← Home ← Back to /g/

Thread 107163140

30 posts 24 images /g/
Anonymous No.107163140 [Report] >>107163160 >>107163195 >>107163468 >>107163486 >>107164392 >>107164508
Sentience soon (tm)
Anonymous No.107163160 [Report]
>>107163140 (OP)
theres just a couple tiny we details we still have to fix
but were there by 2027, for realsies this time
Anonymous No.107163195 [Report] >>107163233 >>107164365
>>107163140 (OP)
jokes aside you can now ask it to compile and run bits of code server side. i think.
hopefully it isnt hallucinating the whole thing
chud gpts becoming a neat little tool otherwise
apart from that its a corelation machine
and an interface for web searching
its not worth 20 bucks
but i could pay like 3 fiddy
and i think that would be a viable-er business model
instead of a handful of people paying 20k or whatever
shove it everywhere bc you charge 3.50

hes in too deep with the whole hype now i think
if he does a hard stop and then turns around people are gonna lose interest
Anonymous No.107163233 [Report] >>107163310
>>107163195
I asked it to remake some old MS-DOS game in Javascript and it gave me a link to Youtube for the full playthrough.
Anonymous No.107163248 [Report] >>107163950 >>107167712
I'm out.
Anonymous No.107163310 [Report] >>107163374
>>107163233
it sucks at programming proper
in practice all you can ask it is bits of code that illustrate concepts, or interfaces youre working with
which is pretty damn powerful in itself, actually
i experienced that with opengl where learning what would take me a week or so took me a dozen hours.

you can ask it to break down the steps of a process, and keep prompting it for alternative ways to explain the thing until you get it
what can work is if you tell it to generate a function within a specific context, but that becomes tedious very quickly, to the point where writing the code directly is actually faster and less work intensive
maybe theres some system prompt shenanigans to be attempted, but i didnt test it yet, what im working with doesnt have enough training data (c intrinsics and pretty advanced optimizations. picrel)

and its still dyslexic. its a machine built out of stats, so it wont provide the best solution, it will provide *a* solution which might not even be correct. otherwise its gonna be the most statistically significant one anyways, not the best or best adapted one

so, its a force multiplier, but it has many limitations
Anonymous No.107163374 [Report] >>107163447 >>107166275
>>107163310
cont
btw, if you want to prevent hallucination, formulate a theory you build based on your current understanding, and ask it to either confirm or deny it.
it shouldnt hallucinate and if youre semi correct or your vision is a simplification it will build on your model unprompted
theres ways to circumvent the limitations of the tech and still make it do useful work
but its certainly not anything near agi lmao
Anonymous No.107163437 [Report]
>sentience means doing what I want you to do
Reading comprehension soon
Anonymous No.107163447 [Report] >>107163521
>>107163374
Nta but have you ever tried Gemini?
Anonymous No.107163468 [Report]
>>107163140 (OP)
another trillion dollars in financing should resolve this
Anonymous No.107163486 [Report]
>>107163140 (OP)
I don't care about text, when can I CLANK a robot maid programmed not to like it and display severe depression and anxiety for days afterwards?
Anonymous No.107163521 [Report] >>107163601 >>107165084
>>107163447
yeah but it was upon its launch i think
nah tfu it was cursor

>inb4 try it, then muh premioom
if you have premioom, showcase it-

ask it to write a program that takes in a path to a bmp as parameter
opens a window, adds a zoom feature tied to mouse scroll, marks the contrast in red, and does it:
-in C
-using SDL2
-and the graphical operations have to be written in glsl

last time i tested chud gpt thats how fart i got
qwen and deepsneed perform slightly worse, they open a window allright but then shit themselves
Anonymous No.107163601 [Report] >>107163661
>>107163521
Interesting. Personally I tried Gemini 2.5 pro to create a "rocket game" in Java (make a triangle move across the screen with basic physics) and it worked but it had a lot of bugs (crashed when colliding with walls, the rocked disappeared after certain time etc), I managed to fix some of them just asking Gemini but it created other bugs and so on
Anonymous No.107163661 [Report] >>107163693
>>107163601
yeah probably what happened in your case is that you hit context limit.
thats the symptom- the chatbot starts to forget things

the counter to that is an attention mechanism to "remember" the relevant parts when context shifts.
but then you lose "resolution of the information" so to speak. its like a summary- you cant cram every detail into a summary
and in programming every detail counts
as the prompter one could re-establish a relevant context, but it becomes tedious very quickly, or even impossible with a big enough project

and the context size problem is one where linear growth needs a quasi exponential increase in computing power
so its not something thats gonna go away anytime soon. even if we reduce token usage
Anonymous No.107163693 [Report] >>107163734
>>107163661
keep in mind that i didn't manually edit any code; I was trying to use it to simulate not being a programmer or having very basic programming skills to see how far I could go "vibecoding" something
>pic related, it's the shitty """game"""" i'm talking about. This was two months ago using gemini-cli with gemini 2.5 pro btw.
Anonymous No.107163734 [Report] >>107163936 >>107163936
>>107163693
yeah thats were talking about. otherwise the test would be invalid
Anonymous No.107163936 [Report] >>107163970
>>107163734
>>107163734
>in your case is that you hit context limit
i thought Gemini had like 1 million tokens, but maybe it's not enough.
it blows my mind when I read that some non-programmer creates a quite large project, but it seems that it must be fake or at least quite exaggerated or perhaps it's a case of survivorship bias, considering that many people try to vibecode something, but only a very small percentage of them manage to create something truly useful.
Anonymous No.107163950 [Report]
>>107163248
He got mindbroken
Anonymous No.107163970 [Report]
>>107163936
They all suck dick for code bases larger than a snake game, but they do excel at one thing which is prototyping so you can generate boilerplate pretty fast and build on top of it yourself, that would be the only use case for ((vibe coding)).
Anonymous No.107164056 [Report]
This is unironically the right time for Ask Jeeves to make a huge comeback
Anonymous No.107164365 [Report]
>>107163195
>hopefully it isnt hallucinating the whole thing
Everything an LLM outputs is a hallucination, that's how they work.
Anonymous No.107164392 [Report]
>>107163140 (OP)
Anonymous No.107164465 [Report]
yeah that shots annoying koboldcpp does that im trying to make it act like alexa/siri with speech recognition and tts but it jeeps talking to itself and ig oring the triggerword context
Anonymous No.107164508 [Report]
>>107163140 (OP)
AI psychosis.
Anonymous No.107164581 [Report] >>107165683 >>107165822
I think new, sentient Grok 4.20 is rolling out as we speak, actually. You might not like it, but this is how ASI looks like. My boy Grok is associating things from across X and then going in for the kill.
Anonymous No.107165084 [Report]
>>107163521
you need triple your width - 1px
Anonymous No.107165683 [Report]
>>107164581
grok 4 got a JQ update
Anonymous No.107165822 [Report]
>>107164581
Anonymous No.107166275 [Report]
>>107163374
yeah you can't just ask it make me minecraf now! and expect it to drop something meaningful
Anonymous No.107167712 [Report]
>>107163248
Sweet, it turned arab in hopes you won't be able to sense its presence.