"The future will come in one minute." -- my three-year-old daughter.

all i'd say is that if i were in Dario's shoes, i'd now need a wellness retreat /stat/.

Whee! Knicks. Fireworks in nyc!

Woah, uh?

switched from a unicomp to a keychron w/ TTC Frozen Silent V2 39g Linear Plate Mount Switches. it's taken me two weeks to type on this without fat-fingering. those on video chat with me are grateful: typing on the unicomp sounded like an act of terrorism on the other end.

awesome, David! everyone, check out the below.

watched @jeremyphoward's recent video. wanted to unpack some thoughts about a future of serious software that nobody reads and what that means for fulfillment and anti-slot machine ways of living life. first off, if you haven't seen it, please do. it's short and great (no surprise). youtube.com/watch?v=SUZwYV5J… i agree that long-term delegation of critical thinking will put you in a funk. i also worry that the generation growing up with AI might not develop the ability to focus. anecdote: i overheard someone in high school stunned to learn that not that long ago, people wrote 8, 12, 20 page conference and journal papers through sheer brain force. (and as someone who saw Google's codebase evolve, i have a deep appreciation of the mind's ability to focus.) where i begin to drift from Jeremy is on whether you need to go deep on every aspect of a problem you're working on. his demo of SolveIt was awesome (amazing work Rachel and team!), and i'd love to try it. but the reality is that human attention and capacity is limited, and time is limited, and technological progress is one of abstractions. what's interesting to me is where the discontinuities happen: what level of abstraction do you have to get to to do novel AI research (as one example) so that certain fundamentals of AI can be "chunked" (in the Hofstadter sense)? surely our ability to innovate will end if we need full representations in our minds at all times. of course, we can extend our attention by delegation / working in teams. Transactive Memory theory is awesome at covering this, and is being updated for our agentic world. i could be an expert in X and not need to go deep in Y if my team-actor (actor == human or machine) knows Y. and then we get these proofs of Erdős problems that likely happened only due to the Langlands Program-like breadth and superhuman tenacity of current models. is the prompting of the mathematicians legitimized only after they dig in to understand the proof? what happens for proofs that will eventually be too hard to understand? the work that my team and i are doing sits at the intersection of software engineering and AI. i recognize in Jeremy someone who loves to code. i do, too. but the thesis question we have at wheelie.dev is: what becomes possible when you haven't just delegated all code writing, but you're not even reading the code produced? i realize this is heretical to some. first, is it even possible to build complex systems this way? we're taking perhaps a more experimentalist than scientific approach. the user asks for what they want built, but not as a complete spec, more as an iterative dialogue with an agent, and we include not only the backpressure of typed languages but the double checking of formal methods analyses. think of it like the bone chair (jorislaarman.com/work/bone-c…). it's almost like they said: "ok, the chair needs to hold x pounds, be comfortable for a human body, etc." something comes out that works, and ymmv on whether you actually want to look at it. if agentic coding abilities aren't going to slow, i think we should use /our human ingenuity/ on how to coax AlphaGo move 37-like behaviors from it. further, we may learn that AlphaZero's same lesson applied to code: remove the human scaffolding, and you wind up with something beautiful, mysterious, and slightly beyond the ken of the human mind to understand. back to the discontinuities: i'm positing that certain problems we wish to solve both (a) require code and (b) require us to take as a chunked primitive the behavior of that code. i don't think we're anywhere close, which is why, back to Jeremy's video, i wouldn't blindly accept choices from a coding agent that i didn't understand. but the line isn't static, and i think we should accept some level of letting go. @wtgowers re: whether the future is a crisis or exciting @blaiseaguera, @bratton re: centaurs. @michael_nielsen re: chunked primitives

Incredible. That last put back was one in a million. (Knicks)

how i feel when an agent chooses python for some ad hoc tooling...

what every older AI researcher is thinking when comparing their $$ from back in the day with what the next gen is making.

so.... (and look, i love coding and driving) ...what will happen first? considering dangerous / antiquated to brain code or to drive a car?

contrarian view: still getting better results w/ 5.5 than Fable5 on large codebases.

the snare on Fugazi's Repeater is lacking the high freqs i remember when listening in my late teens. bad hifi, or bad ears?

movie i can't get out of my mind: Sirāt.

He did indeed. I was flabbergasted by his result, which has made much existing (and continuing!) work in quantum foundations obsolete. But it has received grossly insufficient recognition from the community. They still don't know what hit them and are still ingeniously discussing non-existent things like "quantum non-locality" (and "the measurement problem") in ever greater detail. He deserves patrons.

never rollback