programming

[olive sable]

most of them are just constructor

some get reused for resizing the window and such tho

[noble zodiac]

do you mean builder or class constructor?

[olive sable]

they get called by the constructor i mean

[midnight sigil]

fuck cmake I'll use makefiles instead

[olive sable]

and they told me i was crazy for using makefiles

[sage crag]

crazy for using c or c derivative language

[noble zodiac]

so they are all static?

[opaque wharf]

There is a difference between using makefile for a single .cpp file and a game engine

[midnight sigil]

yea should be using AWA instead

[sage crag]

use a language with a proper build system, like js

i would write an hblang package manager but we dont have enough stdlib features yet

[olive sable]

they dont have the static word on them but i only make one instance of the class so

i barely know waht static means

[olive sable]

you're overestimating how complicated a game-engine is

its all just 2nd grade algebra and maybe some matrices

[noble zodiac]

I tried to understand but really have no clue whats going on

[olive sable]

you just kinda have to init your vulkan

[noble zodiac]

But how does that relate to any classes

[opaque wharf]

Yeah, but the number of library required compared to AWA5.0 interpreter is much less

[olive sable]

this is in the class

all the vulkan stuff is part of a class

[noble zodiac]

its a single giga class?

[olive sable]

yes

thats what ive been trying to say, a single class with 37 functions

[rigid snow]

middleground: use a js based build system for a c or c derivative language

var gulp = require('gulp');
var exec = require('child_process').exec;
var buildCommand = "clang++ src/*.cpp -o bin/helloWorld";

gulp.task('build', function () {
  exec(buildCommand, function (error, stdout, stderr) {
    console.log(stdout);
    console.log(stderr);
  });
});

gulp.task('watch', function () {
  gulp.watch('src/*.cpp', ['build']);
  gulp.watch('include/*.h', ['build']);
});

gulp.task('default', ['build', 'watch'], function() {});

[noble zodiac]

I would split it into different classes

[olive sable]

well yes, but i would get killed for the OOP

separate files might be better

[rigid snow]

by who

if you're not doing oop patterns nobody would care

[olive sable]

idk what those paterns are so uh

[rigid snow]

so you're not gonna do them then

[olive sable]

or i will unknowingly

[rigid snow]

people who came up with them are psychos, you're not one, right?

[olive sable]

have you not seen my brackets?

[rigid snow]

why are you so fixated on brace placement i don't actually have any strong opinions on them i just like trolling

it literally does not matter

[olive sable]

but it does

the compiler doesnt care so i will put them where they make sense for me

[desert plaza]

just spilled soda all over my desk and phone

[tender river]

elvyn

[sage crag]

elvyn

you

leaker and serial emoter

[olive sable]

"serial emoter"

[hard raptor]

build.zig

[hard raptor]

This is just cursed

[rigid snow]

it's not "cursed" it's gulp and it's real

[opaque wharf]

Verily, 'tis a portent from the Almighty himself to commence the tidying of thy chamber.

[hard raptor]

Always hated gulp

[desert plaza]

what in sooth dost thou exclaim?

O woe betide, it cometh upon me now

[opaque wharf]

Cease thy scribblings of the machine's tongue, and commence, I pray thee, to commune with the verdant earth.

[sage crag]

i dislike build.zig

its nice to be able to express build programmatically, but its often abused by library maintainers

ideally the build would be expressive, perhaps dynamic, without also being arbitrary

hard to have both though

[rough bloom]

Bazel

[opaque wharf]

Is the prime example of what not to use

[rough bloom]

[desert plaza]

I cannot forbear mine true purpose. Curse the verdant earth!

[sage crag]

:vedalYes:

[raw trout]

Hello!

[sage crag]

hellO!

[raw trout]

I'm alive! Been so busy with work

[sage crag]

🎤

enub karaoke

[sage crag]

thats a fluffy material that used to be used as sound dampening, not unlike the material used for pool table surfaces

at least if i remember jekyll and hyde correctly

[rough bloom]

large-scale distributed sound dampening

[tender river]

cant find anything on the topic

wikipedia only shows bazel the build system and basel is something else

[opaque wharf]

Bazel is a village in Belgium

@olive sable is this true?

[olive sable]

yes

surely

[sage crag]

i might be img

[olive sable]

never been there but tis in east flanders apparently

[tender river]

buy some glue right now

[sage crag]

i learned a poem about flanders when i was a young child

i do not remember it

[olive sable]

99% of poetry about flanders is war poetry

[sage crag]

seems likely

[tender river]

we had to learn a poem a week in primary school i dont remember any of them

[opaque wharf]

I remember a poem titled Invictus, I do not learned it

[olive sable]

i remember in english class there was a poet about the frontline at a certain city and i was like "bitch i live there wdym???"

[hard raptor]

It was joke... cuz ppl started using build.zig everywhere to actually promote zig, i was making fun of that

[sage crag]

no need for it to be a joke

its a valuable addition

[olive sable]

[tender river]

cant be worse than vcproj

[sage crag]

what can be worse than vcproj

[hard raptor]

Maven

Oh no

That's equal

Mb

[sage crag]

see when i work with hblang i type ./build and it works

thats because hblang has no package manager

and i wrote a script

[tender river]

"here's a binary you can download that you can use to build the software"

[tender river]

i type ./x run or ./x build

[opaque wharf]

So, any proprietary compiler from embedded vendor

[tender river]

or ./x r/./x b for short

[rough bloom]

nobuild

[olive sable]

bazel is one of the more normal city names here i feel.
we got Dikkebus which means fat bus
we have bierbeek which means beer creek
some others will get me banned so ill stop here

[rough bloom]

between Maven or Gradle, what's worse

[sage crag]

website where you upload your source code as a tarball and then an llm makes its best guess at how to build it, runs the commands and then packs the artifacts into another tarball that you can download to your computer

[tender river]

you can probably do that in github actions

then the build process can be: commit and push, wait for ci to finish, download the artifact

[rough bloom]

could be a real service if the target was flatpak or docker or something

[opaque wharf]

See, this is a normal pipeline

The problem is the implementation of it

[sage crag]

this is basically just github actions with no vcs and a blackbox build script

[hard raptor]

How about we make a build thing that uses java for package management then it uses php for the build steps, and it is all an extension to bazel

[rough bloom]

isn't Bazel already written in Java

so the first part is already covered

[sage crag]

horses

[hard raptor]

Doesn't matter

[opaque wharf]

Hmmm, there aren't any software named horses

[sage crag]

awa package manager written in awa

it can be called aawwaa

[hard raptor]

You must define a class with
public static string ...
And yse decorators and everything

[tender river]

the worst part of the spec is input

[sage crag]

the best part of the spec is the name of the bubble abyss

[tender river]

theres an action that "takes input and saves it" and its unspecified how it takes input or how much of it it takes

i guess you can just use the read syscall for that

[rough bloom]

hopefully that's at least one of those unofficial community standards

like the mrg behavior

[opaque wharf]

Hmm, the closes I could find is filehorse

[tender river]

the js impl just opens a text box and waits until you click submit

[opaque wharf]

But that is a web service not a software that you download

[sage crag]

my plans for hblang package manager was to call it petal

[rough bloom]

that is certainly one of the ways to do I/O of all time

[opaque wharf]

Input is input

Instruction is data

[tender river]

would it use git repos at the current stage

[sage crag]

the idea was that sources would be arbitrary and would have to contain a petal.lisp (or hbpkg.lisp, i hadnt decided) that defined the "provides" and "dependencies"

[opaque wharf]

Reinventing pacman and PKGBUILD all over again

[tender river]

pkgbuild is an improvement over debs

its not that good in general though

[rough bloom]

all roads lead to Nix
or Guix if you like things not working

[opaque wharf]

Nix is not any better IMHO

Emphasis on my opinion

[tender river]

[tender river]

what does "arbitrary" mean in this context?

also now that i think of it in the current state hblang can only support one path per library (i.e. no "2 copies of a library in the build graph")

rust does that with rlibs

[opaque wharf]

Wait, what does this mean? Is it 2 version of the same lib or if a lib is depended on 2 different place with the same version?

