High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario.

"One of the very confusing things about the models right now: how to reconcile the fact that they are doing so well on evals. And you look at the evals and you go, 'Those are pretty hard evals.' But the economic impact seems to be dramatically behind. There is [a possible] explanation. Back when people were doing pre-training, the question of what data to train on was answered, because that answer was everything. So you don't have to think if it's going to be this data or that data. When people do RL training, they say, 'Okay, we want to have this kind of RL training for this thing and that kind of RL training for that thing.' You say, 'Hey, I would love our model to do really well when we release it. I want the evals to look great. What would be RL training that could help on this task?' If you combine this with generalization of the models actually being inadequate, that has the potential to explain a lot of what we are seeing, this disconnect between eval performance and actual real-world performance"

With the help of an AI model, sentences from a vocabulary of up to 125,000 words were recognized with up to 74% accuracy. Researchers at Stanford have decoded for the first time the brain activity that occurs when people imagine themselves speaking (known as “inner speech”). These “silent thoughts” were decoded in real time with up to 74% accuracy using a brain-computer interface (BCI). · Microelectrodes were implanted in the motor cortex of four participants with severe paralysis (caused by ALS or stroke in the brain stem, among other things). · They were asked to either speak or imagine sentences. Similar activation patterns were observed in the brain, although inner speech was less pronounced. · With the help of an AI model, sentences from a vocabulary of up to 125,000 words were recognized with up to 74% accuracy. · Interestingly, even content that was not explicitly read aloud was recognized—such as numbers when participants were asked to count pink circles.

Reset.

The recent history of AI is extremely poorly documented, despite its importance. Most of it is not on the news. ~none is on Wikipedia. Slightly more of it is on Lesswrong. Even a little more was posted about on X, but good luck finding those. Much more of it is stored in my mind.