today's @wsj deciding to cover my latest essay on agency is crazy meta. it's proof that the bar to make something that can emerge - the idea of merit through action using AI, in this case - is not beyond reach. it doesn't _require_ you have stamps and credentials. i have a boring bachelor's degree in econ from a random public school in italy, and i get to work in the hottest market and on the hottest tech of the last three decades. i genuinely believe emerging today is primarily about having the willingness to do it. to be hungry, i guess low key im so proud to see my stupid little blog grow up... (also, so refreshing to read a pro tech take in mainstream media)

it's honestly uncanny how much his speaking and rethorical style hasn't changed almost a single bit kinda cuts against my arguably naive intuition that you grow into greatness through will and life and hard work. maybe it looks more like you're predestined for it

so excited to see @the_marwell join an interview (for the first time?) - he's an absolute legend, materially shaping the frontier and the way we make models i know what i'll be listening to during tomorrows car trip

this model has been my preferred choice for most of the things i care about -- in no small part because of they way it communicates and talks to me opus is back! .5 versions really seem to be special

this is truly exceptional work. i had the chance to talk to Steven a bit about how he went about designing and building TB-Science, and i came away really impressed making benchmarks is easy. making an outstanding one is unbelievably hard. the incentives are stacked against it, from diluting task quality to ship fast, to building hill-climbable problem shapes you can beat by buying envs on the market see, many folks think of benchmarks as marketing tools, and in a way they are! but the best ones are way more. to me they're basically the AI version of GDP. heavily gamed and debated, yet still indispensable. a good public eval doesn't need to be ungameable. it needs to be hard enough to game that gaming it looks a lot like actually getting better steven and his team built something special here. measuring the frontier of science and how fast (or how slow!) models are pushing it further is arguably one of the most important things in the world today. doing it with relentless rigor makes it a public good can't wait to see TBScience v02 come out, and if you're a scientist, you should consider contributing!

interesting how certain things are so pervasive and ubiquitous and considered fundamental fabric of society -- and yet can disappear so fast. woke up this morning realizing that i haven't done a "proper" google search in years. remember when we were typing various keywords in google, scanning through first page results, skimming through pages and pages of mostly-unrelated-but-hopefully-mildly-related textual content, when solving a problem? trying to estimate the credibility of a source, typing symptoms of the flu and having to skirt all the "you have cancer" and "you're pregnant" 2007-era forum results. chunking progressively shorter pieces of whatever stacktrace you got, and spending hours (days!) reading blog posts and SO answers and piecing together how things were done. obvious in hindsight, but i only now realize how much all this was still, fundamentally, shaped like going to the library. a faster horse. acceleratory, yes. but not ontologically different than search & retrieval in the 70s. i wonder how much of the world we're seeing today, that we might consider natural laws on how we go about things, is just one model generation from being overhauled. what's the next horse? one of my current bets is medicine as population science. modern healthcare is built around the average patient. we get prescribed standardized exams, standardized protocols, standardized drugs. the whole system is pattern-matching from that average to us, because personalized inference over n=1 is too expensive. but what if it's not anymore? if we can have entities that can reason over our genome, our labs, our history, and literature, all at once... "standard of care" starts sounding very much like "first-page-of-google-results". not just acceleratory - ontologically different! how many other fast horses are out there?

the most goated example i could find when eval'ing Opus 5 was on a frontier bench very hard task. the model is given an engineering drawing of a machine part and asked to write code to rebuild it as a 3D freeCAD model. the fun catch is that the model has no way to view the drawing: mini-swe-agent is a text-only harness, so all the model ever sees is terminal text over bash IO. it's something hard and multidisciplinary even for a human engineer capable of seeing the drawing! working blind, Opus 5 first tried off-the-shelf OCR with Pillow and Tesseract, found the results unreliable on the drawing’s thin numerals, and built its own vision stack on the spot. carved out each printed character as a connected component, rendered the glyphs as ASCII art (?!?), and matched glyph bitmaps to read the dimension labels off the page then it vectorized the drawing itself: run-length line detection with sub-pixel centroids, least-squares circle and ellipse fits to recover radii and bores, arrowhead detection to tell dimension lines from part geometry. it worked out that the four views were printed at different scales, derived each one’s mm-per-pixel ratio, and rebuilt the part, feature by feature, as a parametric freeCAD model welcome to the world, my dude!

interp is genuinely the coolest and in a way underrated branch of ai research. i so deeply wish i was enough long-tailed in math and shape rotation to be able to contribute to it. it's not just telling us something interesting about how a piece of tech we built works -- by both comparison and contrast, it's forcing us to reflect on fundamental human nature itself