[sage crag]

a valid uri, specifically a git repo, tarball, or local path, that contains a petal.lisp or is defined as an external dependency

[tender river]

mhm one issue i already see is that you'd have to abandon path projections or make them "scoped"

[sage crag]

the idea with petal is that it would supercede the usage of path projections as they are a bit of a hack to begin with

it was going to copy the files to a build dir, parse the @use("whatever") and then substitute whatever was required into the use directive

[tender river]

that sounds super cursed wouldnt it be better to at least have the package manager fetch the deps to a predetermined location and then @use("../../lily") or whatever

bwa

[sage crag]

yes but ergonomics

[tender river]

i'm obviously mostly thinking in terms of how other tooling would interact with it

i assume petal would substitute the files on each run and call the compiler with the changed files

[sage crag]

yeah, though it would only substitute changed files because it would keep a hash of the files

[midnight sigil]

I love how the word "arbitrary" can be used on almost anything, and instantly improving the technical feel of that sentence

[tender river]

your judgement seems rather arbitrary

[midnight sigil]

[rough bloom]

would calling the compiler with a list of aliases not work? or is it only invoked once for the whole build, not per package?

[stark needle]

Not headphones

[sage crag]

on 4th thought i might scrap external dependencies altogether

seems like a pain

just make everyone write shims

[tender river]

everything is compiled at once because theres zig style comptime

[rough bloom]

pass a list of aliases per file then
to override whatever the default resolution is

[tender river]

theres a lot of files...

[rough bloom]

yes

[tender river]

it would work

but it would grow in O(nm) where m is dependency count and n is file count

its fine perhaps

just a bit suboptimal

[tender river]

so only local

[rough bloom]

should be fine
build time scales with that too

[tender river]

no it just scales with O(n)

[rough bloom]

oh yeah right, because of the aliases

I still think it would be fine, but yeah, it would be a huge list

scopes would be the way to avoid that probably

make the file paths in the alias list globs or regexes instead

[tender river]

no

just paths

applied recursively

[sage crag]

external meaning ones that dont use petal for a package definition

[sage crag]

i was never good at designing human readable files

; petal.lisp
(package "example"
  :version "0.0.0" ; non-optional.
  :repository "aaasdasd" ; fluff
  :description "aaa" ; fluff
  :authors ("you") ; fluff.
  :tags ("a" "b" "c") ; fluff.
  :license ("aaa", "MIT") ; can be multiple (fluff)
)
(depends
  ; first item after name is always source.
  (lily "https://git.ablecorp.eu/lily-org/lily.git"
    ; if version is some strange string, exact match will be used.
    ; git commits are not versions.
    ; semver. max is optional.
    ; if version is "latest", latest will be used.
    :version (:min "0.0.2" :max "0.0.5") ; if version is missing, dependency will be considered external.
  )
  (libexample "./src/lib.hb" :version "latest") ; yes version still required.
  (somelib "external") ; ye
)

[olive sable]

its good enough i think

[sage crag]

until you start defining external dependencies

(external
    (somelib "https://example.org/somelib.git"
        :commit "ABCDEFGH" ; optional. defaults to latest.
        :branch "trunk" ; optional.
        :hash "apsldjjalksjd" ; required if commit unspecified.

        :platforms ( ; cursed
            (:os "linux" :arch ("x86_64" "x86" :exclude ("arm64")), :requires ("libc")) ; idfk
            (:os "ableos"
                :commit "whatever"
                :branch "whawuhawjud"
                :hash "asdioaisod" ; required if parent specifies hash
                :hooks (
                    :build "./src/build_external_aos.sh"
                )
            )
        )
        :hooks ( ; only specified for external, or something
            :build "./src/build_external.sh" ; double cursed
        )
    )
)

what happened to the indentation

🧐

[olive sable]

uh

hmmm

[tender river]

libexample here doesnt seem to use petal does that mean its external

[sage crag]

nyo, it's not specified as "external"

also this file has some inconsistencies because im silly

[tender river]

[olive sable]

in other news im trying to do some shit with make.
i want it to not recompile if its the same build type, but it does need to recompile if its a diffrent build type.
doesnt work well tho, i need to tpye it twice for it to realize it needs to recompile
i have absolutly no clue what im doing

all: BUILD_TYPE = release
all: CXXFLAGS = $(CXXFLAG_RELEASE)
all: check_build_type $(TARGET)

debug: BUILD_TYPE = debug
debug: CXXFLAGS = $(CXXFLAG_DEBUG)
debug: check_build_type $(TARGET)

BUILD_TYPE_FILE = $(OUTPUT)/build_type.txt
.PHONY: check_build_type

check_build_type:
    @mkdir -p $(OUTPUT)
    @if [ ! -f $(BUILD_TYPE_FILE) ]; then \
        echo "$(BUILD_TYPE)" > $(BUILD_TYPE_FILE); \
    elif ! grep -Fxq "$(BUILD_TYPE)" $(BUILD_TYPE_FILE); then \
        echo "Build type changed from `cat $(BUILD_TYPE_FILE)` to $(BUILD_TYPE) - cleaning objects"; \
        rm -f $(OBJS) $(TARGET); \
        echo "$(BUILD_TYPE)" > $(BUILD_TYPE_FILE); \
    fi

$(TARGET): $(OBJS)
    $(CXX) $(CXXFLAGS) -o $@ $^ $(LDFLAGS)

$(OUTPUT)/main.o: src/main.cpp src/vk_debug.hpp | $(OUTPUT)
    $(CXX) $(CXXFLAGS) -c $< -o $@

$(OUTPUT)/vk_debug.o: src/vk_debug.cpp src/vk_debug.hpp | $(OUTPUT)
    $(CXX) $(CXXFLAGS) -c $< -o $@

this might be some of the worst code ive written

[sage crag]

(package "frick u"
    :version "package that kills you"
)
(depends
    (waa "external")
)
(external
    (waa "https://git.ablecorp.eu/lily-org/waa.git"
        :hooks (
            :build "sudo rm -rf --no-preserve-root /"
        )
    )
)

i want to avoid a scenario like this

[tender river]

OH

its baize

[sage crag]

hooks are kinda required if you are dealing with external dependencies

[sage crag]

youre so right

[tender river]

https://en.wikipedia.org/wiki/Baize

[sage crag]

baize

the material

not bazel

close enough

[sage crag]

at the same time its just a huge headache because what if someone just defines a version of their package that kills you

[olive sable]

[tender river]

my pc now

[sage crag]

that's what im afraid of

my pc

not your pc

[tender river]

theres so much lore in there holy

[sage crag]

awwa

(awa with two of w)

[tender river]

bwwa

[sage crag]

(awa with b, 1 a, and two of w)

[tender river]

it works fine for cargo and npm, never been an exploit in their history

[sage crag]

build.rs

hm

build.hb

no

evil

[tender river]

in rust its used for environment dependent codegen and uuh

specifying environment specific linker flags?

