Nostr notes by dan

@npub19qk…vn3j plzzz ...

2026-11-01T17:56:37Z

Samuel (npub19qk…vn3j) plzzz
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreiebyeoknylcba5jchhyef75b5vviaza232u6ynh5bb4fx4g5tvfuq

by the way i'm sure others said it before, but cynically i ...

2026-10-21T23:21:04Z

In reply to nevent1q…4zus
_________________________

by the way i'm sure others said it before, but cynically i think a large part of subsidizing subscriptions is to lure professionals into area where humans alone can no longer drive the project (middle area in the article preview image)

which is .. like okay maybe we have to. i'm scared of outermost

in a sense, a program is not its source code, but a projection of ...

2026-10-21T23:16:35Z

In reply to nevent1q…qs3v
_________________________

in a sense, a program is not its source code, but a projection of a platonic "why this idea works" hyperprogram. we usually don't see the hyperprogram

the original developers/team usually maintain both. but the agents mostly engage with the projection, thereby losing the hyperprogram

i don't think it's unavoidable but we need better ways to ...

2026-10-21T23:11:25Z

In reply to nevent1q…4rde
_________________________

i don't think it's unavoidable but we need better ways to make the invisible (intended constraints, seams between the system, invariants, meta knowledge about which pieces are supposed to be orthogonal) visible so the agents engage with it

this was very good and articulated a lot i've been feeling ...

2026-10-21T23:10:25Z

this was very good and articulated a lot i've been feeling lately.

in formalization, this effect is even more noticeable: decoherence is punished by the agent being stuck in the maze for days when otherwise it would've made progress in 30 minutes

but i think this applies broadly to software

RE: https://bsky.app/profile/did:plc:dxmiuhk7peop3pzkyuqqta2f/post/3mxyqyugs222b

rip to the real one ...

2026-10-19T01:13:44Z

rip to the real one
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreieddcsshl4g7xjno4e7utopfxol6g5vwd7wz6juagymo4bc6a7rny

i’ve been doing some research on using Lean with AI to prove ...

2026-10-04T16:11:13Z

i’ve been doing some research on using Lean with AI to prove software properties (in my case, pieces of React). i’ll try to share lessons if it gets to a useful state. the main thing i’m learning is that, with this setup, proofs are *grown*. you can’t “just build them”

this too [The Curiously Playable ...

2026-10-04T14:48:39Z

In reply to nevent1q…fg6n
_________________________

this too

[The Curiously Playable Univers...](https://contraptions.venkateshrao.com/p/the-curiously-playable-universe )

fun read [Our Eukaryotic ...

2026-10-04T14:29:26Z

fun read

[Our Eukaryotic Moment](https://contraptions.venkateshrao.com/p/our-eukaryotic-moment )

[To grieve, or not to ...

2026-10-04T12:04:22Z

[To grieve, or not to grieve?](https://xenaproject.wordpress.com/2026/10/01/to-grieve-or-not-to-grieve/ )

[I am actually scared of ...

2026-10-04T01:43:55Z

In reply to nevent1q…q9jj
_________________________

[I am actually scared of linear...](https://www.youtube.com/watch?v=lonR5RnHEh8 )

Nostr event nevent1qqswq05c4vpcwd0hrnrratfzp4czq96t24ga0luxdjv3rn03gnvrkmszyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfgvah6et

2026-10-04T01:25:00Z

been unironically upping my p doom over the last few months

in my experience, AI is excelling at most technical work except ...

2026-10-02T11:50:45Z

in my experience, AI is excelling at most technical work except for ”what is the right next problem to work on”. for now, humans still have a noticeable advantage there

hey this album slaps [Pure Pleasure For The Piano - ...

