Have you tried turning off the hype?
I had yet more conversations this weekend about generative AI, specifically about Anthropics’s new Mythos model. Some people were worried for their website security after reading the news.
If you missed the noise, Anthropic (the people behind Claude) announced a model so good at finding security flaws that they decided they couldn’t release it to the public.
People panicked, journalists wrote horror stories, cybersecurity stocks dropped.
But the story got a lot less scary when people outside Anthropic’s marketing department examined it.
Independent security researchers took the specific vulnerabilities that Anthropic were hyping up, and ran them through smaller, much cheaper AI models. Turns out they all found exactly the same vulnerabilities.
There was some progress in performance for this new model, but nothing unusual. Previous model releases showed similar improvement jumps.
The only unusual thing here was the marketing.
These companies have been telling us (and investors) for years now that they’re months away from Artificial General Intelligence or that their models are going to replace half the jobs.
But, when you get down to it, the best press they could have gotten for their revolutionary new model was “it finds bugs a bit better”. So they chose fear instead. If your product isn’t revolutionary, make it a threat, and journalists will do the rest.
I’d love to stop writing about generative AI. But, as long as the hype machine keeps turning, I guess I’ll be showing up to spoil the party.
Colin