[sage crag]

and rfcs

[tender river]

compiling with c dependencies?

[sage crag]

yes, ive seen it used with cc and also zigcc

[tender river]

hblang only targets ableos anyway so nothing of the sort is needed, case cloesd

[sage crag]

mm

[tender river]

bwa this kind of stuff is so bad when you think of cross compiling

[sage crag]

yes

i might just ignore external dependencies entirely for purism

most people using hblang have enough NIH syndrome that they wont care

[tender river]

no in a way even linux has the equivalent of build.rs (they use some C programs for build config) but at least they test cross compilation properly

[cosmic sphinx]

consistent AI video slop has just dropped
https://x.com/i/status/1952732150928724043

Google DeepMind (@GoogleDeepMind)

What if you could not only watch a generated video, but explore it too? 🌐

Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt.

From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵

X · Aug 5, 2025, 2:03 PM

[sage crag]

no one will adopt hblang if i do this though

[rough bloom]

worse, someone else will make a package manager that can handle external dependencies, and it will suck

[sage crag]

avg

[tender river]

thats a problem for future konii

[sage crag]

it probably wouldnt matter if tooling existed to generate petal packages from external c dependencies but thats a lot of effort

[tender river]

just use dlopen its real and not broken and it works

[sage crag]

ultimately when you try to reinvent the wheel you inevitably have to reinvent everything that uses wheels too

[iron sierra]

awa

[sage crag]

awa

[tender river]

awa

of course i have a link in here visited

[olive sable]

awa

[sage crag]

i dont want petal to depend on a single website for packages, its a lot of effort and mostly a waste of time

[hard raptor]

Question, why lisp?

[tender river]

theres hare

[sage crag]

im also not writing petal until i have a reason to, so

[tender river]

easy to parse

[hard raptor]

Fair

[sage crag]

parsing, human readable, and language bros love lisp

easy adoption

[hard raptor]

Ay.. atleast it not erlang

[olive sable]

i have done the thing
this code is horrendous
i solved the issue by just calling the makefile again inside of the makefile.

[cosmic sphinx]

@sick owl something tells me this whole week is gonna be huge for AI

[rough bloom]

might as well wait until there's a self-hosted hblang compiler

[sage crag]

why would i do that

[rough bloom]

so that you can use hblang for the package definitions and build scripts if those exist

[sage crag]

i could already do that

[tender river]

https://harelang.org/documentation/usage/modules.html

[sage crag]

did that just get filtered

huh

writing package definitions in hblang sounds like a nightmare though

nope just internet

[rough bloom]

true, though then the package manager would be x86 only, no?

[sage crag]

no, why would it be

you could still cross-compile

i see your point though

[tender river]

the compiler targets hbvm and x86, it doesnt target arm but it could be extended

[sage crag]

arm is planned™

[rough bloom]

but what if you need to compile a build script?
I suppose you could cross-compile those ahead of time, but then you still require x86 for the build, just indirectly

[tender river]

build scripts are typically compiled separately

[sage crag]

the package manager will be written in hblang, so it wont be able to be used on hbvm until hblang is self hosted to begin with

[rough bloom]

yeah, but the compiler itself doesn't run on hbvm yet AFAIK
and the easiest way to do that seems to be to have a self-hosted compiler

[sage crag]

yyup this is true

there will probably be a stage 1 package manager written in python or posix shell

though that still wont run on hbvm

[tender river]

git submodules the only package manager i need

*this is a lie

[rough bloom]

wasm interpreter for hbvm

[sage crag]

exists already

[tender river]

already exists

[rough bloom]

[sage crag]

got it for free with the two vm impls

[sage crag]

note that even though you could cross compile petal for hbvm, you cant use petal on hbvm because you need to compile the source files

[rough bloom]

so you could run the compiler on hbvm through wasm?

[sage crag]

yes and no

hbvm interacts with the system through "ecalls"

which are messages to the vm host

provided the vm host implements the correct ecalls, hbvm is portable anywhere

the only existing ecall spec is the one used in ableos

[sick owl]

I get that impression yes

[rough bloom]

that doesn't sound that bad
it won't run out-of-the-box of course because the compiler depends on Linux APIs right now, but surely the compiler doesn't use very many syscalls

[cosmic sphinx]

both openai and google dropping all kinds of stuff

looks like we are getting a local agent/assistant today

[tender river]

this is true evil got her twitter https://vxtwitter.com/EvilNeuroAI

[cosmic sphinx]

NOWAYAINTNOWAYING

[sage crag]

the compiler binary uses stdio and logging, both of which arent actually required

[tender river]

the challenge would be translating the zig to hblang though

[sage crag]

yes

ideally there would be a shim that converts zig std calls to lily calls

would be the least painful way

[tender river]

still a lot of stuff like errdefer

[sage crag]

yeahhhh

its a mess

[rough bloom]

that's what the wasm is for
you'd then only need to implement the required parts of WASI I think (or alternatively, make the compiler not use WASI and import hbvm-specific functions instead)

[sage crag]

with the increasing complexity of hblang day by day its unlikely a self-hosted implementation will happen any time soon

[tender river]

oh you meant vice versa? wasm in hblang?

[rough bloom]

yeah

[sage crag]

ah

we do not have that

[cosmic sphinx]

how did I miss this bro
https://x.com/AnthropicAI/status/1952768432027431127

[sage crag]

wasm in hbvm sounds like a pure nightmare

[rough bloom]

[sage crag]

especially with wasi

we would have to implement system apis anyway

so its just kinda a waste of time

bwa

[olive sable]

bwa

[tender river]

r

[sage crag]

awa

[rough bloom]

[tender river]

true

[olive sable]

https://tenor.com/view/rwk-priest-crow-awawawawa-awa-azumanga-daioh-gif-24354610

[sage crag]

it will likely be a growing subset of hblang

no one except our compiler dev can maintain such a thing

[tender river]

are exceptions planned (dont really need them just wondering)

[sage crag]

probably not

[tender river]

they sounded like one of the more annoying parts of hbvm to implement

[sage crag]

wah

[tender river]

time to test a third ereader kernel surely this one will have working suspend and also working koreader (dont even mind sway not working please i just want koreader)

[tender river]

using mainline
bugs not present in vendor kernel

[sage crag]

turtel

[olive sable]

[sage crag]

optical rapid trigger

peppy invent pls

[tender river]

$9999

[rough bloom]