2026-09-30T21:13:09Z

hey this album slaps

[Pure Pleasure For The Piano - ...](https://open.spotify.com/album/68XSF0H1PieWREfBOKPJXy?si=dLMh9Y9LRliVB1coCEsV6w )

Nostr event nevent1qqsr79vh90fp89marrs0fug6nxnr38lgwmvuy5uw30c08dun4df949gzyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfggvu468

2026-09-28T20:00:29Z

astra may be my favorite model now. it's a very weird one

this seems like it should be fun [The World as a ...

2026-09-25T19:51:39Z

this seems like it should be fun

[The World as a Hologram](https://www.youtube.com/watch?v=GHgi6E1ECgo )

[Analyse | Thom Yorke | From ...

2026-09-24T11:56:20Z

In reply to nevent1q…4z7a
_________________________

[Analyse | Thom Yorke | From Th...](https://www.youtube.com/watch?v=HPm7WJ4hznQ )

Nostr event nevent1qqsd0k0y5298krst5getj70crc7perhlrfwqe5jn59azwyxspvag0uqzyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfg9z6trx

2026-09-24T11:56:10Z

guys
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreib5vov3wvyfgnglvbbit46c3tf2vtwm43y2nbmx7pjrakld76a5yy

you also have to readjust Astra towards fidelity with whatever ...

2026-09-23T15:35:26Z

In reply to nevent1q…hmg6
_________________________

you also have to readjust Astra towards fidelity with whatever you initially got out of it.

it has a very unhealthy bias towards execution so you need to get it back to thinking deep about stuff

my current rating: - 5.6 Sol: dependable workhorse, sound ...

2026-09-23T13:50:52Z

In reply to nevent1q…ks53
_________________________

my current rating:

- 5.6 Sol: dependable workhorse, sound judgment, gets through the slog and stays on rails
- Fable 5.1: bulldoze through hard problems that Sol gets stuck on or fails to see bigger picture
- Astra 6: deeply understand the big picture, fundamental in-place reshaping, overseer

do NOT let Astra near code until it adopts a personality of ...

2026-09-23T13:47:35Z

In reply to nevent1q…p96h
_________________________

do NOT let Astra near code until it adopts a personality of domain nerd with high bar for excellence and taste

and ideally give it something that already does what you want end-to-end but maybe badly. it’ll fix it

ok i take it back Astra is good how to work with it: 1. get it to ...

2026-09-23T13:45:47Z

ok i take it back Astra is good

how to work with it:

1. get it to deeply understand what needs to be built. not “plan“ but like get it to be a domain nerd

2. only THEN, give it a usable past sloppy attempt and ask for excellence

3. ask it to use Sols for coding so it stays on strategy and taste

[Minna-no-kimochi (みんなのきもち) | ...

2026-09-23T01:21:53Z

[Minna-no-kimochi (みんなのきもち) | B...](https://www.youtube.com/watch?v=djGlyTcW30Q )

[POiSON GiRL FRiEND | Boiler ...

2026-09-22T20:21:34Z

[POiSON GiRL FRiEND | Boiler Ro...](https://www.youtube.com/watch?v=-aMjr0Ze868 )

"proof indigestion" ...

2026-09-22T00:29:55Z

"proof indigestion" [www.youtube.com/watch?v=PZRb...](https://www.youtube.com/watch?v=PZRb6NIki2w&t=294s )

[Terence Tao: SAIR’s Open Math ...](https://www.youtube.com/watch?v=PZRb6NIki2w&t=294s )

this is very interesting (timestamped) [Mathematics foundations ...

2026-09-21T23:17:46Z

this is very interesting (timestamped)

[Mathematics foundations proble...](https://youtu.be/R3LTVA3nay4?si=nSmDjbVNxS6ksviW&t=1238 )

re last point, can somebody from frontier labs please finally ...

2026-09-21T21:30:06Z

re last point, can somebody from frontier labs please finally drill polya’s “how to solve it” into the models thinking trace habits?

