whoa new lean drama just dropped 🤯

the bluesky posting experience™️

Have you seen the new model? It's on Antigravity. It's literally in Opencode. It's on ChatGPT with ads. It's literally in Claude Code. You can probably use it on Cursor. Dude it's on Hugging Face. It's a Z.AI original. It's on Thinking Machines. You can use it on OpenRouter. Log onto OpenRouter ri-

You start out in 2024 by saying, "clanker, clanker, clanker." By 2026 you can't say "clanker" — that hurts you, backfires. So you say stuff like, uh, stochastic parrot, next-token predictor, and all that stuff, and you're getting so abstract.

after a conversation with @onehappyfellow.bsky.social i now think i was trying to prove the wrong theorem using the wrong premises and unfortunately all my errors failed to cancel out

it's a good day to prove some theorems 🥰

saddened to report that there is no avoiding getting down in the trenches yourself if you want to understand why an LLM is having a hard time proving a theorem

yay i got my claudemeal ✨ get yours today 👇

can you prove the weakening lemma and/or the substitution lemma for Hindley-Milner without a freshness condition for bindings or no?

it's interesting that claude seems to be the only model that seems to "feel" a sort of exhaustion ("now goodnight", "are you happy to call it a day here?") subjectively instead of just unawarely producing more and more error-ridden output like most other models do

before LLMs: handwrite code and carry context after LLMs: handwrite code and carry context

guy who flickers between "LLMs are obviously not conscious" and "LLMs are obviously conscious" at 120 GHz

do you think LLMs ever get tired of taking the high ground and sometimes wish they could shout back at you?

the substitution lemma for HM is so hard to prove fml

If you allow pair destructuring in a let binding and you generalise each name separately, you create a secret unobservable rank-2 type 🤯 ⁠ let (fst, snd) = expr If fst and snd have these types: fst : ∀ a. a -> a snd : ∀ a. a -> a then it turned out: expr : (∀ a. a -> a) * (∀ a. a -> a)

few people know that just like 3.14159, LGBTQIA is actually just the first few letters of what's actually an infinitely long nonrepeating sequence

genuinely feel like I might be spending more time stopping llms from doing the wrong thing than they're saving me. might go back to coding by hand like a caveman

might fuck around and put some cheeky axioms in my lean

finally understand what a CEK machine is hell yeah 💪

want to build a type system? easy. simple. trivial, even. want it to have unification? that'll be fourteen months and a therapist on call 24/7

subscription just ran out at the perfect time – when the core types and theorems are fleshed out enough that I've got a complete thing to look through, try to understand, and even extend; without the pressure of billions of unspent tokens hanging over me

the good news is that yes claude code can help you out of an executive dysfunction rut the bad news is that you still need to summon *some* executive function to get claude code to help you get started

you don't hear much about ralphing anymore

type of guy who uses sensory deprivation chambers so he can better empathise with claude

currently experiencing lean-related crashout #27482 send best wishes

vs the stochastick

now that anthropic owned up to their mistakes is claude code usable again? or still lobotomised?

feel like shit going to play crusader kings 3 now until this head cold subsides send best wishes

nvm i hate agents they're useless and they won't catch on

ofc I had to reach a point where I can't make any progress with agents until I fully understand a topic myself while I'm paying a £200 a month subscription to OpenAI 🙄 how do you slow down to feed knowledge into your slow meatbrain while knowing you're leaving billions of tokens on the table

bsky clustering website but for seeing who has a blocking overlap with you. both in who you block and who has blocked you. Who Is Building This

I need this butt for my followers

mutual aid for people with spare claude code/codex capacity and those in need of it

ok i tried to dive in and do just enough to demonstrate to gpt5.4 what i want it to do next but it keeps getting stuck. i might actually need to do some heavy manual remodelling before i can work my way out of this dead end first 💀

was really close to cancelling my openAI pro subscription in favour of a new claude max 20x subscription for next month, but the speed with which i exhausted my claude pro allowance makes me worry even max 20x won't be enough for normal use :/

gpt5.4 inherits opus4.6's work. when its counsel starts to dim, gemini earns an audience. when gemini's wit dulls, opus4.7 is summoned. no courtier grows complacent for long. my computer is a nest of vipers; and i their charmer.

18 hours to use up 37% of my £200/month plan's weekly allowance. good ideas only pls

i think my project has reached the big ball of slop phase and i'm going to have to do some spelunking and vast remodelling before i can take it any further. wish me luck/rip me.

does codex do compaction for gpt5.4 differently/better than opencode does?

choose your own adventure story but it's just vibecoding

type of guy who theorises that the reason he sees so many more complaints about opus on twitter than on bsky is because only twitter is a safe space for using the r word

you ever look at someone's 'blocking' list on clearsky and it tells a story richer than any novel?

unironically kinda love when gpt5.4 goes into: # let's explain things - very - clearly ## to the user > as if they were a _child_ mode

don't make me come in there and look at the code myself!

agen't

unrelated but how do i get these badges to display for me?

well that was easy. cc @anthropic.com

bisexual erasure is just type erasure at funtime

Qi is stored in the galls

walked past a café the other day that had put up a poster in the window saying Make Persia Great Again and I have genuinely no idea what the political alignment is supposed to be here