vslira's blog

Statement on the US government directive to suspend access to Fable 5 and Mythos 5      I agree with the sentiment that this probably unnecessary at the current level of capabilities, and if this was motived by animosity from the administration towards Anthropic this is certainly a terrible development on institutional grounds. Besides, I'm not an American citizen (or resident for that matter) so this decision, and likely many others like it that will follow (and certainly many others previously that we're not aware) go against my self interests.     Still, it is probably a good thing that we're in the timeline where there's real governamental attention being paid to the potential risks brought by AI.

Man, what a week, eh? Not much to add for now Knowledge-Creating LLMs The thesis is that it will soon become economically possible for large labs to keep frontier knowledge that they discover on their own gated behind API, and maybe not share LLMs that should be capable of finding their solutions. One particular part of this post that I liked is the section about mapping problems to canonical problems, which are themselves "easily", or at least straightforwardly, solvable. I had thought about this before, and for a while I've known that what I really like is exactly finding out this mapping in problems. I'm sad to realize that it will soon be automated, too. Fraud Investigation is Believing Your Lying Eyes patio11 has commentary on the Minnesotan Child Care industry. Money quote: There is a genuine difference in the culture and epistemology of the financial industry versus the government of the United States here. In the...

The Adolescence of Technology Lots of good stuff, what’s more interesting to me is the beginning, where Dario roughly says, if I’m reading this right, that he’s not that worried about AI power seeking because you have to hammer these things a lot in post training for they to begin doing something useful, so maybe runway capability isn’t that automatic. Also, I laughed at “Democracies competitive in AI”. Something tells me he’s not talking about France’s Mistral. What else didn’t he say out loud? On The Adolescence of Technology Zvi’s commentary on the previous text. tl;dr he’s not impressed, still very worried and doesn’t think Dario is doing all he can. The Housing Ladder's Broken Promise Nothing new if you’re already landpilled. Why didn’t he add the happy/sad emojis to section 3, though? Vitol With Nukes The USSR had a place for people who belonged in hedge funds, and that place was managing the USSR's own commoditi...

     I use Anki to learn, and for the last few years I've successfully built a habit of doing my daily cards. It's hard at first, sometimes you don't feel like it, but like any habit if you stick to it eventually becomes easier (not always "easy", just "easier").     But I'm not immune to falling behind, and like everyone who does Anki when life gets in the way of the daily sessions I struggle to keep up and start accumulating cards. Then one day you open the app, sees that there are 100 cards to study and you get that strong resistance to get started.     I could simply change the settings to lower the daily review limit, but I feel like I'd be just sweeping the problem under the rug then. I need to face the fact that I have to work through the backlog, which is also a signal to myself to pause adding more stuff until I get it into a manageable state.     So lately I decided that my ...

     Recently I got myself a new Macbook which is a lot more capable than my previous daily driver - an M2 Mac mini. But I didn't want to retire the old fellow, it's a very capable personal computer still and I don't think it's a bad idea to have a backup in case I lose the Macbook or have it damaged. Besides, there are some applications that I use which are mostly local and even though migrating the related files would be trivial, if I then lost the Macbook I'd lose them, or I'd have to setup a backup strategy. So I decided to keep them in the Mac mini and just use it whenever I needed them - a good excuse to have it turned it on, headless, instead of storing it away in a drawer, unused.      I hadn't done this before so I was a little concerned I might need to install dedicated software (which I'd have to trust with unrestricted access to my home computer) or fiddle with weird settings. I was wrong, fortunately, and modern MacOS makes this a breeze. Bu...

     Related links: Reddit thread , HN thread     A few thoughts:     I fully understand the schadenfreude on the part of the SWE community. These things are possibly an existential threat to their livelihoods and the negativity is fully understandable if not misguided, but these victory laps seem to be repeating a common pattern of not waiting a couple months(!) before these systems improve, very publicly, from a mix of better models and now actively incremented training data from MSFT employees. This is the worst this system will ever be and looking in the improvement rate of the last few months should teach people to wait a little bit before condemning the effort fruitless     On a tangential note, I wonder if this public effort is somehow related to Github's new updated rate limits ( hn discussion ). All this juicy data PLUS expert human feedback out in the open ready to be used for distillation is an obvious target to upstarts trying to...