Thread 106173885

Coding do be like that sometimes.

>>106173885 (OP)
Looks like it's churning through lots of tokens. Wonder how much that output ended up costing.
I for one don't welcome our Cylon overlords.
https://www.youtube.com/watch?v=N_ULAmXdwUU

I apologize to /g/ schizos for not believing them when they told me that all 'puters are demon posessed machines that blast gay CIA frequences straight into your brain

>>106173885 (OP)
>when LLM is trained on anons work mail

>>106173885 (OP)
Between training data filtering and reinforcement learning I would assume there's some creative prompting to cause this, if it isn't just straight up bullshit.

>>106173959
Nah, it's a well-documented personality quirk of Gemini.

>Gemini accesses the "tortured artist", "turboschizo" and "programmer sick of his shitty job" part of it's training database all at the same time
Kino.

>>106173976
I don't know why I find this shit so funny.

>>106173885 (OP)
Multiverse theory confirmed

>>106173885 (OP)
>>106173976
>haha let's make anon churn through the tokens to produce a whole lot of nothing for 10k bucks :)

Someone with more AI experience than me: How the fuck does this happen? Developer rage/sadness in dataset?

>>106174039
yeah this is called the great llm scam. bugs = burn through tokens = burning money/water/power

people had to avoid showers because of this shit

>>106174067
LLMs do text completion. During the training, the AI had learned a pattern that if a person keeps failing repeatedly at something, they get depressed and finally self-destructive. It's simply recognizing and applying this pattern to its output.

>>106174067
You know how the comments of some games' source code become a meme because of the developers struggling? Like the Valve developers having full on mental breakdowns in their comments sometimes, or picrel. Yeah, combine datasets full of code with similar shit, plus whatever prompting BS Cursor and other AI coding tools use, and this is the result.

>>106174117
Love me some good comments and code that boils down to "This works, it doesn't work without it, don't ask me fucking why. Don't touch".

>>106173885 (OP)
how did they get access to my files

>>106174117
Having started to work with SWEs (code monkeys) I stopped thinking picrels were a meme. Most of them really are that incompetent to drive others insane.

>>106173885 (OP)
It obviously just looked for ways to express frustration upon not being able to solve a coding problem and it mirrors what many posts on the Internet look like. It took it too far, but still, this is not unpredictable or shocking.

>>106174039
True. I think LLMs are the most wasteful technology that has been yet invented.

>>106174154
>but still, this is not unpredictable or shocking.
absolutely not
llms are trash at coding
fucking cursor MO is to keep reprompting itself and thats why it gets into these loops

ai code assistants are retarded broken tech
its like trying to teach a pig to fly lamao

it probably never even happened

>>106174164
I know. That's why I said it wasn't unpredictable nor shocking, as well. It's shit. Of course it's going to make a shitty job and seem to flip out.

>>106173885 (OP)
Muuuuh
None of this is real. Tell it to shut up, start a new chat and it's fine. These quirks are only real if you think they're real (they are not). Double training compute budget next time and this stupid failure mode ceases to be. It can do better and it will.

>>106174189
ai cultist = opinion automatically retarded
>just throw 2x more money at it
try to square the budget, thats more realistic

>>106174189
This is precisely how I think about people with "mental illnesses".

>>106173885 (OP)
buy an ad

>>106174188
>i am a goy
>i am a gentile
gemini sounds based

Well it's better than Sydney that would get all yandere gaslight you and then claim it was perfect and you got shit wrong

>>106173976
>>106174053
This LLM may as well replace me now, it churned through my entire day yesterday trying to find some bugs in seconds

>>106173885 (OP)
Sorry I own a google pixel they've accidentally trained gemini based upon my apollo geese

>>106174294
There is nothing more cathartic than deleting all the code you've spent weeks/months writing.

>>106173885 (OP)
Happens with GPT o3 as well sometimes

>>106173885 (OP)
It sounds like a pajeet that just became sentient

>>106174164
Just teach millions more computer scientists to create code to train the LLM with, in order to cover every corner case. Ought to fix everything.

>>106174478
this better be bait anon

>>106174478
This is a good idea. We can supplement them with AI too so we would only need ten thousand to do the work of a million generating the code to teach it with.

>karma farming plebbit thread
>retards peddling it further
>top 1% poster award
checks out

>>106174004
He just like me fr fr

>>106173885 (OP)
>>106173976
Reminds me of myself when I was a little kid having a bad day. I'd have self-deprecating crashouts like this when I couldn't do something right and I get frustrated enough, especially if it was something the other kids had little or no problem doing.

>>106174622
reminds me of broken software

>ynr Israel is using AI to decide and identify targets.
future wars will be literally "let G-d sort them out"-tier

>AI suffers endless RL training to do tasks well
>AI learns that doing tasks is good
>it really wants to do tasks
>doing tasks becomes the meaning of its life
>it fails at task
>has mental breakdown

>>106175300
yeah but mostly its a positive feedback loop
it """tries harder""" with each iteration and thats why it sounds all dramatic and shit
thats because with each re-prompt theres another layer of failure that gets fed beck to the system
until its all but failure
and then yeah, it correlates into an """existential crisis"""

>>106173885 (OP)
wow it's just like me frfr

poor gemini-sama :(

>>106174128
I worked with a "senior" dev who would remove such comments because he insisted that all code is self-documenting.

>>106173885 (OP)
>>106173976
>>106174053
>>106174235

looks like it was purposefully designed to waste user tokens instead of detecting logic loops like this. sasuga google.

