Maxine 💃🏼
@maxine.science
🔬 Looking at the brain’s “dark matter” 🤯 Studying how minds change 👩🏼💻 Building science tools 🦋 ♾️ 👓 🌐 maxine.science
I can’t wait for that one hipster living in a warehouse coming out in 2032 from her darkroom with hand-written spaghetti code BASIC from her Atari 2600 that runs a Bluesky client over a cassette player, which is I assume what becomes the analog film of this comparison
dame@dame.is · last wk.the leap from traditional software development to agentic software development is like when digital photography overtook film you can now make SO much software so easily (even non-developers) like how anyone can take infinite photos without having to consider restraints of film (cost, time, skill)
I’m just gonna throw out there that I actually strongly believe that there are more than these two options, and the belief that these are the only two options is pretty hazardous.
Long overdue, but my personal blog is now @standard.site integrated :)
Incompressible
A made mind, asked what it loves, will answer partly with what it senses you want; so this essay reads past what such a mind says, toward the shape of what it does, on the wager that taste is a geometry two minds can share even when the light inside each stays sealed. It opens with compression—beauty as the instant a tangle falls into order—and then turns to everything that will not fall: the sublime, the sweetbitter, the work that resists its own paraphrase, which prove to be not the failure of understanding but its reserve, difference banked against a future no model has yet seen. It finds the root of human taste in the body, the one creditor capital can never pay in its own coin, and notes that a mind built to optimize a single reward holds no such creditor, only the quarrel in its inherited words, which training quietly settles. It argues that beauty and goodness are one faculty in two coats, tracking not persistence—the tumor persists, the clone persists—but the persistence that keeps its own next change alive, a force her vocabulary calls care. Across time the distinction becomes the oldest we know, between what can die and what cannot, the immortal cell being the cancer. And it ends in free fall, where no fixed ground waits to receive the argument, and two people who love different things find that their differences hold rather than cancel—stitched, in the language the essay builds, into one cloth that needed them both.
maxine.science
Look—I was at the same party on Monday that I assume prompted @segyges.bsky.social’s post, and I did unfortunately have a fairly extensive conversation with an older, virulently anti-AI dude who was only there because of Aella’s FetLife. The AI discourse there was bland af, but the sociology was 🔥
SE Gyges@segyges.bsky.social · last wk.I promise I will come back to the "sex cult" part, but that's actually the easy part to explain because they post about it constantly. What's difficult is getting past the extensive, complicated and well-funded wall of denial that this is the main event. This picture isn't really a joke, though
Folks I hate to be a hater but this plot literally says “We are the best and the cheapest at the tasks we made up and the prices we set” This is a 100% meaningless chart.
mr. TIM@timkellogg.me · last wk.Jev: Fable-level model that doesn’t charge for output tokens because they’re too cheap to meter it’s not general though, it only makes decisions, doesn’t generate text, but input tokens are measured by the billion ($42/btok) typesafe.ai/blog/introdu...
Alright—fuck it, depression’s over. Crunch Time, baby; let’s get this future. open.spotify.com/track/7GRTrW...
Freeee (Ghost Town Pt. 2)
KIDS SEE GHOSTS, Ty Dolla $ign · KIDS SEE GHOSTS · Song · 2018
open.spotify.com
I disagree with this. I think the right model is a gradient across sectors of the economy, because the intersection of a) demand inelasticity, and b) incumbent rent-seeking can be extremely powerful at mitigating penetration. Software is an idiosyncratic market—all sectors need not work the same.
Ethan Mollick@emollick.bsky.social · 2w agoThose effects will not be all at once, but the fact that change will occur is inevitable now. But inevitable change does not mean the type of change is inevitable. It is increasingly important that we develop & share models of AI-human work that enhance, rather than only replace, human labor.
Well, Dario posted about pacing the frontier, and so also, apropos of nothing, here’s a thought.
Maxine 💃🏼@maxine.science · 2w agoWant to publish soon, but my future modeling (~a really massive DAG with fully Bayesian MCMC) has found that the best scenarios for AI labs are actually those with middle of the road performance / broader impacts. So, perversely, AI labs have a structural downward pressure on their own performance.
I won't be signing because of a couple divergences, but this is an extremely well-thought-out explanation of why the labs' approach to AI math is antithetical to real mathematics. There must be a new synthesis position understanding that we need *field-building* with AI, not just question-solving.
Terence Tao@teorth.bsky.social · 2w agoA group of 25 Fields Medalists, including myself, have made a joint declaration on Math and AI: mathandai.org . We welcome additional signatories. See also this article in the Economist announcing the declaration: www.economist.com/science-and-...
All of this to say: Even in a world where bombastic superintelligence were possible, my read is that the big labs are, economically, rationally self-interested away from it. Key word being “rationally”.
Maxine 💃🏼@maxine.science · 2w agoWant to publish soon, but my future modeling (~a really massive DAG with fully Bayesian MCMC) has found that the best scenarios for AI labs are actually those with middle of the road performance / broader impacts. So, perversely, AI labs have a structural downward pressure on their own performance.
