File: 1748402132784160.gif (762 KB, 487x560)

762 KB GIF

>geothot
i dont care

>>108898942
Agents in my company literally cost more than devs.
I see people wasting over $100 in API costs on 1-2 story point tasks because everyone runs 5-10 agents doing different things, wasting so many tokens especially with a big repo. And AI prices are only increasing. The last month, 30 developers managed to burn $100k worth of tokens.

>>108898942
The recent math proof definitively refuted LeCun and Marcus. AI skepticism is a cope for people who are too cowardly to accept they are actually anti-AI or is a way to grift off of people who are still coping.

>>108898942
That's basically the opinion I hold. Although maybe unfortunately for us, the current approximations are already good enough to satisfy most of the real use cases for code. Not to mention the best models already beat all but maybe the top 1% of SWEs

>>108899062
you don't code.
fuck off.

>>108899062
Maybe math is just different.
You can generate proofs and then test them with lean. There is no problem of slop being accumulated.
Vulnerability discovery too, one can test it, you either find a vulnerability or you don't. No slop accumulation.

>>108899062
yeah but can (You) harness AI to make any progress in Math? No? Do you even understand these math propositions?

I'm not convinced average (that is, retarded) software engineers do benefit from AI coding. They don't know what the fuck they are doing in the first place.

>>108898942
>>108898977
Geohot hasn't been relevant for 18 years.

>Geohot thinks
Stopped reading there. Post the actual arguments next time.

>>108898942
Statistically, I know that commented back to you is worthless, but here I am.

>>108899023
I mean I would do the same if I was a wagecuck in a company
imagine putting effort into your corporate shit j*b in the current year (2026) when you could just proompt the day away
dumbass goyim I swear....

>>108898942
I would just like to say that Go has been slowly declining in popularity over the last few decades, due to a lack of interest from eastern youth. Chess managed to find a new space for itself during covid, and while online Go exists it didn't have nearly the same reach.

>>108899985
>when you could just proompt the day away
But you may end up getting dumber and thus totally unemployable.

AI will make garbage look nice and corporate structures will reward features over polish.

>>108898942
the obvious answer to slop is training the model on the way the user is using the model and fixing up mistakes. Western AI companies claim not to train on your data are simply lying the more you use it the better it will get so while the slop still will be there it won't become that bad.
Anyway so far I have seen nothing impressive from the vibe coders not even the math problems solved were useful in practice why don't they do some real world problem instead of these homework questions?

>>108898942
geohot gets double anal penetrated by SonyChads

What nobody seems to be talking about is even if senior and experienced developers can make good use of AI and can tell slop from valuable code, what happens when they retire? What about current students and future students? How will they get the experience and knowledge that the current seniors have? I mean the short answer is they won't, universities are shoving AI down students' throats and vibecoding and cheating is rampant. So even if AI doesn't destroy the world right now it will in 50 years when we have multiple generations raised by AI.

>>108898942
That is not a very smart person if he only realized that now. It was always obvious.
Although LLMs are very useful, they cannot create anything. They are good at predicting what YOU want. They don't want anything on their own, they're markov chains guessing next token based on prompt and a compacted version of the Internet, aka AI model. There is not magic inside the model. It is data downloaded from the Internet (decades of data from archives) with lossy compression applied to it. That's what training is. Lossy compression of data. utf-8 -> tokens -> vectors and matrixes -> weights and biases.

>>108899062
The proof was not a counterexample, they literally say repeatedly in the paper they had to have humans clean up the idea. The LLM found the right construction but mathematicians needed to actually make it a proper proof and they even comment that this ignores the many cases when it gives you something that just sounds right but isn’t… nobody seems to read anything and nobody seems to know anything on this board

>>108898942
True. It's weird, but the problem is that it works to a "just good enough" level at a rate that's beyond human QC capacity. If we're locked in a room playing russian roulette for 8 hours a day to get a paycheck, which makes more sense to play? One round an hour with basic 1/6 odds, or infinitely many rounds an hour with those same odds? Even if we make the infinitely many rounds game much safer per play by dropping the odds to 1/60,
what do you think is going to happen when you're moving at that speed? You'd need to have a clean up contractor on speed dial.