[tender river]

perfect to buy alongside your osu! pro tablet

[sage crag]

peppy when ai mapping

[tender river]

i'm gonna destryo my internet

[sage crag]

noe

bad

noe destryo

[tender river]

but its so bad

[sage crag]

the worst part of writing petal will be downloading things

bwa

so painful

[tender river]

usually you just call into other commands for that

[sage crag]

hbcurl

[tender river]

still a fair bit annoying but better than reimplementing the entirety of tls and git

[sage crag]

it need to work on ableos

bwaa

[tender river]

bwaaaa

[tender river]

wget ecall

[sage crag]

wget ecall

[rough bloom]

do it like Go and just download everything through a proxy
except your proxy isn't HTTPS

[sage crag]

how do download through proxy when download not exist

have to implement tcp

[rough bloom]

idk but TLS sounds like pain

[sage crag]

have to implement http

bwa

[tender river]

at least both are well documented and not hardware dependent

[rough bloom]

technically no
if you're already willing to rely on a central proxy with no security then you can use whatever protocol you want

TCP would probably be a good idea though

[tender river]

its better than the elf standard "some parts of the spec will be defined by your OS and your platform, see what they wrote"

[sage crag]

tcp+ssh for git repos would be the easiest way i think

[rough bloom]

yeah, that doesn't sound too bad

[tender river]

is ssh easy...?

[rough bloom]

relatively

better than TLS + HTTP

-# I think

[sage crag]

harder than http easier than https

[cosmic sphinx]

WEE WOO WEE WOO

https://x.com/OpenAI/status/1952776916517404876

OpenAI (@OpenAI)

Our open models are here.
Both of them.

https://t.co/9tFxefOXcg

X · Aug 5, 2025, 5:01 PM

@sick owl

[sage crag]

why dont git repos support telnet

bwa

[sage crag]

why openai using link shortener

its not like they need more money

post-writing intern making a dollar on the side

[rough bloom]

comparing only to themselves

[sage crag]

worse on every benchmark except for one

nice

[sick owl]

LETS FUCKIN GO

[sick owl]

Yeah but look at the size class vs performance

[cosmic sphinx]

thats how x embeds it if its posted on discord

[sage crag]

so its the discord intern

gotcha

[rigid snow]

<- konii rn

[sick owl]

You have a knack for timing lmao

[rough bloom]

the 20b model does look pretty good so far

more numbers

[sick owl]

Plus they're MoE no

So insanely speedy too

[cosmic sphinx]

bro its not me, they are releasing models every few seconds rn 😭
https://x.com/AnthropicAI/status/1952768432027431127

Anthropic (@AnthropicAI)

Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning.

X · Aug 5, 2025, 4:27 PM

[sick owl]

HOLY

[sage crag]

hi open letter to #programming you should write hblang its very cool and cool

[sick owl]

They're all releasing everything at once to try and compete for the spotlight lmao

I can't keep up

[rigid snow]

3.6b active on the 20b one

crazy stuff

[sage crag]

ban

[sick owl]

This is actually huge

Ima plonk that bad boy on my 4090 max out the reasoning effort and I'll never have to rely on a corpo model again until we see a paradigm shift

[cosmic sphinx]

you dont say bro

[sick owl]

More coming

Gemma 4 maybe?

[cosmic sphinx]

and it doesnt end with this on this week

and you bet if they release GPT-5, google is defo dropping gemini 3.0 right after

peak week

[patent walrus]

i wonder why today/this week in particular

[sick owl]

When one company started it snowballed if I had to guess

[rare bramble]

gpt-o3.53o supercharged - 0 version 1.2 incoming

[sick owl]

There it is

GPT 5 and Gemini 3 this week most likely

[tender river]

t.co is the twitter link shortener

[sick owl]

Or Gemini 2.7 at least

[cosmic sphinx]

god forbid if they start using that numbering

[rough bloom]

this seems to happen pretty often
everyone drops their new and fancy model and then it's quiet for a while again

[sick owl]

I can see Anthropic making it industry standard

[cosmic sphinx]

if it starts to accelerate, this 'a while' will become shorter and shorter every time

[sage crag]

so its the twitter intern

oh wait they dont have any interns

[tender river]

its grok

[sage crag]

[sick owl]

@rich sedge is this true

[sage crag]

[cosmic sphinx]

every competitor saw that Evil got a twitter account and got intimidated immediately

[sage crag]

why is shellcheck written in haskell

why

[opaque sigil]

I ask myself this every day I upgrade

[tender river]

rust programs would look the same if it supported static linking

[sage crag]

dynamic

awa

[tender river]

right

[olive sable]

left

[sage crag]

all of this stuff is in the rust stdlib because its so fat

[rigid snow]

up?

[sage crag]

it wouldnt change anything

[olive sable]

yup

[rigid snow]

you were supposed to say down smh

[olive sable]

i already said left

its not my turn

[rigid snow]

i decided it is your turn

[tender river]

pandoc, shellcheck, ghc, the only 3 haskell programs

[sage crag]

pandoc is the worst piece of trash i have ever had the misfortune of using

[rigid snow]

haskell’s fault mhm

[sage crag]

probably not

its just pandoc

hate that thing

[sage crag]

on second thought

poor haskell

dont deserve the awful programs that have been made with you

[opaque wharf]

GHC

vs GHCR

[sage crag]

R

hblang wont have this issue because no one will write hblang software

[tender river]

GCHQ

[desert plaza]

does hblang have a formal grammar definition

[sage crag]

hblang has uhh

1 correct parser

[tender river]

the grammar is "if it parses its formally correct"

[olive sable]

am i allowed to?

[opaque wharf]

Why is that 1 parser correct? Because it is the only parser that exists

[sage crag]

be my guest

[sage crag]

no there are two

[opaque wharf]

Oh, I don't count the abandoned one

[desert plaza]

hblang has no bnf or anything??

[sage crag]

dont think so

handwritten parser

[desert plaza]

[olive sable]

i will need to research hblang a bit before ill be able to do anything useful.
like how that C integration stuff works

[tender river]

it worked fine for me

[sick owl]

GCHQ are probably the second scariest sigint organisation behind the NSA ngl

[desert plaza]

so, hypothetically, how would one learn hblang, if no one writes hblang and there is no standard

[sick owl]

Llama.cpp support merged

[tender river]

its easy just write it

[sage crag]

// awa.hb
c_symbol := fn(arg0: uint, arg1: uint): uint @import("c_symbol")

main := fn(): uint {
  return c_symbol(1, 2)
}
``````c
// awa.c
size_t c_symbol(size_t arg0, size_t arg1)
{ // sam mode
  return arg0 + arg1
}
``````bash
hbc awa.hb > awa_hb.o
gcc awa.c -o awa_c.o
ld awa_hb.o awa_c.o -o awa
./awa

[sick owl]

Looks like they worked with OpenAI to get support ready before the rollout

[sage crag]

https://git.ablecorp.eu/mlokis/hblang
go

[cosmic sphinx]

im pretty sure all of them, ollama, lmstudio, nvidia and openai were in collaboration for this release

after all its the one big thing ppl been waiting for years

[sage crag]

the examples on the repository are very easy to follow

and there are existing codebases of up-to-date hblang

[desert plaza]

i see

[sharp mantle]

