Since nobody took me up on my $60 bounty for solving the alignment problem, I made a website for it: sixtybucks.org Would someone please just solve it? I'm getting kind of tired of this whole AI risk thing.

At AI Impacts, we started looking into evolution vs engineering solutions to problems, and (as I recall), we only looked carefully at flight efficiency. The short answer is that human engineers and biology hit similar lift:drag ratio limits. wiki.aiimpacts.org/doku.php?…

Open-weight LLMs ship with safety training that can be stripped in a few hundred fine-tuning steps. Can current defenses stop this? We built and open-sourced TamperBench, the first unified framework for evaluating tamper resistance, and the answer is mostly no. 1/7

I’ve said enough about my disagreements with some of the ideas below, or at least with the certainty Pope Leo is expressing about them, but can we also reflect for a moment how wild it is that the Pope is tweeting stuff like this? It feels lifted from a screenplay about takeoff.

I don't think it's a log scale thing. Power output on a bike spans maybe 3-4x from median to world record. It's just that 1.5x in athletic performance seems huge. If it takes you 30 min to climb a hill, and your athletic friend 20, it's hard to imagine anyone can do it in 10.

Hell yeah.

It's been more than a week, and @sleepinyourhat still hasn't answered this question. What is he hiding?

I checked it out today and it's a nice little shop. Most of the items aren't things I particularly want to buy, and some of the pricing seems a bit steep, but on the other hand, I bought the same mug for the same price as Daniel, so who am I to say it's wrong? Luna called while I was there and we had a short, awkward, friendly conversation. After that, I chatted with the employee who was there for a bit and she seemed pretty good natured and happy to talk about the experience.

The card reads to me like Anthropic understands there are nontrivial issues with their current process for ensuring safety and alignment. They clearly put a ton of work into evaluating Mythos, and they clearly know a lot of not-so-obvious places to look for shortcomings in these evaluations. Also, they shared many pages of quantitative results and anecdotes, which is awesome. But, FWIW, I did not read all that much concern in the language. I think there is good reason to suspect this process and anything like it are wholly inadequate for the next generation of models, much less whatever comes after that, and I did not get a "this shit is not gonna cut it" vibe from the model card at all. Did I just not read it charitably enough? Did I not apply a strong enough adjustment for "this passed through some PR team's editing"? IDK, maybe. But either way, that's how it comes across to me and ~everyone who read it with me yesterday.

I'm at the Munich Security Conference. It's interesting. It feels a lot like a normal academic conference, except it's packed with generals and senators and there's armed police pouring out or every orifice. 🧵

Name one thing @davidshor doesn’t have polling data for. I’ll wait.

Man, I really don't like wading into things like this on twitter, but I also don't want there to be a false consensus here. So: I am in strong opposition to ICE's conduct over the past year. I'm not informed enough to have an opinion on ICE's conduct before that. Large-scale deportation is not a priority for me, but if it is for the American public, I think probably it should be a priority for the government. I don't have "a plan" because I don't know how to do things like this. But if the current tactics are all that DHS can come up with, that's a skill issue.

I think what Epoch did here is very reasonable. But the earlier data doesn't show a clear trend, and I'm not really sold on the ~doubling of the rate of progress. Based on my very lazy Google sheets analysis, I'm inclined to model it as one trend. That said, I suppose my best guess on the best rate for modelling near-future progress might be the second slope?

we’re going to have a century of christmas cheer compressed into five years

I don't think I'm an "Ai bro", but: It's the same reason I don't learn carpentry, lens making, farming, tailoring, welding, masonry, baking, medicine, metalworking, basket weaving, etc, etc, etc Occasionally, I encounter people who take their bike to a mechanic for something very simple and think "why don't you just learn bike maintenance. It's not hard and it's pretty fun". But of course the answer is that there are way more things you could learn than you have time to learn, so you will need to pass on almost all of them.