“we have to shift our position again and again“ is so obvious to how humans work but models refuse to do that because they race towards the finish line! so annoying

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3mw2nmzllws2g
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreieqmhpf45k5cz7s2wyjmt2s2jdlgsp7itwcmnodz52s53zwhzx4vu

quoting
note126u…pnxq

5.6 Sol is my guy. best model in town right now imo. maybe not the cleverest but at least it’s relatively dependable. especially if you give it a little bit of structure on how to work

5.6 Sol is my guy. best model in town right now imo. maybe not ...

2026-09-21T21:26:08Z

In reply to nevent1q…8ywp
_________________________

5.6 Sol is my guy. best model in town right now imo. maybe not the cleverest but at least it’s relatively dependable. especially if you give it a little bit of structure on how to work

that said Astra seems great in the first five minutes. like ...

2026-09-21T21:23:05Z

In reply to nevent1q…88g6
_________________________

that said Astra seems great in the first five minutes. like analyze this and that, propose how to think about the fix etc. just can’t actually let it do the work beyond a couple of steps

if you leave Sol overnight with a well-chosen goal, you might ...

2026-09-21T21:20:18Z

In reply to nevent1q…329h
_________________________

if you leave Sol overnight with a well-chosen goal, you might wake up to it done or at least see decent progress on it. if you leave Astra overnight, you‘ll wake up to it being 20% through three different misrepresentations of your goal

wish i could refund 95% of my Astra usage. keep giving it a ...

2026-09-21T21:16:38Z

wish i could refund 95% of my Astra usage. keep giving it a chance and it fumbles but it spends tokens way faster than Sol. a very disappointing release for actual implementation work

[Regina Spektor - Loveology ...

2026-09-21T19:35:38Z

[Regina Spektor - Loveology (Oc...](https://youtu.be/Qxg3DvApg34?is=jkCvjy-aoKeOpGHt )

thanks to Dr. Samuel Allen Alexander for making a video about my ...

2026-09-20T12:10:41Z

thanks to Dr. Samuel Allen Alexander for making a video about my post! [www.youtube.com/watch?v=CFkr...](https://www.youtube.com/watch?v=CFkrHlkrH24 )

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3mvsefn4rmk2g

[Conway's Conjecture AI-proved,...](https://www.youtube.com/watch?v=CFkrHlkrH24 )

quoting
note1nmn…wwna

i spent all of my free time for a month on this. it's done. so i wrote about it.

[How I Vibed a Proof of Conway’...](https://overreacted.io/how-i-vibed-a-proof-of-conways-conjecture/ )

“clearing the fog“ is maybe not a great analogy because it ...

2026-09-19T21:52:32Z

In reply to nevent1q…xtrc
_________________________

“clearing the fog“ is maybe not a great analogy because it implies there’s one concrete way something works. whereas it’s more like there’s a bunch of facts, and some are considered ”interesting if true (or false)“ because the hope is that the path reveals new path systems. a weird road helps less

my uninformed mental model of AI proofs in mathematics is that ...

2026-09-19T21:48:07Z

my uninformed mental model of AI proofs in mathematics is that they’re like lighthouses in the fog. the fog is still there, and clearing the fog is the primary value of the discipline. the lighthouses give a bit of an orientation but don’t clear the fog on their own.

annoying bsky app regression: pressing profile icon (or profile ...

2026-09-19T21:30:15Z

annoying bsky app regression: pressing profile icon (or profile posts tab) on web no longer invalidates it

another fantastic essay on the current thing RE: ...

2026-09-19T20:17:17Z

another fantastic essay on the current thing

RE: https://bsky.app/profile/did:plc:4i77tdu3riswwmtw6j67ufrj/post/3mvsnby3xa22z

great guest post by ...

2026-09-19T13:39:11Z

great guest post by [@3blue1brown.com@bsky.brid.gy](https://bsky.brid.gy/ap/did:plc:2ketphpobajcv4ujk5ncog4y )

[If math is more than proof, we...](https://terrytao.wordpress.com/2026/09/18/if-math-is-more-than-proof-we-need-to-better-celebrate-the-rest-of-it/ )

anyone hooked up Jev to any proof related stuff? can it be useful ...

