11:14 |
phf |
since everyone appears to be dead, |
11:14 |
phf |
happy new year, log |
| |
~ 1 hours 32 minutes ~ |
12:47 |
awt |
happy new year! |
| |
~ 28 minutes ~ |
13:16 |
shinohai |
Feliz año nuevo |
13:16 |
shinohai |
$vwap |
| |
~ 3 hours 19 minutes ~ |
16:35 |
signpost |
happy new year y'all. |
16:37 |
signpost |
http://logs.bitdash.io/pest/2022-12-30#1019998 << w/e sbcl uses by default for seeded randomness via sb-ext:seed-random-state, though asciilifeform recommended xorshift because cheap to reseed, which happens very often in encode/decode |
16:37 |
bitbot |
Logged on 2022-12-30 23:49:44 phf[awt|deedbot]: signpost, what's the prng you ended up using? |
16:39 |
signpost |
will probably start using xorshift on the python side too if it proves worthwhile. currently w/e default, have tried several which were all about the same. |
16:42 |
signpost |
and in lulz of old, wrecker gets wrecked |
| |
↖ |
| |
~ 36 minutes ~ |
17:18 |
phf |
http://logs.nosuchlabs.com/log/pest/2023-01-01#1020083 << yeah, i ran into this issue with the luby prototype. cmucl uses XOROSHIRO, sbcl MT19937, python also uses mersenne twister, but whether it corresponds to sbcl's mt remains to be seen |
17:19 |
dulapbot |
Logged on 2023-01-01 16:36:18 signpost[asciilifeform]: http://logs.bitdash.io/pest/2022-12-30#1019998 << w/e sbcl uses by default for seeded randomness via sb-ext:seed-random-state, though asciilifeform recommended xorshift because cheap to reseed, which happens very often in encode/decode |
17:19 |
bitbot |
Logged on 2022-12-30 23:49:44 phf[awt|deedbot]: signpost, what's the prng you ended up using? |
17:20 |
phf |
well, since i'm just exploring the concept, i'm just `sending` vectors along. i.e. i punted and defer to you on account of you spearheading this whole idea |
17:22 |
phf |
the whole shared prng is an optimization anyway, so that you don't have to send the values, can just send seed instead. to understand the behavior saving the values works good enough™ |
17:25 |
phf |
but ftr prng choice has to be part of reference implementation, since it has to produce identical sequences between server and client |
| |
~ 16 minutes ~ |
17:41 |
* |
asciilifeform waves, wishes folx happy newyr |
| |
~ 54 minutes ~ |
18:36 |
signpost |
phf: in onlinecodes two seeds need to be sent along. one's a "message ID" which generates the association between message blocks and their "aux" checksum blocks. |
18:36 |
* |
signpost can't see a reason not to use the hash of original payload for this, two birds one stone. |
18:36 |
signpost |
sorry, two classes of seed. |
18:37 |
signpost |
second class is the check-block-id, which is the product of the second (final) pass of encoding. one of these gets sent along with each block. |
18:38 |
signpost |
that seed produces the edges of that check block which point to the message+aux blocks the check block is made of. |
18:38 |
signpost |
since the sequence of check blocks can be infinite one doesn't wanna do seed+offset for rng state for a given check block |
18:39 |
shinohai |
In Luke Jr lulz: https://archive.is/YlIyj |
18:39 |
* |
signpost was thinking a "literate programming" pass over the py implementation would be nice |
18:39 |
signpost |
it's really not a complicated algo; simple enough my dumb head could implement |
18:40 |
signpost |
re: standardization on prng, seems simplest that is capable of producing the desired distribution wins, and would wager xorshift will do fine. |
18:40 |
* |
signpost will grunt the py version forward to updated vpatch + blog poast hopefully sometime this week |
18:41 |
signpost |
shinohai: I scooped ya! http://logs.bitdash.io/pest/2023-01-01#1020018 |
18:41 |
bitbot |
Logged on 2023-01-01 16:42:29 signpost: and in lulz of old, wrecker gets wrecked |
| |
↖ |
18:43 |
shinohai |
bwahahahaha |
| |
~ 31 minutes ~ |
19:15 |
asciilifeform |
lol, had nfi he was still around |
19:15 |
asciilifeform |
( and, at least to some rubes, chief lolcow? (... what happened to gavin ? ) ) |
19:18 |
asciilifeform |
signpost: any prng can shit out the desired distribution if you transform the output. ( afaik the key attributes of a prng usable for lubyism are -- long period (i.e. not absurdly short), uniformity, and speed , in that order. ) |
| |
↖ |
19:19 |
asciilifeform |
( seems like signpost has a workable one already ? ) |
| |
~ 2 hours 18 minutes ~ |
21:38 |
crtdaydreams |
http://logs.bitdash.io/pest/2023-01-01#1020038 << lol "happy new years" |
21:38 |
bitbot |
Logged on 2023-01-01 18:41:16 bitbot: Logged on 2023-01-01 16:42:29 signpost: and in lulz of old, wrecker gets wrecked |
21:39 |
crtdaydreams |
LMAO |
21:43 |
crtdaydreams |
dood had it stored on coinbase, whatdya expect? |
21:43 |
signpost |
as I recall the guy was endlessly full of shit. there's no telling what happened, except that he's butthurt, which is funny. |
21:45 |
signpost |
http://logs.bitdash.io/pest/2023-01-01#1020042 << figured producing a particular distribution faithfully would depend on the quality of prng output, but probably here period being way higher than number of check blocks to realistically be produced before heat death of universe is fine? |
21:45 |
bitbot |
Logged on 2023-01-01 19:18:56 asciilifeform[jonsykkel]: signpost: any prng can shit out the desired distribution if you transform the output. ( afaik the key attributes of a prng usable for lubyism are -- long period (i.e. not absurdly short), uniformity, and speed , in that order. ) |
21:46 |
* |
signpost cribbed some guy's impl of the alias method for the lisp encoder; I expect numpy does something similar under the hood for biased selection |