max

are these canary strings a real idea? www.alignment.org/canary/ that someone would republish docs with this string, without this string, on random websites that get ingested into training data, seems like something that people might do maybe i am missing something, if someone could explain?

if i had to bet a nickel, i bet we are on the way to it becoming easier to write secure software, rather than the reverse as dependencies get more secure, intuitively it seems like total system stability would converge to me

w hard problems, if i dont watch the agents at all, they often do crazy stuff, build a shadow architecture of ever-growing complexity, and eventually stop making real progress but if i pay attention, to occasionally suggest how not to do that, things converge to being ordered and efficient enough

was messing around with @attie.ai and managed to make a feed idea I've been wanting for years its a round-robin feed that shows posts from people I'm following with a maximum of one post per author so that frequent posting doesn't mean more visibility

alternatively, in the meantime, I just tested it out and you can just paste a link to this thread into @attie.ai and say make a feed based on the idea in this thread, and it was able to do that

was messing around with @attie.ai and managed to make a feed idea I've been wanting for years its a round-robin feed that shows posts from people I'm following with a maximum of one post per author so that frequent posting doesn't mean more visibility

this is the core idea, although also trying a couple other variations: - same idea, but posts from mutuals show up before non-mutuals - same idea, but uses an "algorithm" to choose which post from each author (valuing likes or recency, and also another one that specifically shows posts w/o likes)

anyone know of experiments with llm-coding-ides that look something like, as the programmer, you look at and review the following and nothing else: - types - in which folders code that interacts with certain types lives - lean proofs about certain relationships between types

anyone know of experiments with llm-coding-ides that look something like, as the programmer, you look at and review the following and nothing else: - types - in which folders code that interacts with certain types lives - lean proofs about certain relationships between types

its funny how for stuff that its already kind of clear how it should work, you can dispatch the llm plan in like 30 seconds and it will churn away for hours and its all correct, and then there is one hairy design detail in one piece, that takes two hours of back-and-forth and head scratching

arguably, if expressivity requires real knowledge of the world and what is being expressed, then ux and mechanism are never really separable, and they only appear separable where there are actually good abstractions

its funny how for stuff that its already kind of clear how it should work, you can dispatch the llm plan in like 30 seconds and it will churn away for hours and its all correct, and then there is one hairy design detail in one piece, that takes two hours of back-and-forth and head scratching

probably i am not sure because i am working on things with a ux and meant to be perceived by a person if you just want a medicine to get rid of the cancer, have a strictly verifiable result, or want the starship to take you to alpha centauri, seems like a different thing

its funny how for stuff that its already kind of clear how it should work, you can dispatch the llm plan in like 30 seconds and it will churn away for hours and its all correct, and then there is one hairy design detail in one piece, that takes two hours of back-and-forth and head scratching

its kind of funny that something named clippy also casually moonlights as a very useful formal verification tool (in so far as intent can be encoded into disallowed methods)

in a sense: - disallow any function F - create new wrappers which call F in some specific ways that exclude some possibility Q - this is now a way to specify that Q never happens