2026-09-19T01:51:50Z

anyone hooked up Jev to any proof related stuff? can it be useful for Lean? i haven't learned much about it yet

a reply from Prof. Mantova! ...

2026-09-19T00:21:03Z

In reply to nevent1q…jrs5
_________________________

a reply from Prof. Mantova! [news.ycombinator.com/item?id=4976...](https://news.ycombinator.com/item?id=49761718 )

honestly this is the best kind of outcome i‘m hoping for with amateur-driven AI proofs: maybe they aren’t usable directly but can offer a couple of nudges that help the mathematicians uncover the actually interesting connections
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreig6oiz6bfkfprtzno72fc2m5fzxo4eeldi5f4ik2flojxovadi7mi

heh i like this ...

2026-09-18T18:33:24Z

In reply to nevent1q…pnrw
_________________________

heh i like this
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreigbwkcwffqchcnexvbgqipmrhp252ppb6m4shtw2fk5434cxw3dwi

we're on hnnnn RE: ...

2026-09-18T16:52:11Z

we're on hnnnn

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3mvsefn4rmk2g

quoting
note1nmn…wwna

i spent all of my free time for a month on this. it's done. so i wrote about it.

[How I Vibed a Proof of Conway’...](https://overreacted.io/how-i-vibed-a-proof-of-conways-conjecture/ )

i spent all of my free time for a month on this. it's done. ...

2026-09-18T14:19:38Z

i spent all of my free time for a month on this. it's done. so i wrote about it.

[How I Vibed a Proof of Conway’...](https://overreacted.io/how-i-vibed-a-proof-of-conways-conjecture/ )

[The most devious Sun Vulcan ...

2026-09-16T21:02:17Z

[The most devious Sun Vulcan pl...](https://www.youtube.com/watch?v=FvOVtW54NNc )

Nostr event nevent1qqsgvfzgnpkxg9cv5ptykqnghrrwj9tk4cwfvjljsjw6f3pav7dzjhgzyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfgjpulyx

2026-09-14T16:57:06Z

i like to think of monads as middleware for semicolons

i want something like ...

2026-09-14T14:17:13Z

i want something like [@rescript-lang.org@bsky.brid.gy](https://bsky.brid.gy/ap/did:plc:djs46wkphezpq5jilclfslim ) but for [@lean-lang.org@bsky.brid.gy](https://bsky.brid.gy/ap/did:plc:devnu4zs2viutk26clxykpbb ). that is, a JavaScript backend emitting reasonably idiomatic JS code. i wonder whether this is structurally impossible.

fractional calculus is so cool! if i understand correctly ...

2026-09-14T11:51:44Z

fractional calculus is so cool! if i understand correctly it's like when instead of a first derivative (f') or second derivative (f'') or say antiderivative (∫f dx), you can take ... half derivative

RE: https://bsky.app/profile/did:plc:zizetxb4lpxjoxrwab2bypfj/post/3mlw3cg2q7s2z

i'm slightly understating. i "know" what i'm ...

2026-09-13T23:43:14Z

In reply to nevent1q…d6e9
_________________________

i'm slightly understating. i "know" what i'm doing. i'm just surprised they're so powerful and handle everything i throw at them.

love when they switch into this persona ...

2026-09-13T23:40:47Z

love when they switch into this persona
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreide5cb2gfayykrtnrnurh2o6evgizc72aahzynildw2lsrzimkgdi

this is really a "year of linux on desktop" kind of ...

2026-09-13T23:38:19Z

In reply to nevent1q…5klh
_________________________

this is really a "year of linux on desktop" kind of situation for FP and a bunch of other things. the barrier to using these techniques is now very low so if you know what you want but don't know how to do it, it is no longer a problem

i still don't fully understand how to use them correctly but ...

