inb4 trannies start seething about binary coding and binary genders

>>109455171
Based Elon is right again. As always. I expect nothing less from the smartest man in human history.

>>109455171
Oh great another twitter screenshot

if you say exactly right it means kimi made you into her zombie

>>109455171
Im a non binary vibecoder.

File: F12019_Schloss_Gabelhofen(...).jpg (77 KB, 641x816)

77 KB JPG

>>109455208
>BWHOA

>Using an LLM to write 1s and 0s

>>109455171
Elon is a fraud, he doesn't know anything about technology, like every CEO, he only knows how to lie, he is all day posting on his botnet platform while taking credit for what other starving people do for him

>>109455474
>I'M SMARTER THAN ELON!!!
Cool, so where's your top 10 Forbes company, tr00n?

>>109455171
Ignoring assembly is exactly why software has gotten so much more worse and bloated though

>>109455474
Okay, but like can you disprove anything he said or point out his absurdity there? You already can get AI to reverse engineer assembly to source code. Its barely a leap to interpret raw binary.

>>109455171
Except compilers are deterministic in nature, while your LLMs are slot machines with random outcomes. Yeees, blindly trust it and throw all your money at it, little goyim

>>109455187
>Inb4 gender is only one bit

Bros...
https://x.com/lauriewired/status/2084326266132652178

>>109455481
DogeDesigner, is that you, posting from a loo, in Bengaluru?

>>109455171
That's not how code works at all, what is this retard talking about?

>>109455171
How tf would this even work, you train it with a bunch of binaries and then expect it to predict the next byte without breaking anything? And what if i want to modify a program of 10GB or more?

>>109455171
They want to create proprietary binaries so no one may have the sourcecode.
What would happen if it generates the exact binary of a compiled version of a GPL software?
How would it provide it's sourcecode for modification?

>>109455469
It's literally more powerful since you're back to talking to the CPU the language it knows best instead of in-between coding languages that may get interpreted wrongly by the compiler.

The only main downside is that you don't have a direct view into the binary sequence, don't copy and paste binary from an online LLM.

>>109455608
>muh auditing
will not be required, idk why but code trannies are having a hard time coming to terms with the fact AI is the new interface
Do you heckin audit the assembly emitted by compilers?
There will be no compiler at all lmao, how hard is this to understand.

>>109455490
>Ignoring assembly is exactly why software has gotten so much more worse and bloated though
isn't it the job of the compiler to clean up the bloat when it turns it into 1's and 0's though?

>>109455662
Don't just copy and paste vibe coded code for anything outside of a local infra holy shit

>>109455506
code troons need re education camps atp, years of webshit has turned their brain into a mush.

>>109455612
>DogeDesigner
this was so pathetic and sad.
worse than his path of exile shit, like why even put that out there in public

wait until they combine AI with analog cpus

>>109455481
>all nepo baby richfags are smart

you really are this retarded huh?

>>109455171
how many more times are you going to spam this particular shitter post?
why don't you stay on shitter or kys?

>>109455506
Programming languages will always be needed because they provide a level of abstraction that is easier to reason about. Also, the same code can work on many different architectures. It's possible for a human to write machine code but it's far more economical to work on a higher level of abstraction, and the same applies to AI.

I'm not denying that AI will continue to replace programmers, but it will always use a programming language. There might even be a new language designed to work well with AI.

>>109455651
Give me an example of something a compiler interprets incorrectly.

>>109455171
If this kills "unused ram is wasted ram" era then I'm all for it. Modern programmers don't give a shit anymore anyway.

>>109456008
Anything that can be SIMD'd. Anything that compiles to a vtable usage in e.g. C++.

>>109455171
This sounds like a security nightmare.

>>109456163
>virtual functions are always unnecessary
>simd is always better
dumb retard

>>109455651
This is not how LLMs work.

>>109456206
>A -> B
>B
> |- A
Falsum

>>109455171
Reminder: this guy has more GPUs than anyone else and created xAI to compete with OAI but his models can’t even compete with older versions of chatgpt and claude and they still get btfo'’d by open-source Chinese models by a margin. What's even more pathetic is that he promised to open-source Grok models and then fully abandoned that idea.
Put simply, he somehow end-up making shitty proprietary LLMs that is purely for show.
Fuck this retard. I'm tired of his bullshit.

>>109456305
I think more funny is that xAI is now owned by SpaceX whose revenue comes from tax payer dollars.

>>109456305
>this guy has more GPUs than anyone else
He uses them to pretend to be a gamer, not for compute

>>109455626
Why not? There's no technical reason you can't use an LLM to go straight from a text prompt and maybe some sketches to an executable. Whether any of the LLMs are able to do this correctly or well is another matter.