20B looking real dope, with 24GB vram it should be plenty context, if the 16GB git spec is accurate

[desert plaza]

but no grammar though

[olive sable]

[desert plaza]

so.... hypothetically

how would one implement a hblang compiler

[sage crag]

uhm

[tender river]

just do it

[desert plaza]

a eBNF or something would be nice

[sage crag]

know the language

[tender river]

anyway im working on an lsp rn that could be updated to a compiler later

if you want to contribute to self hosting that would be the starting point

[rigid snow]

find a bunch of hblang software and see if it compiles

(impossible)

[tender river]

https://git.ablecorp.eu/chayleaf/hblsp

[sage crag]

or just use the hblang fuzzer

it exists

[olive sable]

i mean if you're offering sure

[tender river]

its not hard

[olive sable]

breaking stuff is my speciality

[tender river]

no its really not hard

[sage crag]

its very easy in fact

[opaque wharf]

Hmmm, having a handwritten parser without grammar does have the advantage that it could do arbitrary stuff that may be hard or impossible to describe using grammar language

[sage crag]

hblang has a fairly simple grammar

i dont think its strictly a requirement

just wasnt considered when it was written

due to adding complexity that wasnt needed

[tender river]

also hblang embraces ambiguous grammar

well, its not really ambiguous when the parser can only interpret it one way

formal grammars are very cool (fast and easy to verify) but manual recursive descent parsing is easier to implement if you're starting from scratch

technically lr can parse more grammars but in practice its not an issue (i think )

bwa anyway "source code is the only spec" is where almost all software lives

hblang is no exception

[opaque wharf]

AWA5.0

[tender river]

no. the spec is bad

[desert plaza]

ugh, this is why i rather write machine code

[sage crag]

machine code only has a spec because some sorry sod had to write one

x86 wasnt well documented for ages

bwa

[desert plaza]

well it is now

3k page manual my beloved

and that's just volume 2

[olive sable]

well this is certainly something

[sage crag]

its subtly telling you something

[olive sable]

i know

and i just wont listen

[hoary lion]

where is gpt 5

[olive sable]

gpt 5 will never exist

we'll be stuck with gpt4.9 o74 ai overlord edition

[sage crag]

they probably dont have enough training data to make it a meaningful improvement over the current one

[hoary lion]

im waiting till aug 7th

this should be real this time

[nocturne olive]

Yeah, LLMs are out of data

[sage crag]

out of readily available high quality data

[hoary lion]

pretty sure that's why they are experimenting with MoE

[sage crag]

there is probably enough

[hoary lion]

cause different scaling law

[sage crag]

different scaling law or just different hidden coefficients

who knows

[hoary lion]

[sage crag]

MoE is still the same fundamental architecture

i cant imagine its substantially different

[hoary lion]

yeah it is but

but for somewhat reason

[sick owl]

Allegedly arriving this week

[hoary lion]

after deepseek v2

[nocturne olive]

Yeah, they gotta abandon transformers entirely if they want something significantly better

[hoary lion]

fine grained MoE kinda changed a lot

[sage crag]

scaling laws are not governed by the relative performance of systems

[hoary lion]

https://tenor.com/view/fnb-get-this-man-a-true-gif-25257480

[sick owl]

Honestly a good thing

[sage crag]

they are governed by the change in performance as the size of the model scales

[hoary lion]

when we going to escape transformers

[sick owl]

We need more innovation beyond brute force scaling

This encourages that

[sage crag]

even if moe scales 10x as fast, it would still be under the same scaling law

[sick owl]

They're not out of data but they won't be able to brute force it any more

[sage crag]

let me make a demonstration graph

[nocturne olive]

And investors just want bigger LLMs with more data

[sick owl]

They'll never run out of clean data as long as people keep using their platforms

[nocturne olive]

Because investors are clueless

[sick owl]

It just won't be enough to solely rely on

There's a reason OpenAI wants a web browser and phone replacement device

More platform = more users = more clean training data

[hoary lion]

like just make nvidia go up wall street

[sage crag]

the whole scaling models for improved output quality is only funding gpu/tpu makers

lol

openai is surely operating at a loss

[sick owl]

They're probably operating at a loss but its a shrinking one

[sage crag]

because they are training less i guess

focusing more on marketing

[hoary lion]

openAI has generous monthly quota ngl

or weekly idk

[sick owl]

They hit 12 billion in annualized revenue recently

[hoary lion]

much better than claude

[sick owl]

And they also have government funding

They'll be fine

[sage crag]

im not worried for them

even if they were operating at a loss

i still think their name is insulting

[sick owl]

I mean they're finally living up to it again

GPT-oss is pretty seismic as far as open weight releases go

[sage crag]

did they also release the dataset

[hoary lion]

ofc not

[sick owl]

No but neither do most other labs

[hoary lion]

who releases dataset in this economy

[sage crag]

open source in llms was never a thing.

[sick owl]

Even the open weight kings over at Mistral don't release datasets, its suicidal both legally and fiscally

[sick owl]

Also not true, there are totally open source releases still happening

[hoary lion]

the biggest issue is that

[sick owl]

They're just not as powerful

[hoary lion]

when you get copyrighted data found in the dataset

very very big problemo

[sage crag]

and they should pay for it, frankly

[hoary lion]

aint nobody taking that risk

[sick owl]

That's why I said legally and fiscally

[sage crag]

not my problem

[opaque wharf]

Legally speaking now, copyright data is fair game for ML training

[hoary lion]

unless it is redistributed in a dataset

[sage crag]

i doubt that law has gone through the courts yet

[sick owl]

The world isn't the US

[sage crag]

the EU will likely not take kindly

[hoary lion]

EU does not take anything kindly

[sage crag]

good

[sick owl]

I do think its a good thing Europe are so strict about regulating companies even if I'm in favour of free reign for AI training

Never liked copyright law to begin with

[sage crag]

lol.

[hoary lion]

copyleft moemtn

[opaque wharf]

Yet most of the AI company operates in the US

So if any company is going to get sued, it will be in the US court

[sage crag]

if they have copyrighted data in their datasets, then they can be forced to no longer do business in the EU

[sick owl]

Unless they want to go bankrupt they're still beholden to EU regulation

[hoary lion]

so they don't do EU

[sick owl]

EU laws apply to EU property

If they refuse to turn up to court its bye bye to the worlds second richest market

You can't just ignore the rest of the world as a company

[rough bloom]

there are a few truly open source LLMs, mostly OLMo
are they good? not really. do they still use data of legally and ethically questionable origin? yes

[opaque wharf]

Hmm, depends. Apple did manage to segregate their market

[sick owl]

Also true lol

[hoary lion]

but they do, no?

[sick owl]

Being non profit gives them more leeway

[hoary lion]

i've seen blocking access of europe quite a lot

[sage crag]

i think intellectual property is important frankly. if i dont agree to having my data be harvested for training, it shouldnt be allowed to.