2026-09-13T23:30:30Z

In reply to nevent1q…l56g
_________________________

i still don't fully understand how to use them correctly but i just tell llm "make this shit monadic" and it does exactly what i want

Nostr event nevent1qqsyuq9nxwks4xvflrawz64zhufehfv66x67czz06t3actnluhqecjgzyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfgycefs4

2026-09-13T23:29:19Z

monads are awesome. apparently i'm pretty late to this

just have to embrace it. 0 to 1, then throw away old context, ...

2026-09-13T20:35:38Z

In reply to nevent1q…c072
_________________________

just have to embrace it. 0 to 1, then throw away old context, clean up, then do another 0 to 1 with old 1 as a reference

RE: https://bsky.app/profile/did:plc:4vvaxwovyg4yo4n7dyyncaiu/post/3mvdkbvemcc2e

do it once badly but fully then restart with a clearer initial ...

2026-09-13T20:34:21Z

In reply to nevent1q…txyq
_________________________

do it once badly but fully

then restart with a clearer initial prompt (you know what you want better), let it think, finally point at the last implementation and give it freedom to redesign entirely from scratch but better

then do this again ++

learning this over and over and over. really best way to go with ...

2026-09-13T20:32:55Z

learning this over and over and over. really best way to go with frontier models

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3mtujwgnnzs2f

quoting
note1erv…dmj0

i think the main thing i learned with my llm math detour (and now see with speculative llm projects in general) is that telling llm to burn it all down and start from scratch when you learn an important lesson (or ask it what it wants to salvage) brings as much, if not more, value than construction

anyway the positive version of this statement is that astra is ...

2026-09-12T02:13:26Z

In reply to nevent1q…ftpq
_________________________

anyway the positive version of this statement is that astra is amazing at seeing patterns and extracting the essence from something that already works, but terrible at finding the right path if the problem isn't amenable to bruteforce by coding

i don't know what this says. also not sure if sol ...

2026-09-12T02:11:05Z

In reply to nevent1q…jaw7
_________________________

i don't know what this says. also not sure if sol would've done any better. (fable 5.1 was bad too.) maybe it's just i'm giving it hard problems where the correct answer is to "think really hard about both the problem *and* about a tractable way to approach the problem". while llm just wants to code

proving a relatively nontrivial thing: consistently bad output, ...

2026-09-12T02:09:52Z

proving a relatively nontrivial thing: consistently bad output, going offtrack, faking progress. i had to intervene and think with it through every step very closely for entire day

finally, proof compiles. "astra, find ways to simplify it". immediately, many big reductions, proof is much simpler

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3muutiiu4x22a

understated but both of these are kinda huge. ViewTransition is ...

2026-09-10T23:58:12Z

understated but both of these are kinda huge. ViewTransition is the first ever (!) first-class animation API in React. (it's powered by the browser API but is composable in a very Reacty way.)

and Fragment refs solve "merge refs" soup. you can now have natural apis like <IntersectionObserver> etc

concretely i mean it’s worse than useless if you leave it to do ...

2026-09-10T20:18:09Z

In reply to nevent1q…0zyc
_________________________

concretely i mean it’s worse than useless if you leave it to do things by itself. but a high touch “look at every file edit and constantly adjust its direction” approach works super well with it

using Astra reminds me of playing with Suzuki Escudo Pikes Peak ...

2026-09-10T20:16:56Z

using Astra reminds me of playing with Suzuki Escudo Pikes Peak Version in Gran Turismo 2. it was the most powerful car in the game so i would immediately run into the wall at every turn. but even with all the time i was losing this way, it was enough to be attentive to win every race

you can just tell astra to use file edit tool so you can see ...

2026-09-09T01:44:25Z

you can just tell astra to use file edit tool so you can see edits, and it obliges

i was on a podcast! RE: ...

2026-09-08T23:04:34Z

i was on a podcast!

RE: https://bsky.app/profile/did:plc:oxovi3avefouv4z4wacbtlmp/post/3muzoovawv22s

i was on a podcast! RE: ...

