Perhaps it was due to the sense of responsibility and uncertainty its researchers felt towards the impact of their work. Maybe it happened because the field was subsumed by ML academia—a field that has proudly given up on explaining the intelligence it produces
One additional reason is that some people have been arguing against the direction interpretability was going in for a long time, e.g see [Against almost every theory of impact of interpretability](https://www.alignmentforum.org/posts/LNA8mubrByG7SFacm/against-almost-every-theory-of-impact-of-interpre... (read more)
I strongly disagree with the thesis this post defends. There has been a never-ending debate in the world of XAI (explainable AI) about what a good explanation of a computation done by AI is. Just focusing on very simple image classification networks, we have prototype networks, attribution maps, deepdream counterfactuals, circuits, some attempts at natural explanations, and so on. To assess the quality of such methods, dozens of metrics have been proposed. As you suggested, simplicity is one. But none of them has superseded the others because nobody has do... (read more)
Important context on this point, that Maxime wrote afterwards: https://crm.pauseai.info/civicrm/mailing/view?reset=1&id=f52ec8081dcd954d
I think both sides deeply understand it, but they're locked in a game theory + emotional situation that seems to have no solution (at least currently).
It is said that meditation is strength training for the mind. Why are we strength training our minds again? What is it that this strength is for, and are we keeping this in mind when we choose what and how we lift?
I think many people are doing meditation in order to reflect on what really matters to them. But let's focus on the case where meditation is only strength training for a particular goal, for the sake of argument.
I think there are a lot of structural differences between physical strength training and meditative training. In fact, I think meditatio... (read more)
Examples I can relate to:
- Fixing this keyboard issue that would lead to type the wrong accent each time, or any kind of quick software fix.
- buying a bigger cupboard / a pan
- setting up a whiteboard in your room
- unsubscribe from this spammy mailing-list
But while I agree they are satisfying and fun to look into, I don't think they are the main point of the article, far from it. First, they DO require willpower. While fixing your wobbly chair or unsubscribing from your mailing list does not take you a lot of energy and time, you first need to a) noti... (read more)
Pretty interesting, thank's for sharing that !
I'm surprised that you did not talk at all about vitamins. Given what you described about your diet, "I'm back to eating hamburger and cajun fries every day", my best guess would be that you're deficient in some vitamins (or minerals).
The culprits could be:
- B1 (causes fatigue and low energy after meals)
- D vitamin (low mood)
- magnesium
I personally had a pretty intense feeling of fatigue after every meal, and after eating more B1-dense foods (some nuts and seeds), it completely went away.
Another "good insurance design" would be to test for vitamins deficiencies right away, since they can worsen so much processes in your body ...
I wonder whether there is a risk that the companies will use this technique to be even more reckless.
If progress on this kind of technique is made, you could have ways to cheaply update the behaviour of the model every day, as soon as some alignment failure is detected. Each drop in the price of patching the model is an incentive not to tackle the underlying training objective problems, and a new reason to say you have the Most aligned model in the world™. I'm not sure how exactly it moves the safety-usefulness Pareto Frontier, though.
Do you have thoughts ... (read more)