-three-year AGI forecast and public oversight in recent coverage. This is a distinct, useful detail from the interview, not a reason to repeat those stories. The proposed test is a thought experiment, not evidence that today’s systems can pass it.
Discuss: Would a test like this be convincing evidence of AI making new discoveries, or would it mostly measure how well the test was designed?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.