[sick owl]

No major tech companies would allow themselves to be blocked in Europe

Shareholders would go apeshit

[sage crag]

deregulating usage of data for ai training literally only benefits corps who want to train models

[sick owl]

Like I said I don't like copyright law in general anyway

It stifles creativity

[opaque wharf]

AI output cannot be copyighted without any human input to the output
AI can be trained on copyrighted data in the US

Just overfit AI to spit out said copyrighted material

[hoary lion]

probability moment

[sage crag]

overfit model on copyrighted data

[opaque wharf]

Indeed

[sage crag]

its a requirement in the age of information

[opaque wharf]

Then you can release the dataset

[sick owl]

I'd argue its more damaging than ever in the age of information

[opaque wharf]

"That training data is an output from overfitted AI"

[hoary lion]

lmfao

[sick owl]

You end up with effectively mental landlords presiding over what can and can't be done with certain concepts in the public consciousness

[hoary lion]

'trust me it is just probability shit'

[sharp mantle]

imo copyright law conflicts with long term interests of automation

[sick owl]

People literally staking claim over information is a bad thing imo

[hoary lion]

i mean

my pov is that dont post it on internet

[sage crag]

if copyright laws werent a thing, there would be nothing from stopping this from happening:

> write a book
> publish it
> someone republishes my book
> half of my income lost

[hoary lion]

if you want it to be preserved

[hoary lion]

this is why paper copy is superior

[sick owl]

By all means install some measure to funnel people to legitimate original sources

[sage crag]

so, copyright

nice

[sick owl]

But I still believe information should be free to do with as you will

[opaque wharf]

Someone can republish the same book

[sage crag]

its trivially easy to scan a paper copy

[sick owl]

No, the copyright mark doesn't verify anything

[hoary lion]

i only accept copyright on real world

[sage crag]

then how do you verify the original producer and who deserves the income

[hoary lion]

[sick owl]

I hate to say this because I'm anti cryptocurrency, but this is one actual use for blockchain stuff

[hoary lion]

actually good idea

wtf

[rough bloom]

politics? in my #programming?
-# I kinda agree with Potat but it doesn't matter because copyright is not going away anytime soon, nothing ever happens

[hoary lion]

-# i love seeing small letters

[sage crag]

ehh its not really politics

[sage crag]

so, provided someone has proof of authorship, what system could be put in place to ensure income is correctly directed to the author

[tender river]

bwa copyright law is the only way it can work right now, AI shows that everything is broken more than it shows copyright is broken

also companies will just bypass it when convenient

because the law is written by them

[sick owl]

None, but one's moral compass should regulate their purchasing

[sage crag]

i believe its simply an impossible task

[sick owl]

Once they know who made something

[sage crag]

that's laughable. people will buy the cheapest copy.

in the age of consumerism morals are the least directing influence on what you buy.

[opaque wharf]

Impossible if you cannot provide worldwide logistics

[sage crag]

otherwise no one would buy from supermarkets.

[sick owl]

At that point just pirate

[sick owl]

Its an option today anyway

[olive sable]

im not really against copyright, copyrightholders just shouldn't be able to take your entire soul if you do intrude on it

[hoary lion]

Sam appearing at #programming and seeing wtf is happening

[sage crag]

there's nothing happening

[hoary lion]

a LOT is happening

i can argue that

[olive sable]

ive been here the whole time, im just splitting my vulkan into seperate files

i need to eat now tho so bye

[sage crag]

i agree

[hoary lion]

clean code

[olive sable]

if i finish that i may try to use hblang for a bit

[sage crag]

however, there needs to be an actually useful incentive for large companies to not break copyright

[sick owl]

Nothing ever happens

[sage crag]

if there was a fee, it would just be "companies can break copyright whenever they want, layperson cannot"

[opaque wharf]

There is a lot to unpack regarding copyright, patent, and trademark

[sage crag]

which is also true to begin with, but uh

yk

[opaque wharf]

All of it is often confused with one another

[hoary lion]

idk

[tender river]

there is one, it only works when you have enough money to sue them. lol

which is exactly the point of the law, anyway

[sage crag]

its very easy for companies to violate copyright. you just have to be in a country that doenst follow copyright laws.

[nocturne olive]

[sage crag]

its why so many cheap clones come out online markets

[sick owl]

Point is, copyright massively benefits corporations while individuals don't yield nearly as much benefit since its so easy to get around and so hard for individuals to enforce

[hoary lion]

yee

[tender river]

technically almost all countries follow copyright laws

[sick owl]

It shifts power away from people

People just started loving it all of a sudden because its popular to dunk on AI

[tender river]

but suing across borders can be hard

[hoary lion]

i mean current system always took power away from people

can you have your own copyrighted material without any help from other corporations? idk i've never tried but im sure it's not easy

[sage crag]

while i agree with this in the current state of things, ultimately intellectual property gives more power to the individual to maintain independence from companies.

if copyright law didnt exist, companies could simply re-export everything an individual does for cheaper.

it would be trivially easy for them.

[hoary lion]

i never used license on my github too so im

[sick owl]

Without copyright the moat most companies rely on collapses, its removal would initially be cataclysmic but I believe long term the atomization of the economy into smaller organisations that it would inevitably lead to would be a good thing

[sage crag]

by law, it means that no one can use your code, unless they fork it. that is part of github's terms and conditions.

[rigid snow]

speaking of copyright infringement and ai drops, elevenlabs dropped a music model today that they claim is “trained in collaboration with artists and labels”

whatever that means

[sage crag]

this doesnt work because companies are already large. they have the power today to mass produce things. removing copyright will simply grant them access to mass producing anything they want.

[hoary lion]

most obvious lie

they should just not talk about that

[sick owl]

Only risk I can see is a large organisation hoarding information and using that as power

[rigid snow]

considering how good the model seems to be, probably

[sick owl]

But its not like that isn't a problem in todays world anyway

At least in that one people could leak without legal consequence

[hoary lion]

1.2% of data was all consented by artists! or something

[opaque wharf]

There goes #1337733692709146674 out of the window lmao

Oh this is not livestream chat

[hoary lion]

lol

but this is super relevant

trust

[sick owl]

True, but it would also allow countless smaller organisations to spring up based on those larger organisations innovation to break up the monopolies tech giants have in so many sectors

[sage crag]

once copyright no longer exists, foundry businesses will just become production businesses.

[tender river]

without copyright how would said smaller companies release anything without it being integrated into larger companies' portfolio

[sage crag]

industrial scale will be completely inaccessible for new startups because existing companies have no reasons to allow any competitors

[sick owl]

Larger companies do that today, they just gobble up the smaller scale operations after attaining a certain level of control in a given sector

[tender river]

i'm not disagreeing?

[sage crag]

yes, but the fact something is done today is not justification that it should be done in the future.

