GPT Passes Turing Test

Figure: Jones & Bergen, arXiv:2405.08007

I lost a bet with a friend last year because I expected someone to set up a Turing test for Large Language Models, and I thought the models would pass. Alas, it didn’t happen – no one even bothered with a Turing test, but recently, researchers at UC San Diego tested whether people can tell humans apart from GPT in five-minute chat conversations. They found that the study participants could not reliably identify whether GPT-4 was human. In about 54% of cases, it passed as human. In comparison, humans identified humans as humans with higher than random chance, in about 2/3rd of cases (mean 67%). Just exactly what counts as passing the Turing test was never really settled. Some researchers have argued that convincing 30% of humans is sufficient. Others said it should be 50%. But either way, GPT passed – though it still falls short of actual humans, which is somewhat of a relief. These results also make clear that as AIs become better at mimicking humans, it’s not that we will mistake them for humans, rather, we will readjust how we identify humans until we can identify both only at random chance. Paper here.

This episode of Science News covers gravity without mass. I recently read that you can have gravity without mass. Ha, no way, I thought and had a look at the paper. It's wild, people. The author says that there are hollow spheres scattered across the universe that have no gravitational attraction and that these spheres explain dark matter. I will try to sort it out for you. You can now leave comments on our quizzes!

Hopes for Mega-Collider Shift to China

Planned location of the new collider. Image: CERN

The German government recently rained on CERN’s parade of big collider plans when a representative from the German Ministry for Education and Research (BMBF) made it clear that they do not plan to contribute. Germany is one of the highest-paying member states of CERN and will continue its annual support to CERN. However, the bigger collider – dubbed the Future Circular Collider (FCC) – would require significant additional contributions. Germany’s refusal to contribute makes it even more important that other countries support the FCC, most importantly the U.K. and France, both high-paying member states of CERN as well. The U.K. decision will likely come later this year (though with the upcoming U.K. election it might take somewhat longer than expected). If neither the U.K. nor Germany supports the FCC, then the project is likely to be cancelled. The only way to get sufficient money together in that case would be from the USA and that’s unlikely because the Americans would rather build a muon collider. And Americans being how they are, they’ll prioritise doing their own thing. The Chinese government will likely make a decision about their own collider plans next year.I believe it would be a big mistake to build the FCC. It would commit a lot of resources for decades into a project that’s unlikely to help physics progress.

Learn Science Through Storytelling

You all know that I like my science without all the unnecessary confusion. Nautilus is a free newsletter that uses the power of storytelling to break down complex scientific concepts into fun, intriguing stories. Delve deeper and rediscover the joys of learning with Nautilus.

Canada Plans for Nuclear Fusion by 2050

The Canadian Nuclear Laboratories (CNL) have put forward a roadmap for their path to nuclear fusion. It foresees a transition to commercialization by 2050 and a global expansion afterwards. It’s a remarkably optimistic plan considering that to date we don’t even have a nuclear fusion reactor that produces net energy, not to mention net money. More here.

Support Sabine on Patreon!

Join the discussion on Patreon where you will find transcripts of our videos with links to references.