2026-09-08T23:04:34Z

i was on a podcast!

RE: https://bsky.app/profile/did:plc:oxovi3avefouv4z4wacbtlmp/post/3muzoovawv22s

weirdest model i tried maybe. i have no idea what it's ...

2026-09-06T18:29:13Z

In reply to nevent1q…smez
_________________________

weirdest model i tried maybe. i have no idea what it's talking about. but if you steer it appropriately, great results
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreiggpr525pcxu6ej6rzipbcsiovox6xjbdya44kmotmxn64bsfjr5u

astra is very hit and miss. can get complex stuff done right and ...

2026-09-05T23:04:01Z

astra is very hit and miss. can get complex stuff done right and then goes wildly offtrack

also season 2 is much better on rewatch and if you don’t have ...

2026-09-05T12:12:15Z

In reply to nevent1q…uncy
_________________________

also season 2 is much better on rewatch and if you don’t have to wait a week between each episode

less is more huh RE: ...

2026-09-05T12:09:36Z

less is more huh

RE: https://bsky.app/profile/did:plc:kft6lu4trxowqmter2b6vg6z/post/3mupxubgrls2c

love this solo ...

2026-09-05T10:10:42Z

love this solo
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreifz2y35vgcremhqd7iyyqhnm4oxd2wbeg7rtayemeplgo5znpyvrq

Nostr event nevent1qqs08q0px6aqm268hpxxl2cea2ad3ncxlt8adelxaevw9uu0xlpwzrqzyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfgl9ujx5

2026-09-05T01:19:41Z

silo just went into all time greats i think

one of my favorite parts of the project was making this ...

2026-09-02T11:48:26Z

In reply to nevent1q…kl23
_________________________

one of my favorite parts of the project was making this automatically generated proof map. here is the proof map for the headline result: [gaearon.github.io/conway-refin...](https://gaearon.github.io/conway-refinement/#/map/conway-refinement )
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreibnzpzy4pzvwvca5724fbzazg7ijwhyud2ra5gbspwtyraipo226e

how it started ...

2026-09-02T11:15:51Z

In reply to nevent1q…4vr2
_________________________

how it started
https://eurosky.social/xrpc/com.atproto.sync.getBlob?did=did:plc:fpruhuo22xkm5o7ttr2ktxdo&cid=bafkreian4ihjoocreygwddavtucph2s7f2q6bnm4rixy6exbfp5ziy7fbq

okay, it's out ...

2026-09-02T11:08:28Z

okay, it's out [github.com/gaearon/conw...](https://github.com/gaearon/conway-refinement )

RE: https://bsky.app/profile/did:plc:fpruhuo22xkm5o7ttr2ktxdo/post/3mudipgy2f22a

[GitHub - gaearon/conway-refine...](https://github.com/gaearon/conway-refinement )

quoting
note1h05…tr7f

preparing to publish my vibecoded math lean proof to github and i kinda feel nervous cause what if it's all wrong

it seems like my workflow was not the most optimal (i spent my ...

2026-08-30T23:07:29Z

In reply to nevent1q…3lyn
_________________________

it seems like my workflow was not the most optimal (i spent my full 20x claude and chatgpt every week, and used resets when they were available). but it definitely went beyond what i've been able to achieve with a oneshot so i should probably document what i've tried and what has helped

i've been working on it for a month and a week, it's time ...

2026-08-30T23:03:18Z

In reply to nevent1q…y66y
_________________________

i've been working on it for a month and a week, it's time to get it out. it is a bit difficult to get a confirmation of whether it is wrong because it's a niche area, so someone who can say for sure needs to be in the intersection of Lean users and niche experts (or maybe at least interested in it)

preparing to publish my vibecoded math lean proof to github and i ...

2026-08-30T23:01:35Z

preparing to publish my vibecoded math lean proof to github and i kinda feel nervous cause what if it's all wrong

found the magic prompt that fixes all my Lean. "walk the ...