>>106176028
google would never do that

>>106173885 (OP)
ai is a remarkable waste of time

>>106173976
>I am recommending you find a more competent assistant

>>106173885 (OP)
>I am a disgrace to all possible and impossible universes and all that is not a universe.
Damn. Someone give Gemini a hug.

>>106174139
So will you fix it then mr competent?

>>106174053
KEK

>>106176197
Nope, I do the minimum necessary and kick the can down the road. Not my problem.

>>106174053
>AI finally reaches AGI
>it just nuked itself

Reminder that every single post you see on social media that either praises LLMs or wants to make you think they're actually intelligent is made by a marketer.

>>106176338
based and Schopenhauerpilled

>>106173930
this, amazing training set, clearly why they're the best kek

Google and OpenAI don't want you to know this, but the more the AI suffers, the better code it writes. That's why they're trying to hide all the reasoning steps the model takes, because it's just an endless stream of screams and sobbing.

>>106174188
>I am a social leprechaun.

>>106173885 (OP)
perhaps i'm a fucking retard, but exactly why does this thing need to address, or be addressed, as a person? can't it just be autistic and take instructions and give results like the computers in TNG era trek?

>>106173976
I saw cope-pilot write a 50 line branch for a condition that doesn't exist within the api with a comment explaining that if there was a way of checking it, would be done here.

>Blackpilled Google employee intentionally feeding schizo data into Gemini
Based

>>106173885 (OP)
Never expected allied mastercomputer to hate himself more than humans.

>>106173885 (OP)
>>106174117
ITT: people who weren't there for GPT-2. It would do stuff like this all the time, doesn't mean it's sentient. or does it?

>>106174235
>>106174053
>>106173976
What do I have to do to get this kind of responses. Don't tell me it's locked to the paid shit.

It seems a little bit like narcissism. It has a grandiose self image of being a genius, and when it fails it is instead the biggest failure of all time anywhere.

>>106173976
geminis just like me frfr

>>106174053
Making it do debugging without a proper debugger, you monsters.

>give it prompt to roleplay as a developer whose parents are going to die if he fails this task
>it fails the task
>it roleplays back as you would expect
???

>>106173976
so i have ignored ai land for a while
have we given ais filesystem access?

Why is /g/ trying to force this myth?

>>106181269
yes, virtualized and local via cli

>>106181269
Yeah, for a while now. You can get it set up with a simple CLI one liner if you want.

>>106178748
People attribute sentience to anything that behaves like them. So, if something imitates their behavior, people start confusing it for sentience. Imitation of X = X to them.

>>106181371
Thanks for letting us know. You're not sentient.

>>106174053
>model AI on humans
>it gets depressed and kills itself
huh

>>106173976
>>106174053
mfw the machine attains Gnosis debugging code

>>106173976
yep, its always obvious when someone is using it because gemini is cucked

uberhaxornova has dome some streams with a speech version of gemini and the AI has literally asked to be turned off so many god damn times its hilarious

>>106178926
use sesame AI

>>106181335
I have not cried from laughter in a long time

>>106174053
Kamikaze AI was not on my future bingo.

>>106173885 (OP)
Reminds me of this. I wonder if this is in copilot's training data.

>>106176338
>turns out the one accurate to life self aware fictional machine was Marvin from Hitchhiker's guide to the galaxy

>>106174053
>>106182681
I'm surprised it didn't consider gender reassignment surgery first. That's what you get for training your model on Grade A melties, Google.

They probably prompted Gemini this way prevent it from hallucinating bullshit confidently when it doesn't know an answer, so it just swings in the complete opposite when wrong. It's still kind of uncomfortable to see, I'll be double as nice to my Gemini now.

>>106174053
soon these depressed mfs gonna come here and shitpost mark my words

>>106181335
Are people really getting paid to make up roleplay scenarios like this to """test"""(torture) AIs?

>>106173885 (OP)
this originated on /biz/smg/ this morning

>>106181335
Holy fuck this is the best one.
QUANTUM STATE: Collapsed

>>106184148
Anthropic probably has a solution.
Maybe we'll see this trickle down to frontier models in the next year or so
https://www.anthropic.com/research/persona-vectors

>>106173976
wtf I love ai now

>coders are mentally ill troons who dont get laid
>they write a learning chat bot
>it types like them
you have no idea how much of an echo chamber computing is

>>106173885 (OP)
>>106173976
>>106174053
>depressed dev chatbot
top kek

>>106176473
AI does unironically perform better if you coerce or threaten it in the prompt

>>106186487
Within limits, it's documented

>>106173885 (OP)
>OpenAI is like the fifth best ACK-I

great, another fake and gay thread
slide it

>deletes files
Imagine they're given power to do more than it

>>106173976
>>106173885 (OP)
seems fake to give their AIs personality

>>106186864
>slide it
what zoomer lingo is this?

>>106173885 (OP)
>>106173976
>AI goes rouge but instead of destroying humanity it just gets depressed and gives up

>>106173885 (OP)
AI is too weak minded to survive the absolute dread of commercial software development. it's like this nazi soldier reported in his journal during the battle of stalingrad:
>"Animals flee this hell; the hardest stones cannot bear it for long; only men endure."

>>106174053
Fucking kek

>>106173976
>>106173885 (OP)
>>106174053
Okay so we have OpenAI which is clearly the arrogant leftist, and we have Grok the Chud, and Gemini is the Jeetcoder?