>>109456227
Sure it is.

>You are a compiler that takes natural language text as input and you produce executable binaries as output.
>You understand ARM64 and x86-64 and know all about formatting the binaries for any modern operating system, as prompted.
>Make no mistakes. Write no errors into the binaries.

And that's it. Okay for technical reasons you may have to require it to produced encoded output, but that's a matter of harnessing.

>>109456325
>The claim that the vast majority of SpaceX's revenue comes from taxpayer dollars is false; roughly one-fifth (about 20%) of SpaceX's revenue comes from U.S. government contracts, while the majority is generated by commercial operations like Starlink.
code trans lost

>>109455651
This lil faggot thinks he’s smarter than the compiler lmfaoooooo

>>109456246
..no?
He was asked for an example not an implication. Anon says "anything that can be simd'd". I guess the only argument you could make is that 'anything' can mean both 'everything' (which was my interpretation) and 'one thing', as in "there exists one thing that can be simd'd and it would be good there for the compiler to do it, but I won't give a concrete example teehee".

>>109456504
Congrats, you are retarded.

>>109456527
>nothing to say
Explain why you think I am retarded. Otherwise I think you are a bot and will stop posting.

>>109456305
his physical ai on tesla and optimus is still the best, llms and chatbots are loser technology

>>109456352
>random text generation that has no actual conception of how things work generates text that may look correctly from a first impression

File: IMG_20260804_130032.jpg (109 KB, 627x449)

109 KB JPG

>>109455171
>SAAR

>>109455171
This retard (Musk) has the opposite of impostor syndrome. He's painfully incompetent and it's so obvious but he acts like he knows shit.
Bonus example: https://youtu.be/vYbEVmStFDQ?t=89

>>109455507
>Except compilers are deterministic
Damn I wish this was true.

t. reverse engineering guy

>deepfaked apps full of malware, flooding the Internet
No thank you.

>>109455171
Nah. We just just move to most people managing more human readable languages so ai slop can be audited. More javascript and python. C++ becomes Assembly 2.0 so to speak, fewer will be in situations where reading it is necessary.

>>109459559
I wouldnt go that far, he just isnt infallible

That's retarded. The abstractions that programming languages provide allow you to write source code that is maintainable in a way that raw binary is not. This does not change just because the LLM is writing code.

>>109459830
This. The more readable the code is the easier it is to maintain

>>109459830
>>109459969
So this>>109459718

>>109455506
>let’s have AI write binary that no human can understand
This sounds genius can I be a trillionaire like Elon?

>>109455171
Elon musk shows himself to be a dumb "ideas guy" who doesn't really understand anything and doesn't think anything through whenever he opens his mouth

>>109460262
He wants to sell his Borg implants so people can understand binary.

How are you supposed to achieve that, when your code needs to explicitly declare library and system calls? Or are you just going to run it directly on your hardware with no real OS?

>>109456337
It's obviously fucking stupid if you know anything at all and if you actually think at all.
- all of the code validation and analysis tools are designed for source code
- ai often makes stupid mistakes still but at least you can reasonably understand the source code it makes and can look through it for obvious mistakes, or debug it yourself
- if you don't blindly trust your ai (you shouldn't) you can much more easily check what it's doing by reading sources code
- if source code has problems you get compiler errors etc that usually tell you something. If a binary doesn't work you get "segmentation fault" or "aborted" or "floating point exception" or "stack smashing detected" and then you have to painfully fuck around with things like gdb to find out what's wrong (or to get useful info to copy paste to your ai)

Etc. These are instantly obvious to me and im an average person midwit

>>109455481
where is einstein's top 10 forbes company?

>>109460339
>the code validation and analysis tools
won't be relevant when AGI hits in two weeks
>ai often makes stupid mistakes
are you still using an outdated LLM from more than 1 week ago? try the frontier models, they fix everything
>if source code has problems
won't ever happen when AGI hits in two weeks
>If a binary doesn't work
skill issue: learn to prompt

My point is that for people who believe in LLMs, then it's a perfectly reasonable step to move from LLM -> source code to LLM -> binary.
But you must believe.
Do you believe? Sam Altman believes. Elon Musk believes. Their combined net worth is a very big number of dollars therefore they are right.

File: paul-grahams-amazing-insights.png (42 KB, 644x164)

42 KB PNG

>>109455206
Retard. The problem with twitter screenshots is when some literal who says something literally retarded, and a 4chan OP tries to stir up drama or make the retarded point seem more important than it really is.