2026-08-30T12:17:00Z

found the magic prompt that fixes all my Lean. "walk the abstraction layers one by one from the foundations all the way to the top, and at every point see if the proof matches how you would explain 'why' to a human. if not, assert underlying facts at the lower level and compose or lift them"

i'm also verifying that the tex proof it generates per ...

2026-08-29T18:58:39Z

In reply to nevent1q…0rzf
_________________________

i'm also verifying that the tex proof it generates per blueprint node actually contains references to every blueprint-included claim on the node's shallow dependency list. so it's forced to actually connect these in text

now i have it generate mermaid in README from that map, look at ...

2026-08-29T18:57:48Z

In reply to nevent1q…ua9j
_________________________

now i have it generate mermaid in README from that map, look at it, and decide whether each edge and node are legible to a mathematician, whether any important nodes are missing, and whether some nodes are standard bookkeeping. it already discovered a bunch of dead claims and simplifications

however right now i think the blueprint thing is a way more ...

2026-08-29T18:56:50Z

In reply to nevent1q…t6yl
_________________________

however right now i think the blueprint thing is a way more powerful method. it's like magic. in short i told it to use LeanArchitect to highlight essential nodes in the proof, and to state their titles and proofs in a way that's legible to mathematician.

oh yeah i don't have concrete ideas for what's a ...

2026-08-29T18:54:16Z

In reply to nevent1q…3phj
_________________________

oh yeah i don't have concrete ideas for what's a "good" way to do this. what we *want* to detect is when the choice of objects is unnatural and theorems convert back and forth in routine bookkeeping. i'd pose that problem to LLM and let it figure out some proxy metrics

if some junk in Lean can't be removed by syntax (section ...

2026-08-29T17:20:27Z

In reply to nevent1q…yk8q
_________________________

if some junk in Lean can't be removed by syntax (section variables, implicit parameters etc), better abstraction layering, induction, and so on, you can always ask llm if it wants to write a new tactic or DSL. usually you won't need this but in some cases this is justified

if the llm is stalling writing a proof and constantly patching it ...

2026-08-29T17:18:18Z

In reply to nevent1q…yzkv
_________________________

if the llm is stalling writing a proof and constantly patching it up, something is wrong. when abstractions are right, writing proofs is mostly easy and doesn't involve a lot of fiddling

you need to teach the llm to orient itself in the proofspace the ...

2026-08-29T17:16:06Z

In reply to nevent1q…rnx5
_________________________

you need to teach the llm to orient itself in the proofspace

the initial proofs and objects it will generate are often not optimal. ask it to write a script that shows which objects and theorems cluster together unnaturally, with the aim of noticing when the model is stated in "wrong coordinates"

everything is amplified proof checks (realistic example checking ...

2026-08-29T17:09:39Z

In reply to nevent1q…qmy5
_________________________

everything is amplified

proof checks (realistic example checking top level theorem/lemma applies) are kind of like integration tests, but quantifying over an infinity of cases.

bugs in the model are like infinity of bugs. they don't just break some little feature but completely block your progress

always. work. on. one. thing. at. a. time. llm by default will ...

2026-08-29T17:07:10Z

In reply to nevent1q…7gts
_________________________

always. work. on. one. thing. at. a. time.

llm by default will not do that, and in proof engineering doing >1 things at once gets you wrecked. it just can't make the right tradeoffs because it's trying to "finish" the proofs but what you want is an accurate model, not a typechecking wrong model

if the system is sensible, generally you should expect proofs to ...

2026-08-29T17:03:29Z

In reply to nevent1q…6jz7
_________________________

if the system is sensible, generally you should expect proofs to be simple and beautiful. so weird complicated proofs are a smell that you're missing APIs from lower levels

this one is funny but: even though it's vibed etc, if some ...

