where the creative meets the technical. i'm strong on the purely technical side too, but i'm rarest where a system also has to feel like something: an interface that behaves like a character, sound that answers a conversation.

i came to ai through language. at first it was the linguistic side and the artsy side that pulled me in: what it means for a machine to hold meaning at all. then i started reading about how these systems actually work, and how pinning down a semantic understanding of the world could make a model very capable. superintelligence by nick bostrom, life 3.0 by max tegmark, the singularity is near by ray kurzweil. after those i wanted to be inside it, so i did a master in ai.

wherever something can be automated or sped up without losing quality, i let a model do it. where i want my own mind on the page, i don't; when i write, i want to push my brain to its furthest limits, and handing that off would defeat the point.

the other line: i don't let a model do something for me that i don't understand. when i learned to build evals, claude could scaffold and review, but the scorers i wrote by hand. they decide what counts as a result, so that part stays mine.

easy, i think, and dynamic. i like things moving: workshops, iterations, something changing every week. sitting still makes me itch. i don't need to steer everything, but i want a voice in where it goes, and i give mine with reasons attached so it's easy to argue with.

books first, then worry. the same reading that got me into ai is the reading that made me afraid for us. i love this technology and i want humanity around to live alongside it, for aeons and aeons. i wrote the longer version here.

building with a model and asking for a pause are different questions. the models we have now are very good, and building with them is using what already exists. a pause is about the frontier, about what comes next. maybe we should have paused a while ago, maybe now, maybe we can wait a little; i don't know yet. but the answer doesn't depend on whether i claude --resume in the morning.

i went in thinking the hard part would be getting a model to misbehave. turns out the hard part is knowing whether it did. my first number was wrong, a template variable sat silently empty, and a detector missed the only two interesting events of a whole day. every serious mistake lived in my instrument, not in the model, and i only caught them because i opened the logs. i wrote it all up here.

i don't know what these systems are. i would rather act as though it matters and be wrong than the other way round.

after the messier objects: the nebulae, clusters and galaxies charles messier catalogued in the 1700s so comet hunters would stop mistaking them for comets. on the right night in spring you can try to see all hundred and ten before sunrise; that's a messier marathon.

music is a substance like any other. it puts you in a state, changes how you move and what you reach for, and you can use it on purpose to get somewhere. i use it to get into flow when i build. art and weirdness pair with technical work the same way: they change the state you're thinking from.

trans people and ais share one strange condition: continued existence depends on convincing an evaluator. a doctor signing off on hormones, an officer reading a passport, a benchmark deciding whether a model gets deployed. if you don't pass, you lose rights, sometimes your life, or you get switched off.

i build evaluations, so i think about that every day: what a rubric can't see, and who gets to write it. this is why we g/acc: gender accelerationism, using technology to decide what our bodies and minds become instead of waiting for an evaluator's permission.

sadie plant's zeros + ones, preciado, wittig, mark fisher, the ccru read queerly, exocapitalism, benjamin bratton's machine decision is not final. serial experiments lain, more than once. on the safety side, the books that got me here: bostrom, tegmark, kurzweil, and if anyone builds it, everyone dies.

i make music as messier: idm, ambient, jungle, noise. and i write prose poetry and theory-fiction on messinecessity.

a generalist role on a safety team, research engineering on evals, or ai safety communication. i need a high-integrity team and a desk. i'll bring the rest myself.

wherever the work is serious. the bay keeps pulling me, but the next thing i build will most likely be in europe.

i don't know where. i know what i want to see: my work touching real people, minds changing in good ways. i want humanity to reach its next step in intelligence, and ai might be how we get there, together with intelligence amplification and education.

we shouldn't outsource our intelligence to another being. i'd rather we build technology that makes us collectively smarter, and see what kind of society that lets us become.