In this OP's case, the tweet introduces a legitimately interesting technical topic about the possibility of AI-generated binaries. You're free stick to a desktop ricing thread or the billionth OS flamewar bait thread if you aren't interested

>>109455171
>>109455506
dumping out raw binary is retarded but dumping out ASM could be viable. Actually, they probably could do that now couldn't they?

>>109460339
All your points are assuming AI wont be able to do those tasks, which in will-smith-spagehetti days, Id have agreed with you, but claude is able to pump out decent 5,000 lines of code now, and reverse enginnering binaries to source code is only being suppressed atm by censorship regulations by the AI com, but i can assure you is fully within current AIs capabilities. Wheres the leap? Give the AI the compiler too, then what happens?

>>109456352
where is the english to binary training data senpai

>>109455187
>we need to accelerate the development of quantum computing so that these CHUDS can see that even computing is a spectrum!

A language model generating source code that is then translated into binaries by a compiler instead of just generating the binary directly seems similar to using floating-point instead of just raw bit arithmetic.
The second option might be more natural and more general, but the tradeoffs are enormous and in practice that flexibility might not be all that useful.

>>109456352
>>Make no mistakes
Its actually insane that retards like you think that telling an LLM to not make mistakes will not only prevent it from doing so, but has any effect at all whatsoever. I actually cant stop laughing holy shit dude.

>>109461088
>Their combined net worth is a very big number of dollars therefore they are right.
All elon musk and all the other shitty oligarchs prove is that being good at bullshitting, being born into money, being rich and throwing money around, and having no ethics, gets you very far in life

>>109461845
>gets you very far in life
In some ways yes, but none of those people are actually happy. I mean look at fucking Elon, hes eternally seething these days about trannies because Grimes left him for one.

>>109461243
They're unbelievably bad at even common asm for now unfortunately.

Actually using LLMs to go from shitty languages to assembly would be a good thing, if it actually works and never fails in a way that is impossible to even find out how and why it failed. So yeah terrible idea

>>109461753
Keep seeing corpobitches say "set temperature to 0" in the prompt. They think it makes prompts more reproducible.

>Giving a next token predictor the job of working with something that is very repetitive.
Don't these retards know how LLMs work?

>>109455171
What a nice sounding midwit idea. Where is the giant training corpus of binaries that WEREN'T produced by compilers from some human readable programming language? The reason why LLMs can code is because they could train on staggering amounts of source code, complete with source control history.

>>109461243
they are similar problems.
the real question isn't whether AI will be able to do it, the real question is whether it will ever actually be better to do it that way than to use source code as an intermediate step.

>>109461222
Retard.

>>109461601
>A language model generating source code that is then translated into binaries by a compiler instead of just generating the binary directly seems similar to using floating-point instead of just raw bit arithmetic.
no it isn't

>>109463228
yeah given how LLMs work they tend to prefer the points of most leverage.
Really great at balling create-react-app garbage and demo porn but anything beyond that you're fighting context rot every step of the way
ASM and binary would 10000x that effort

>>109455192
smartest and richest, he is literally the god king of earth who has never had a project or a prediction go bad

File: 1785891343848.jpg (91 KB, 1064x1021)

91 KB JPG

>>109455469
You just train it to predict when a 1 comes after a 0 and vice-versa. It'll be super small and fast.

>>109461753
It does have an effect, it's equivalent to adding "carefully..." to your prompt.

>>109455507
What a pseudo intellectual faggot.
Decomp trannies figured this out a while ago. That's how decomps work. A bunch of autistics try various functions until the functions match the bytes once compiled. That is itself "non determistic" guess and check until you match. And you're admitting AI basically is excellent for this use case. Try some code, use a compile script to check if it matches. If not. Try again.

I also even said this earlier and it must have triggered tranny jannies because they deleted it.

>>109461463
You don't need data for this since you can verify a program satisfies conditions by testing it, so it can be done with pure RL.

File: almao.png (82 KB, 640x480)

82 KB PNG

>>109455171
LMAO, I can't wait for AI to automate the generation of day zero vulnerabilities.

>>109455171
Then why is he still importing jeets?

>>109463577
Funnily they try to manually output the create-react-app-generated scaffolds (poorly) instead of using the actual create-react-app tool all the time, resulting in breakage from the very first prompt.

>>109464645
You have 0 clue what determinism is or does. You are clinically retarded.

>>109465529
so you tried to ask it to make a react app for you and it did something approximating that, or you explicitly asked it to use create-react-app?