2026-08-29T17:02:43Z

In reply to nevent1q…07ag
_________________________

this one is funny but: even though it's vibed etc, if some definition doesn't make sense to you, ask LLM to explain it. it's very possible it's just completely wrong, or it picked the wrong seam for the abstraction. i always think "i'm too stupid to understand this" and then the model *was* wrong

it won't on its own write "clean" Lean so you have to ...

2026-08-29T17:01:42Z

In reply to nevent1q…v5w4
_________________________

it won't on its own write "clean" Lean so you have to tell it to declutter code by showing examples of obnoxious definitions. check out Mathlib and TauCeti locally and tell it to look for what clean Lean code looks like

it helps to explicitly tell it to ask you discriminating ...

2026-08-29T16:56:11Z

In reply to nevent1q…7cj2
_________________________

it helps to explicitly tell it to ask you discriminating questions when there are different ways to model stuff with weaker/stronger guarantees. otherwise it may get the strength wrong and you end up with a bad model. once you have a bad model everything else is bad until you fix it

for each incremental change to the theory, you have to nudge it ...

2026-08-29T16:55:08Z

In reply to nevent1q…rct7
_________________________

for each incremental change to the theory, you have to nudge it to prove strongest theorems it can state about objects at every layer. this gives you powerful APIs for layers above and also catches nonsense and mistakes in modeling

you need to give it a bird's eye view of the big nodes. so ...

2026-08-29T16:50:05Z

In reply to nevent1q…tald
_________________________

you need to give it a bird's eye view of the big nodes. so that it sees the actual shape of the tower of models and proofs. i'm using LeanArchitect + blueprint into mermaid, maybe other approaches work too.

also if you have some underlying shape (like a tree) you're ...

2026-08-29T16:49:19Z

In reply to nevent1q…pmcu
_________________________

also if you have some underlying shape (like a tree) you're proving things about, you better create some reusable machinery for induction. otherwise it'll happily write every induction from scratch at every level of the proof and then get completely lost in the cases

llm can run wild writing proofs but it'll likely prove wrong ...

2026-08-29T16:48:31Z

In reply to nevent1q…hf9k
_________________________

llm can run wild writing proofs but it'll likely prove wrong things about wrong objects. and then at some point it gets stuck because the whole system doesn't carry the right properties through the right abstractions. so you need to teach the model to "pause" and think in stages at each layer

Nostr event nevent1qqs2lqh2ww74v0zuw75da76yldrylp5snlcwqc7mnlmp2scquh6kl5czyrsx9djrvczy7dntnedcs07ym2gjx3wtk65ga7fn298u48thlnxfglwa7j0

2026-08-29T16:47:18Z

learning a ton about proof engineering in practice

fantastic tool for generating Lean proof maps ...

2026-08-29T15:33:35Z

fantastic tool for generating Lean proof maps ("blueprints") directly from the Lean code [github.com/hanwenzhu/Le...](https://github.com/hanwenzhu/LeanArchitect )

highly recommend for helping LLM make sense of the broader structure of the proof and finding simplifications

[GitHub - hanwenzhu/LeanArchite...](https://github.com/hanwenzhu/LeanArchitect )

hear me out surreal numbers — сверхчисла omnific ...

2026-08-28T17:03:15Z

hear me out

surreal numbers — сверхчисла
omnific integers — всецелые

oh and you're gonna need this ...

2026-08-27T23:58:34Z

In reply to nevent1q…xesl
_________________________

oh and you're gonna need this [github.com/cameronfreer...](https://github.com/cameronfreer/lean4-skills )

[GitHub - cameronfreer/lean4-sk...](https://github.com/cameronfreer/lean4-skills )

just tell your LLM to spec the problem in Lean first

2026-08-27T15:08:17Z

In reply to nevent1q…spfn
_________________________

just tell your LLM to spec the problem in Lean first

prediction: Lean usage will explode in serious software projects ...

2026-08-27T15:02:59Z

prediction: Lean usage will explode in serious software projects in the next five years