>>108899083
No math is the same. Arguably worse even because fuck ups there permeate into everything else. Similar to how people turn the other way when a dashboard says tests are green and code coverage is 100%, math slops can accumulate by churning out AI driven proofs with flawed validation procedure - peer review becomes a rubber stamp and people offload validation to an LLM judge. 300 years later humanity finds themselves wrestling with hallucinations in things they would have long considered to be foundational mathematics and the consequences of sheer volume of those flawed assumptions leaking into the rest of STEM.

>>108902716
go seems to require patience, which is antithetical to how people experience everything nowadays

>>108902716
> I go to go
eheheheh

this is faggot who volunteered to work for twitter for free to fix search and then started begging his followers for ways to fix it right? lol
washed ass ps3 nigga

>>108898942
>https://geohot.github.io//blog/jekyll/update/2026/05/24/the-eternal-sloptember.html
>They are a highly sophisticated statistical model designed to mimic the distribution of programming.
>Things can be broken in ways that weren’t previously possible
wow, geohot finally realized the obvious! and he's supposed to KNOW about this shit.
I'm not even an expert (I don't even fully understand how LLMs work), but this was clear to me ever since I learned about machine learning and then about LLMs.

anyway, obviously I agree with him, and I think the future looks grim. we don't know where all the slop code will take us (as in, code quality, labor quality and demand, ...).

>>108902567
so it's a glorified assistant. just like some of us have been saying for years.

>>108903387
you ever feel like an idiot typing stuff like that out? ever read it back to yourself?

>>108903387
still, the argument about things being "broken in ways that weren’t previously possible" feels kinda ironic coming from him, maybe even hypocritical... didn't he make a whole self-driving platform? how is that different from LLMs?

>>108902227
>So even if AI doesn't destroy the world right now it will in 50 years
I think the technological decline already started. things are too complex, even tech companies switched to cheap labor, and apparently they aren't hiring juniors. plus, as you said, this covid generation is completely messed up.

>>108903441
hey george, no need to feel offended. we know you love attention whoring, and that you are trying to show yourself as contrarian here just so you can get 5 more minutes of attention and a new dopamine hit.

>>108898977
fpbp /thread

>>108903459
nah man, that's guy's an idiot too.
i was talking about the part where you're like
>i actually don't know how any of this works
>but i am very confident that it works like this

you do that sort of thing a lot?

>>108898942
>>108899657
>>108899857
>>108903468
gn sars

>>108903486
so, admitting that you don't know the little details of something (even if you do understand the fundamentals of it) makes you an idiot?
I don't give a shit about my self-image or about having to explain things to retards. this is a anonymous forum after all.
/g/ is the "tech" board, where people are supposed to understand tech (at least conceptually). I will assume that you know what I'm saying, but if you don't, then lurk moar, I guess.

btw, tell me how I know you are american without saying it.

>>108903534
aight man, tell me how the attention mechanism in a transformer works. that's like basic shit, not little details.
you got 1 minute to reply with details.

>>108903592
again, I don't know the details. I know that they work with statistics, just like many other ML algos. that's more than enough to understand that LLMs are glorified, unreliable assistants. they are useful and might improve over time, but they will always be non-deterministic, and therefore unreliable.

>>108903618
lmao clown as nigga

>>108899062
>The recent math proof definitively refuted LeCun and Marcus.
it did not. they never claimed generating wholly abstract mathematical statement sequences required a world model. math is explicitly designed to abstract context away and leave only formal symbolic manipulation games behind.

the real breakthrough is that the model can bring out the big guns from one field of math to tackle problems in another while humans typically spend their lives learning just about one of these

>>108903631
I'm sure you can prove me wrong :^)

>>108898942
>I’m calling it now, the adoption of AI agents into software development will be one of the most costly mistakes in the field’s history.
I highly doubt it can outdo NPM.

File: 1778696021016013.webm (1.25 MB, 502x540)

1.25 MB WEBM

>>108898942
I 'agree'
but this
>it's a slot machine
line is dumb. it isn't, you just suck at conveying your thoughts and intentions through natural language. typically the errors with polish arise because you make some assumption about what the LLM knows, and so your prompt is misleading. this involves conditional hypothetical thinking, you have to plot ahead of the LLMs quirks and lack of context.
there's also context rot, too, if it can't implement something for the life of it you just condense the current chat and then bring it over into a new one.

>>108899023
I am now wasting tokens on purpose, so when they finally realize that the prices go up next month, the MBAs shit themselves.

Bitch faggot

>>108899062
Robbins conjecture was proven with an ATP in the 90s. LLMs are definitely not a milestone here.