Tessera
Folgen

@danabramov.bsky.social

123 Posts240 Follower135 Folgt
@danabramov.bsky.social · last yr.

i wonder what it would look like if blocklists were kind of voluntary. like “yeah i talk about ai” but you add yourself to it, rather than get added by someone. like a badge you wear. and you can choose to block out all people wearing some badge if you want

@danabramov.bsky.social · last yr.

spent ten minutes debugging why an element's border was showing underneath another element despite the latter being fully opaque. turned out it was a hair on the screen. i only noticed after i scrolled the page a bit down

@danabramov.bsky.social · last yr.

claude code is giving me the courage to do things i would never do (preprocess SVG illustrations for my blog with the gnarliest regexes to slightly tweak Excalidraw's output)

@danabramov.bsky.social · last yr.

lmao i'm asking claude questions about how regularity prevents russell's paradox and claude is getting so confused it asked *me* to help it untangle it

@danabramov.bsky.social · last yr.

ok i didn't realize this before but it's kind of wild that in Lean, which is strongly-typed (!), the return type of a function can depend on the input values (can even be extracted to another function to be calculated). i guess that's obviously needed for proofs to work but it's still wild

def pickType (x : Nat) : Type := match x with | 0 => String | _ => Nat def getValue (x : Nat) : pickType x := match x with | 0 => "zero" | n + 1 => n #eval getValue 0 -- "zero" #eval getValue 5 -- 4
@danabramov.bsky.social · last yr.

does claude code never "undo"? seems like as a human i often "undo to last checkpoint where things seemed to work better". would be nice to codify this without git commits

@danabramov.bsky.social · last yr.

has anyone tried making agents emotional? i feel like it would be nice to have some kind of a state machine that makes it more motivated to dig into something or kind of backtrack when it gets bored, vs getting excited at chasing the solution when it feels it's close etc

@danabramov.bsky.social · last yr.

ok so this is kind of very interesting. you need to actually "design" MCP tools to work well. for example the Lean one is actually getting helpful now that we're sorting the autocomplete output by relevance. but then it might help to include extra hints into the output: github.com/oOo0oOo/lean...

@danabramov.bsky.social · last yr.

if a bluesky fork with its own independent appview adds a feature, does it mean it can be first-class in that fork? not just a "client" feature but server too. you would index your own records in your own forked database, so in principle you can add any features. like "communities" or whatever