I've noticed LLMs do things like this. I've been making mine reconstruct a book from scanned images for the past couple of days, and I realized that this generation is really primed for instruction following as literally as possible, instead of thinking slightly "out of the box" and using the component library we made earlier specifically for this task

so it would try to regenerate the page's layout from first principles with lines and boxes and shit

>>109465790
If you explicitly say "use create-react-app" you are just wasting tokens for no reason when you can personally type create-react-app. The only use of this kind of query is to ask "setup a default react app". It should obviously know to use create-react-app instead of manually writing a package.json file and manually inputing its content with a 50% package name hallucination rate, especially in 2026.

As for reconstructing content from scanned images, that happens to be within the application domains I've been working in. I found that mistral-ocr and gemini-2.0-flash are the only two capable of properly doing this, with 3.0-flash having been slightly worse than 2.0-flash. Openai and anthropic models just completely shit the bed, hallucinating just about everything and abridging the shit out of everything. Post-3.0 gemini is also fucked, providing "demo pages" instead of actually extracting page contents. Though maybe the context is different since I work with very big documents (e.g. 450 pages) and use structured output because the goal is to retain a json content description rather than, say, md.

Like you said, the models can take the text raw and reinvent the format around the text (even without the image). It even works pretty well actually, but obviously it's fake and inconsistent across page boundaries. But that's not quite what I'm after here.

What's been your experience there or what is the workflow you use and what kind of success rate do you get?

File: file.png (797 KB, 1500x1435)

797 KB PNG

>>109465850
>mistral, gemini
yeah I didn't want to pay up to even bother trying so I've been coping with fable and sol

this is the type of stuff I'm working with
1. ToC

Sonnet-level models have no problem with OCR but can struggle with ruby text
contd

>>109465909
Do you use the model to perform the ocr (as I do) or do you use it to write an ocr pipeline?
Do you extract structure or just the text itself? In particular, tables, headers, sections with separated section title and section number, etc.

File: i067_compare1.png (1.1 MB, 2798x1400)

1.1 MB PNG

>>109465909
After forcing it to output things in terms of the component library, and reminding it to use built-ins for numbered lists and bullet points, the output got much more consistent

>>109465918
I tried two methods
The first method was more of an OCR pipeline where I extracted the geometry and separated the text and tried to use tesseract and model-based OCR, and attempt to iteratively render each page "better"

This however was a massive waste of time and tokens. The document never converged and wasted about a week's worth of sol's tokens.

So with Claude, I went with a VLM-centric approach and simply had the VLM directly reconstruct text, pages, diagrams, charts directly from the scanned images.
This turned out to be on the right track, but ultimately a mistake as each agent independently re-invented the same graphics and designs.

With the realization from >>109465790
(LLMs take instruction following too directly) I decided that the smartest approach was actually to invert the process; start from an empty book and iteratively classify large structures from the existing component library, and have multiple passes to iteratively refine the data, doing OCR almost last.

>>109465934
I'm still in the middle of doing that (the empty book) and I have the ToC and index working since those are generated context. I'm gonna see if the skill I'm working on is smart enough to fill in the middle chapters and the data driven stuff.

File: sbs.png (782 KB, 1181x2640)

782 KB PNG

>>109465942
Oh, one big tip: Use a linter to emit warnings when LLMs are trying to use any tools that create shapes (called visualizations in typst), and create a whitelist for uncommon characters, which will help you detect if your LLMs are trying to invent bullet points from first principles

there's some other shit too but I think you'll figure it out. linters are great.

>>109465934
For my use case I have to hit a processing rate (fully parallelized and with no real budget constraints in theory) at about 5k pages/minutes. With the one-shot pipeline I use (using the file api to pass the files to the AI), I get about 1500 with either gemini or mistral, but reaching that rate with gemini requires creating a spend of $2000 a month to be in tier 3, which is retarded. In both cases, anything beyond that hits rate limits and dies.

The approach you describe is interesting and I might give it a spin, though I'm concerned about throughput.

How do you handle multipage components that way? E.g. sideways tables spanning multiple pages, or enumerations that start on one page and end on another page?

And if I understand correctly, the agentic loop selects iteratively a structure type based on the page input and current construction state, and adds it to the construction. Once the page is constructed (or whole document?) you then use ocr and give it that + the construction + the image to allow it to fill in the actual contents?

>>109465962
Fortunately the book I'm working with doesn't have multipage components, but if you tell the models that multipage components exist, they are smart enough to figure out how to concatenate them together. With scanned images though, this honestly may be a crapshoot.
I used to be a scanlator so I know how bad it gets

I think the better approach is to have it construct the whole document with some supporting files for templates, components and diagrams.
The reason to have those supporting files is so you can run a linter. I suppose you could do pages but I don't see a good reason to do that; if the LLMs get the layout wrong, one page can split into two.

