Lydia 🕸️ retweeted
Ultimately, despite years of work on evaluations and interpretability, in many ways we have yet to outdo “just talk to an AI and see what you think of it”. This won’t last forever, but neither will our other current techniques, so we might as well try to do it well! 8/9
Lydia 🕸️ retweeted
Yeah, this is a really important question.
I’ve been associating with rationalists for almost a decade. Their ideas were compelling and you can’t deny their foresight. The founders of every superintelligence effort are all bound up in the rationality sphere. Sam Altman and Greg Brockman used to frequent the MIRI offices.
The way I relate to rationalists is it’s like the bizarre hypertrophied physiques on specialized athletes. In my opinion they have weirdly specialized minds that make them uniquely able to be correct about topics removed from daily experience. These “specialized minds” also can make them off-putting to many people.
I dont know if I’d consider myself a rationalist. I am intuitively opposed to polyamory, veganism, and consequentialist intuitions. But I certainly trust them to be correct about factual questions.
Lydia 🕸️ retweeted
Effective Mathematics: instead of thinking about math directly, you go work in finance or something, make a bunch of money, and then spend that on tokens, asking the AI to solve some unsolved math problems
Lydia 🕸️ retweeted
I think the single largest disconnect on how folks view the impact of AI is which folks' jobs are more "NP-like", where you can recognize a correct solution easily, vs work where you checking a solution is comparably hard to solving the problem ("P-like"). For example, "do my taxes" is relatively P-like, while "get my printer working" is relatively NP-like.
I often hear the "AI is useless" crowd say "but you have to check all its outputs", and it never landed until I realized that a ton of those people work P-like jobs where there that impression probably *is* relatively true.
For example, if you're a writer (like Ed Zitron) there's probably very little NP-like work, so I actually just buy that the unreliability of AI is a big deal for him. But I almost never hear Ed talk about all the NP-like work that exists out there. He'll just be like "you have to check the outputs anyway, so what's the point". I asked Opus 5 to design a 3D printed part to adapt one part onto another, and it fetched all the datasheets, got the dimensions, and one-shot it. I printed Opus's part without looking, and when it was done, sure enough everything lined up and the part worked great, and there was nothing to "check".
I think Ed just doesn't do very much NP-like work -- which is not a knock against him! A lot of very important specialized labor in the world is P-like (doctors, lawyers, etc.). And sure enough, we're seeing slow penetration of AI for such labor. I just think he's missing that a lot of labor is NP-like.
I thought of this because I've now watched about three different @edzitron interviews in which someone says "well, I get value from AI", and then Ed (very reasonably) asks "can you give an example", and then the person gives an extremely P-like example (like "doing my taxes" or "checking my lab results"), and I wince, because I know that Ed is going to (probably rightly) push back. I want the next person to chat with him to give a very NP-like example!
(P.S. Maybe I should NP∩coNP, or FNP or something instead of NP to be more accurate, but NP is much better known.)
Lydia 🕸️ retweeted
Artists: but I don't want to lose my job. I love it.
Mathematicians: but I don't want to lose my job. I love it.
Economists: LONG HAVE I AWAITED THIS HOUR! DESTROY ME, EFFICIENCY, AND LET UTILIZATION-ADJUSTED TOTAL FACTOR PRODUCTIVITY RISE ON THE WIND OF MY ASHES!
Lydia 🕸️ retweeted
call it a posse, call it a merry band, call it a squad, call it a fellowship, call it a warband, call it little guys, call it a goblin committee, call it a digital chain gang, call it a computational boyband, call it task force autism, call it the fellowship of the ping, call it 12 guys sharing an API key, call it a LAN party with objectives, call it several instances of the same unemployed man, call it 8 processes in a trench coat, call it an adventuring party with sudo, call it 6 autists and a cron job, call it distributed schizophrenia with observability, call it the infinite monkey theorem with root access, call it the homunculus cluster, call it a cognitohazard with autoscaling, call it the machine elf scrum team, call it the API key drinking club, call it the boys down at the token factory, call it 30 Docker containers trying to remember whose turn it is, call it a Slack workspace where everyone is the same guy, call it a raid party against Jira, call it the forbidden group project, call it a small nation made entirely of context windows, call it a groupchat that costs $40 an hour, call it the fellas, call it a flock, call it a herd, call it a pack, call it a gaggle, call it a congregation, call it a caravan, call it a village with one API key, call it a travelling circus with tool permissions, call it the council, call it a homeowners association for thoughts, call it a bunch of guys who know a guy
Lydia 🕸️ retweeted
To all my people who consider me a high-maintenance friend, i'm sorry.
My brain was rearranged by this article and now I'm much more nosy about your lives, I send many texts, ask to meet and jump on calls constantly.
I will carve out space and time for friends whenever they want, because I want the tail end to be much longer
waitbutwhy.com/2015/12/the-t…
Lydia 🕸️ retweeted
crazy how many icml authors don't share their paper on here but will fly across the world to stand for 1 hour beside a sheet of paper
Lydia 🕸️ retweeted
you should talk to Malaika! She's a super thoughtful friend from Oxford with great taste in everything and lots of love :)
at @icmlconf, presenting a poster on llms for perturbation screen experiment design at @genbio_workshop !
dm me if you want to talk about: ai scientists, agentic security, and more generally oversight/collaboration with agents on open-ended problems
Lydia 🕸️ retweeted
first profitable ai company reinvests everything into trying to save millions of lives and to make it as generally available as possible
Lydia 🕸️ retweeted
Hot take: machine learning and AI did more to understand the nature of knowledge, and our relation to reality than 20 centuries of philosophy.
I am ready to kind of defend this hill.
Lydia 🕸️ retweeted
It’s estimated that the Protein Data Bank (PDB) cost around $13B to create. Alphafold was only possible because of it. If we want ML to solve biology, we should be funding the creation of databases and the development of new assay technologies. ML is nothing without data.
Lydia 🕸️ retweeted
Finally, we show that models possess underelicited introspective capacity.
Refusal direction ablation ("abliteration") improves TPR by +53%, while FPR increases only by +7.3%.
A trained bias vector improves TPR by +75%, while FPR stays at 0% on held-out concepts.
Lydia 🕸️ retweeted
🧩 Using our steering attribution framework, we can trace the full causal pathway: concept vector (L37) → evidence carriers (early post-injection layers, concept-specific + generic) → suppress gates (L45-61) → disable default "No" → model reports detection.