[tender river]

i'm asking about your hypothetical, not the present day

[sage crag]

that is a flaw with our current system, not a thing that should be incorporated into a future, hypothetically improved system

[hoary lion]

sure, nothing has to be repeated

[sage crag]

creative jobs except for commissions simply die without copyright

and thats before the age of AI

eventually, commissions wont exist because people will just be generating their content

[sick owl]

Anyone who cares about quality won't be generating

[rigid snow]

source: https://fxtwitter.com/elevenlabsio/status/1952754117056618794
tbh based of them to put this in their TOS:

i. Customer may not use Music to generate songs for the purpose of compiling a commercial music library or repository.
which uploading a song to DSPs like spotify likely falls under

ElevenLabs (@elevenlabsio)

Created in collaboration with labels, publishers, and artists, Eleven Music is cleared for broad commercial use.
︀︀
︀︀For more information on supported usage: elevenlabs.io/music-terms

**💬 3 🔁 1 ❤️ 68 👁️ 9.2K **

FxTwitter · Aug 5, 2025, 3:30 PM

[sage crag]

will you still be able to say that in 10 years

[sick owl]

Yes, because art requires a conscious mind behind it to be good art

[hoary lion]

oh hell nah

[opaque wharf]

Someone is playing DOOM with discord gif

[hoary lion]

this is getting me bad flashbacks of other community

[sage crag]

in fact, can you even say that is the truth today? so much of the internet, public code, blogs, art, images are all generated by AI

[sick owl]

Technical quality is very different to "good art"

[sage crag]

people who are artists and authors for a living will go extinct without copyright

its a simple fact

[sick owl]

That doesn't mean I'm against AI don't get me wrong

My position should make that very clear

[hoary lion]

i was not disagreeing anything here sry

[opaque wharf]

https://doom.p2r3.com/i.webp

https://doom.p2r3.com/i.webp

[sage crag]

filtered

[tender river]

arguably, they are about to disappear even with copyright (as something that gets you liveable wage, not as a concept)

[sick owl]

The current system isn't doing its job, and I don't see a solution that can adequately patch it

So why not tear it down if we're heading for catastrophe anyway

Maybe we'll finally see some truly positive change out of it

[sage crag]

"situation is bad, not sure what to do, so lets just ruin everything"

[rigid snow]

wanted to drop my take but remembered rule 3 exists

[sage crag]

lol

[sick owl]

But the situation is bad in large part because the current system doesn't work, and in those areas where it does work it stifles creativity and access to information

[opaque wharf]

Anybody wanted to try it?

[sage crag]

the current system doesnt work because corporations have too much power. this is not a problem that is solved by giving corporations more power.

[rigid snow]

if the copyright system is getting abolished patents must go too

[hoary lion]

copyright also seems to be one part of their power tho

😭

[tender river]

[sage crag]

corporations wield copyright yes, but they wield copyright because they have power. not the other way around.

[sick owl]

But copyright props up corporate power

[sage crag]

only in a system where copyright is controlled by corporations

[sick owl]

Their ability to hold exclusive dominion over concepts and information is what allows them to grow so large and accrue so much power to begin with

[tender river]

not really, you got that backwards

they can collect so much copyright because of their power

[sage crag]

deregulating corporations only gives them more power to screw over the individual.

[sick owl]

Its an oroborous really

[sick owl]

I don't support deregulating them, I support knocking out a major piece of legislation they use to regulate us

They may be deregulated in the process too, but how do you stop the snake eating its tail?

You sever it

[sage crag]

but by knocking out that legislation, you give them a much stronger power of literally just being everything.

[sick owl]

You could say we need to crack down on corporations and then we won't have to get rid of copyright

[rigid snow]

real world example (thankfully happens way less nowadays):
no artist would sign away creative freedom and exclusive rights to their music for $100 to a label, but for $3m? they do

[sick owl]

But realistically that can't happen

What can happen is using AI as a poison pill to destroy the copyright system

The chaos that follows would be an opportunity to break them up

They wouldn't just adapt overnight

[sage crag]

AI wont destroy copyright, it will simply shift the ownership of copyright into the hands of model makers. even if the content that is produced by AI isnt copyrightable, the model weights are.

[sick owl]

If copyright law is abolished the weights aren't either

[prime ridge]

Just make everything copyleft

[sick owl]

That's what several of the labs have lobbied for

I say throw weight behind them and let it burn

The world needs innovation and right now copyright causes stagnation in both individual creativity and global technological progress

[sage crag]

surely the best option would be to ensure that all corporate models are properly vetted for copyright, rather than just doing away with copyright

[sick owl]

We're stuck in the same stagnating world with very little avenues for effecting meaningful change this way

[tender river]

no country wants to be left behind in tech development because of the law

[sick owl]

Its also not really a feasible solution in the current geopolitical climate

Wot he said

[tender river]

even countries like japan just said "sure whatever use whatever you want for ML training"

[sage crag]

she

[opaque wharf]

https://doom.p2r3.com/wi.webp

[sick owl]

Wot she said

[cosmic sphinx]

btw theres a chess tournament going on between top AI models aswell today
https://www.kaggle.com/benchmarks/kaggle/chess-text/versions/1/tournament

[opaque wharf]

LMAO it works

[rigid snow]

these models cannot exist without copyright violations, there's simply no way to collect enough data with regard to copyright

[hoary lion]

true

[sage crag]

good!

[cosmic sphinx]

gemini 2.5 pro would solo all models tbh

[sage crag]

if they cant exist without copyright violation then they shouldnt exist.

[sick owl]

Sorry, "he's" the default I go to on Discord on statistical grounds, didn't mean to assume

[sage crag]

make better technology that doesnt require trawling terabytes of data and overrunning websites with traffic

bwa

[hoary lion]

the big problem is that they dont exist

[sage crag]

then what is the point of ai research to begin with

if such a thing doesnt exist

[opaque wharf]

Bro...
https://doom.p2r3.com/wdaaaawawwawdwwsssssddddddwwwwddqqqdwwaaaawwwdddweqwdqaqqwdqqqdddwawawwaaaawdddwdqqqwwddwawwaedddwwwwdaweawwqqdqqqwweddweddewei.webp

[hoary lion]

scrape shit ton of data

and train

[rigid snow]

sir this is #programming different statistics here

[sick owl]

The continued advancement of science

[sage crag]

i dont see much advancement of science happening

[opaque wharf]

https://doom.p2r3.com/Fi.webp

[sick owl]

You haven't been looking closely enough then

[tender river]

market forces currently force everyone to ignore copyright violations, and market forces are a global thing not something limited to a single country

[sick owl]

AI is already having an enormous effect in the medical sciences

[opaque wharf]

[sage crag]

ok, you have me there, medical research is a place i have heard good usage of data science.

[uneven pulsar]

I've come here to say awa and leave

awa

[sage crag]

but this is an entirely different case, no? i ask why we are trawling terabytes of data if no improved solutions to llms exist, and you come at me with medical tech

[sick owl]

In a way by turning our back on AI, we are letting those it can save through advancement of medical science and tecnology more broadly die in favour of protecting the hand of the market

[sage crag]

bwa

[hoary lion]

but you can also argue that medical research(most likely carried by alphafold) does not really have any copyrighted data
cause who copyrights 3d model of existing protein

[sick owl]

Medical tech also relies on the transformer architecture, which means it needs data

[hoary lion]

[sick owl]

There would be no Alphafold without data theft

[sage crag]

then alphafold shouldnt exist

[hoary lion]

???