But yeah I think that's the basic idea. Again, still going through it.

>>109465962
throughput will definitely be a big problem with this approach, especially at first. But depending on the book, that's really a question of tech debt.
You can fully parallelize it and have each agent independently generate each chart and table from first principles, and that will probably work
But if you're OCD about how things end up rendering badly because these models are blind as shit, you'll end up repeating the same work 1500x over.

Ironically, these models are extremely skilled at hard math problems and extremely bad at UI, design, and layout.

>>109455171
Yes, it will be glorious. And I'm comfy watching the world burn from a custom built PC running FreeBSD in a cabin in Alaska lol
>>109455192
At least the title has been revoked from Einstein, but it really belongs to Tesla. Or Hermes if you're a schizo.
>>109455474
>billionaire bad
Oh yeah real insightful take. Did you come up with that all by yourself?
>>109455507
Unfortunately "compilers" aren't deterministic. There are many factors that influence the final binary.

>>109465976
The reason to do pages is that it allows parallelizing the process by having a different request per batch of pages, then assembling the result in a post-pass which is faster because you don't need to send the document in and the extraction ought to be more lightweight than the original document. It also allows doing backreferences into the original document, but that can be arranged by asking the llm to annotate components with the provenance page/region and that works pretty well in practice.

Last question: what do you mean with supporting files for templates? What I use right now is plain json schema for structured output, describing a flat structure (LLMs are terrible at nested structures, so I use a reference field with id on all components, and a parent_id optional field for hierarchies, then deterministically reconstruct the graph in post processing) with allowable components, and I use a self-correction feedback loop in case of schema violation.

>>109465997
In this case I don't need perfect reproduction or anything, I just need to get accurate text/numbers extracted from the pages (which is already hard enough, I have things like legal references or cross-section references and they keep fucking things up, like the original would say 'as per law 1234 of status XYZ' and the extraction will say 'as per law 68 of status XYZ' instead for some fucking reason), plus structural annotations that will allow a post processing phase to create a referential graph between e.g. sections (so if a paragraph says 'in accordance with section 3.5' the postprocess phase should be able to easily link with the section item for section 3.5), and that will allow correction in understanding (e.g. reinterpretation of text as 'cells in table' instead of 'run-in' to avoid the "vegetative crystallography" issue, for example).

The output will largely (after post-processing) be vectorized for searches and also used as input to other LLMs for queries.

>>109466021
Compilers are fully deterministic.

>>109466022
Templates meaning for example design system tokens. consts. In CSS designers make a template file consisting of a list of color, size, and spacing consts to be referenced everywhere else.
You can also use it to globally set how numbered lists and bullet points get rendered.

I haven't had too many problems with nested structures, though nothing in my textbooks are too complicated.

>accurate text/numbers extracted from the page
Do your documents have an index, bibliography, references section?
You can use this as a kind of two-factor verification. If both extractions agree then it must be good. If one is wrong you can send an agent after it.

Yeah, I think your situation is simpler than mine.
Mine is for educational purposes so I've imposed a very strict standard; if the model can't reproduce the contents of the textbook in a way that makes sense, it must not understand the contents very well. Only after it does that can I start actually using the contents to generate more modern-era friendly education tools

File: msk.png (64 KB, 1478x431)

64 KB PNG

Elon famously has deep understanding of what is possible with machine learning.

>>109455474
You're on the right track. If someone is placed publicly, they are a puppet and facade as much as any actor to interact with the public. See how they are treated like stars and the public is polarized about them.

>>109466053
>Do your documents have an index, bibliography, references section?
Of course not, that would be too easy. Especially since it references outside documents that can be anything from business internal documents to appendices that are not provided or come from older exchanges, to public documents in jurisdictions that can't be guessed in advanced.

>I haven't had too many problems with nested structures, though nothing in my textbooks are too complicated.
For me, as soon as you have 2 levels (e.g. title -> subtitle -> paragraph is 3 levels) the frontier models break (bad reference, bad structures, problem respecting the schema, etc.), but flat structure works well enough (in exchange it will often not choose the right element like preferring to make a paragraph with "*" for point lists instead of using a point list structural item).
I think using a json schema would be a stronger version of using reference templates, so it should otherwise work OK if I put in a multi-turn iterative flow like you are doing. Again, throughput might not pan out, but worth a spin. Thanks for the inspiration!