Hi, a bunch of very cool AI stuff this past couple of weeks!
By the way all the past e-mails can be accessed here: https://neuroailab.ucsf.edu/news/
1. AI self-improving and risk mitigation plans
The Darwin Gödel Machine: AI that improves itself by rewriting its own code
Deterrence with Mutual Assured AI Malfunction (MAIM)
2. The gap between the hardest problems powerful systems can solve vs. the simplest problems they can’t is increasing!
OpenAI o3 can solve stereograms!
It is fascinating to observe the reasoning trace: This was the image:. This is the model’s reasoning:
Click on Thought for 2m 36s to expand the reasoning.
OpenAI o3 can’t tell the time of a wristwatch!
This was the image: This is the model’s reasoning: It took almost 5 minutes to give a wrong answer!
