乱世中求最优解
而梦想永不可及
Omne
Karma: 25
re. Claude Mythos release:
I’m scared.
This seems to have aged really poorly in light of the dispute between Anthropic and the US Department of Defense. Given that around an hour prior to this post Pete Hegseth has name-dropped EA as an enemy, it’s going to be difficult to keep the movement outside of partisan politics.
https://x.com/SecWar/status/2027507717469049070
In retrospect, AI safety is often seen by laypeople as a matter of restrictions on what AIs are able to say and do, which makes it a convenient target labeling as “woke” once it enters public discourse.
Epistemic status: untested hypothesis but seems obvious now that I’ve thought of it.
When choosing your note-taking method, you should take your entire learning system into account.
I’ve been taking handwritten notes until now, mainly due to research indicating that handwriting notes improves retention relative to typing notes. At the same time, I use Anki for active recall and spaced repetition. However, this combination makes writing paper notes superfluous—the retention benefits obtained via handwriting notes are probably dwarfed by those of active recall. In that case, I should optimize for the speed at which I can take my notes and move them into either summaries or Anki decks, a goal to which taking notes on paper is antithetical.
The “handwriting improves retention” is probably accurate but incomplete—it doesn’t account for the time spent on note-taking, or for the interactions between note-taking and studying.
Personal context: taking notes at a summer camp where far more of the content is unfamiliar to me than in classes at school. The notes I take during class come out more disorganized if I don’t have the time to fully process the content during class, and written on paper, incurs the overhead of having to sort through and rewrite them. In school that’s at most 5-10% of my notes for a given class, but it’s closer to 50% in the summer camp. Observing the terrible inefficiency made me consider overhauling my note-taking system.
Has anyone done this?
Yes. Computer chess enthusiasts have created a rating list known as CCRL that covers chess engines ranging from the strongest ones we have to somewhat simple amateur-developed chess engines. Most of the engines on that list are open-source.
Two catches:
The weakest engines on the list still exceed random play by a substantial margin.
Implementing some chess engines to fill that gap won’t be difficult, since libraries are available for move generation, and you don’t need efficient search techniques if you’re trying to make a weak chess engine.
CCRL games are typically played from openings that give a small advantage (usually around +1.0) to White, rather than directly from the starting position. Two games are played for each opening, with the engines switching colors between the two games. Unequal openings are used because in high level engine play, playing games from the starting position will result in too many draws.
If you have questions about basic chess programming, please feel free to ask me. I recently started work on a chess engine (as a personal project) and thus am familiar with some of the concepts.
Thank you for asking :)
I didn’t expect this to blow up, but I guess here we are.
The standard AI fear seems possible: we see another abrupt increase in AI capabilities this year, recursive self-improvement happens before anyone is ready for it, and we get paperclipped.
I’d previously assumed that LLMs would plateau somewhere due to architectural inefficiencies such as the fact that they run concepts through language, creating overhead compared with processing concepts directly. Mythos is an update away from that since AFAIK its architecture does not appear to be something other than a scaffolded LLM. Regardless of architecture, I did not expect Mythos-level capabilities to appear so quickly; I expected incremental improvements rather than a model that cleanly surpasses all existing ones in almost all tested benchmarks alongside finding vulnerabilities in extensively audited software like ffmpeg and OpenBSD.
Regarding personal outcomes: I currently evaluate that all of the positive utility in my life occurs in the future, and that makes me anxious about having my plans cut off by transformative AI. I’m stating this vaguely because my personal circumstances are rather complicated and I don’t know how much I want to publicly disclose.