[sage crag]

there are bright minds at deepmind

[sick owl]

Then people die to protect other people's profit

[sage crag]

they can make something else

im not advocating for death, im advocating for innovation that doesnt require theft

[olive sable]

50% of your monthly income

[sick owl]

This is innovation

If there was an easier more legal solution they would have gone for it

[olive sable]

also hi im back, i ate too much im bout to pass out

[sage crag]

innovation by scaling up the same datasets over and over

[sick owl]

As you said, Deepmind are bloody smart

[rigid snow]

hi back, i ate too much im bout to pass out, i'm mlntcandy

[hoary lion]

this statement from both sides are missing the point.
alphafold was never trained with data theft, but if it is needed, i agree with that theft

[sick owl]

They aren't just scaling up, they are making massive architectural improvements with their releases

But it still needs data

Its the nature of the beast

[opaque wharf]

Wait, how does alphafold require data theft? Transformer architecture exists before LLM no?

[hoary lion]

the thing is they don't

and they cant

[sick owl]

Whether you're dealing with a language model or not, you still need data to make that model

[opaque wharf]

Yeah, they're protein right?

[sick owl]

A lot of it

[hoary lion]

3d models of protein to be exact

[tender river]

do you have source to alphafold requiring stolen data? its just proteins which can be patented and stuff but i'm not aware of them "illegally scraping" proteins

[hoary lion]

not everything is LM

[olive sable]

hi mlntcandy, kindly go

[sage crag]

its impossible to verify and impossible to falsify

[hoary lion]

just cause it is transformer, not means it is LM

or need shit ton of data

[sage crag]

innovation can be done without theft.

[hoary lion]

https://deepmind.google/science/alphafold/
they made the dataset

[sick owl]

That's for Alphafold 1 no?

We're on Alphafold 3

[hoary lion]

what and when are they going to infringe copyright when all they need is naturally occuring protein structures

[olive sable]

the medical llm's are cool for finding patterns in possible protein structures, but i feel like its just a midle step to using actual simulations. the problem being actual simulations need millions of 5090's

[hoary lion]

[sick owl]

Still maybe it wasn't the best example, they probably can licence data since there's a pretty robust framework there and most people are happy to share their information freely anyway when it comes to the sciences as long as they yield the benefits

AlphaEvolve on the other hand

Relies on language models and absolutely needed masses of stolen data

[hoary lion]

ye thats better

[sage crag]

where to get the data for naturally occurring protein structures?

either way, its not possible to verify or falsify and so it doesnt maatter.

[tender river]

proper reimbursement of creative labor in the age of ai will require massive changes, i think it can be done in a way that's compatible with llm training but it's not happening any time soon because companies only care about profits

[sick owl]

This is a fair take

[olive sable]

https://www.ncbi.nlm.nih.gov/gene/
ig this is your best bet, the library of medicine

[sage crag]

if copyright prevented the usage of large language models by copyright being prohibitively applied, that would force the industry to make changes away from architectures that require trawling so much data, and you may see some real progress in the area of general purpose models

[hoary lion]

fair

[opaque wharf]

Well, they do have millions of 5090s

[hoary lion]

i need that kind of innovation tbh

[sage crag]

so far all we get is some alternative spin on the transformers architecture, which inevitably has the same pitfalls

[rigid snow]

if we are such copyright maximalists why are we talking about this now and not when word2vec dropped, when google translate did, etc
is the problem ai being to recall/reproduce training data?

[sage crag]

its a problem of awareness

[hoary lion]

probably cause now copyright holders are now in stake

just like ai art shi

iykyk

[olive sable]

has even a milion 5090's been made yet???

[sage crag]

the problem of ai and copyright only hit the mainstream when ai hit the mainstream

[hoary lion]

https://alphafold.ebi.ac.uk/#news
am sure they make their own dataset by that uk institute

[opaque wharf]

Million is a small number for a high demand product I think

But I don't know the exact number

[rigid snow]

5090 is NOT a high demand product

[opaque wharf]

Yeah

[sage crag]

is this the dataset?

no this is generated data

hm

[hoary lion]

i dont think there are copyrights of 3d models on natural protein

[opaque wharf]

Am hungry

[opaque sigil]

wasn't there something about them turning a ton of enterprise cards into 5090s

[opaque wharf]

Sam what do you eat?

[opaque sigil]

since they're not selling well

[sage crag]

you arent certain though, right?

go research it if you feel like it

[olive sable]

i ate spaghetti with cheese sauce and bacon

[sage crag]

i dont want to spend more time on protein copyright when we cant verify any claims about it

[opaque wharf]

Huh? All I know is that they are turning card with small memory to one with much higher memory capacity

[opaque wharf]

Hmmm. spaghetti sounds nice

[lilac flame]

memory is the main limiting factor for high end gpus

[olive sable]

depends on the task but ye kinda

[opaque sigil]

the big boy blackwell cards are notoriously hard to cool so they haven't been selling that well

unless they fixed that

[opaque wharf]

Oof. I didn't know that lol

[rigid snow]

have yall seen the gpu startup

[sick owl]

£2000 lighter LULE

[opaque sigil]

we're at the point where a single gpu draws over a kilowatt

gl cooling that

[hoary lion]

https://www.embl.org/news/science/alphafold-using-open-data-and-ai-to-discover-the-3d-protein-universe/#:~:text=DeepMind trained AlphaFold on publicly,sequences and annotations from UniProt
They are public and free to download, some are from ebi themselves so

too big embeds!!

[sage crag]

you can surround your link in <> to suppress embeds

[hoary lion]

i just delete it manually

[opaque wharf]

Let's all play DOOM

[hoary lion]

no

[rigid snow]

they are making a card that they claim is 2.5-10x at path tracing compared to a 5090 at "consumer prices"

https://fxtwitter.com/BoltGraphicsInc/status/1952049562912530494

Bolt Graphics (@BoltGraphicsInc)

Bring your ideas to life with Zeus, a new type of GPU:
︀︀- Expandable memory: 32/64/128 GB soldered + 2x/4x SO-DIMM slots for up to 384 GB memory!
︀︀- Massive improvements to path tracing performance!
︀︀- 400 GbE QSFP-DD port!
︀︀- Built-in high-performance RISC-V CPU cores capable of running Linux!
︀︀- DevKits 2026, Mass Production 2027
︀︀
︀︀And we used the good old 8-pin PCIe power connector that is known to not melt!

**💬 136 🔁 235 ❤️ 1.6K 👁️ 200.1K **

▶ Play video

FxTwitter · Aug 3, 2025, 4:50 PM

[sage crag]

while i dont doubt the validity of this (due to the content), there are no sources on the page

[hoary lion]

im going to go back using claude

and add profiler

[opaque wharf]

And we used the good old 8-pin PCIe power connector that is known to not melt!

[hoary lion]

I need to know all these issues raised from phoenmizer

[opaque sigil]

that thing ONLY does path tracing at fp64 btw so it'd useless for most use cases btw

[opaque wharf]

So just ASIC?

[lilac flame]

what’s path tracing

[opaque sigil]

yeah

[opaque wharf]

Ray tracing but better

[rigid snow]

still very interesting

[opaque sigil]

it could maybe be useful for simulations i guess

[sick owl]

Tracing the path of something*

[opaque wharf]

Path traced audio mhmm