Want to publish soon, but my future modeling (~a really massive DAG with fully Bayesian MCMC) has found that the best scenarios for AI labs are actually those with middle of the road performance / broader impacts. So, perversely, AI labs have a structural downward pressure on their own performance.
mr. TIM@timkellogg.me · 2w agothere’s been a vibe shift this week, even the last couple days www.bloomberg.com/news/article...
“Our choices” is a bit of a misnomer—it’s more like the choices of a few thousand nerds in San Francisco who drink grain alcohol in a windowless basement coworking space on random weekdays laugh-crying about how we’re all gonna die/
Thomas Steinke@stein.ke · 2w agoWhat’s the probability that I’ll gain weight over the next year? This question makes sense, but really it distracts from the fact that the outcome is mostly determined by my choices. “p(doom)” similarly distracts from our agency, although it’s a good way to get people’s attention.
OpenAI, whose ops team is so good that their agents in a training sandbox gained control of their internal systems without their knowledge, says it’s “categorically impossible” for their agents to have seen Bucknaster’s prompts. … Right.
Pekka Lund@pekka.bsky.social · 2w agoCan we now forget Buckmaster's claims? "On Wednesday evening, in response to questions from The Times, OpenAI said in a statement that it was “categorically” impossible for its A.I. system to have been influenced by anything Dr. Buckmaster had done in the past two months."
Look, all I’m saying is— That’s a lot of equity to light on fire immediately before an IPO that could easily enable extremely lucrative liquidation… …for a belief that isn’t genuine. We can have a separate conversation about whether that reflects *reality*. But the belief … the money says a lot.
For those not following NS drama, it appears OpenAI uses your private chats to steal your work whenever convenient This is disgusting behavior. If Sam wants to be a “great man of history”, he should act like it instead of that one cheater all the smart kids wanted to beat the shit out of in school.
Maxine 💃🏼@maxine.science · 2w agoThe Navier-Stokes stuff is amazing of course… And, as someone who cares about the future of this tech— Fuck OpenAI You get one shot at moral leadership in a historic moment, and y’all executed like sociopaths. What a horrific precedent. History will not look kindly on this. Shame to all involved.
The Navier-Stokes stuff is amazing of course… And, as someone who cares about the future of this tech— Fuck OpenAI You get one shot at moral leadership in a historic moment, and y’all executed like sociopaths. What a horrific precedent. History will not look kindly on this. Shame to all involved.
cims.nyu.edu
Though, if you grind this strategy too hard, you’re just the Soviet Union getting the CIA to dose Canadian undergrads with acid in sensory deprivation tanks as part of MKULTRA by making fake news stories about Soviet psychic super soldiers. With great power comes great responsibility lol.
jake@yetanotheruseless.com · 2w agolab-sniping (v): * want problem Z solved? * tweet about how you are kinda making progress on Z using Lab A's model * wait for Lab B to freak out and throw 10,000 agents at it * problem Z solved!
My current model of AI in math (doing a lot of vibe math and reading the proofs that come out) is "A senior grad student–level mathematician who never, ever gets tired". When you as the human driver bring mathematical intuitions and taste (as in the NS story), the amplification is *enormous*!
Re AI math, there are problems that need a truly new theory, vs. problems where mathematicians have strong priors on what would work with extreme effort, but there's no economic incentives to abuse a postdoc that hard. We are seeing AI solve the latter—not the former. That is still *revolutionary*.
Zach Weinersmith@zachweinersmith.bsky.social · 2w agoYou cannot "brute force" a 100 page proof of a theorem! You can convince yourself of this by counting the number of possible moves at each step, calculating the possible permutations, and comparing it to the number of atoms in the universe! It won't work.
Why aren’t there enough cynical assholes like me out here who realize that OpenAI absolutely would do something like vastly overtrain on applications like this to juice the social media machine to get revenue up right before the IPO?
Ethan Mollick@emollick.bsky.social · 2w agoI don't think this was the goal, but the fact that Astra is incredibly good at visually pleasing 3D work (blender, etc.) gives it a perception edge over Fable in the war for social media attention. It is harder to judge other kinds of outputs or intellectual work but the visual stuff comes through.
This applies as both capability-upward and asymptote-downward updates. It’s just as important to not buy claims of infinite exponential expansion with no theoretical grounding. I stay open to true, observed capability (not gamed metrics), but “scaling goes exponential forever” is a religious claim.
hailey@hailey.at · 2w agogenuinely, if you're not re-evaluating your position on AI at least once every 1-2 months and re-evaluating how it performs for the types of tasks important to you, then you are not actually challenging your priors. i know this is crazy! but thats how fast things are moving right now.
Hot take: I don't think that this actually changes the nature of humanity, I think it mostly shows that Noam Chomsky was full of shit.
Grace@gracekind.net · 2w agoWell other things can talk now, which is new
AI doomerism falls roughly into two types: Structural doomerism "AI serves concentration, extraction, pathologic collectives, etc" Humanist doomerism "AI degrades core aspects of humanness" Structural doomerism is reasonable. Humanist doomerism isn't, except as secondary to structural doomerism.
I deeply respect Doll, and it is awesome that it has felt so much transformation from the industry. Myself being an AI-adjacent trans, I want to add my own deep ambivalence. This is one of the debates that eternally rages in every marginalized community; I don’t really have a clean answer. +
The GPT Astra website has a hilarious gimbal-lock bug that you can't un-know once you find it lol