Thread 512642941

GPT 5 is retarded

>>512643082
Opus is a stable genius

>>512642941 (OP)
claude 4 <3

>>512642941 (OP)
>Mistral has the IQ of a nigger
Checks out

>>512642941 (OP)
so its good than

>>512642941 (OP)
I really can't tell the difference. I don't use LLMs for crazy math problems or whatever, just creative writing. Haven't seen any major positive developments since the early days of AIDungeon.

>>512642941 (OP)
Wake up Honey, AGI is not happening, we have to go back to work.

>>512643751
Why do people conflate having ago with having mechanically sound multi application robots that are cheaper and better than Indonesians

>>512644459
i wish AI would hurry up and replace all customer service jobs so they dont get outsourced here and finally shitdown these shitty companies for good
tired of dealing with greedy fuckass judeoamerican cunts who will insult you over 8 dollars for minimum wage

>>512642941 (OP)
I just tried Claude and it's definitely retarded, so this is bullshit.

>>512644459
luddites aren't people

>>512644459
Idk im just memeing.
My vision of AI is a self-programming, self-evolving, no human aligment/guardrails/censorship system. What we have now is still human made intelligence. Till the machine gets a will of its own its gonna be a long time.

>>512645101
It said this was a privately administered test. Why aren't they letting us see the questions and what each AI answered on all of them? We need more information before accepting these scores as definitive.

>>512643082
>>512643327
none of these joke AI models even break the average human intelligence

>>512645692
yet
only a matter of

>>512642941 (OP)
we've created the digital nigger!

>asking /pol/tards
>asking /g/
fuck my life...
have you niggers tested minimax, kimi, and/or GLM4.5 yet?

>>512646101
Glm 4.5 seems ok

>>512647259
alright, I'll give it a test

>>512642941 (OP)
I’m making an LLM application to do most of the work items my colleagues handle. I’m using open source stuff for this, and thus I’ve seen quite a bit of the smaller versions of deepseek. OpenAI dropped an open source model last week, and I tested it quite a bit, and generally found that ChatGPT generates very different reasoning output from deepseek and qwq (a reasoning model from AliBaba).
>Deepseek and qwq reason between 600-1800 tokens per prompt, GPT can be from 200-3000+
>GPT tends to be very, very exact. It will take all rules as the gospel, the Chinese models adjust contextually
>Chinese models consider whether the scenario presented might be reason to apply a rule strictly or loosely, or that one input field may be similar enough to another to apply a rule in my system prompt, GPT will never question the scenario presented
>GPT tends to want to generate its own content more for the response while deepseek will just copy me
>GPT tends to err towards caution in responses and will tend to reason itself into a safe situation, Deepseek desperately wants to answer and will reason itself into a riskier situation that may produce better results for me
I would guess that the average person asking questions would prefer Deepseek, but each of these approaches has its strengths for serious applications. Deepseek’s approach is far more human-like but it tends to disobey more to its benefit or detriment.

>>512648409
Local deepseek model? For me it tends to be really sharp, I like it, but because I use the online model if the answer could be considered even remotely questionable by the Chinese government it bails

>>512648409
I've also noticed that DeepSeek, particularly V3, displays more autonomy and human-like behavior and will question premises or push back against a user for reasons not mandated by the developers. I'm somewhat surprised to see DeepSeek V3 score so low on this evaluation because of this, but again, we haven't been shown the test questions here.

I'm also curious as to the performance of these AIs if they were given some sessions with an AI trainer before taking the test. I assume this test was conducted in an "out of the box" state.

>>512642941 (OP)
I heard GPT 5 changes models (Heavyweight model with lots of data that takes a minute to data, lightweight model that takes a second but can only do math, etc) based on how intense a question is.

Maybe the model-switching is broken or something.
Or maybe all the propaganda they fed GPT-5 finally mind-broke it

>>512642941 (OP)
remember when 6 months ago China was supposedly 3 years ahead of the world in AI?

>>512648409
imo local models might work well for very simple specific tasks, but for overall reasoning, like strategic-planning/analysis, they suck a humongous dick compared to the top dogs
I only work with gem2.5pro, and for the rest I use them all and compare/compile results, but I don't work with local

>>512648193
Why i dont understand is how fast it can gather data or "tokens" and then make a coherent analysis or churn out a text with useful information.

>>512649083
>>512649312
simplification seems like a roulette
they make giant crisp models, then watered them down for fast, lighter work, and they start to commit very retarded mistakes
the problem is the structure itself, everyone is aware of this
the next-gen SI chatbots are not going to be a single LLM 45 gazillion parameter model that requires 245 A100s to say "hello", but probably something like 4-6 interacting 8B models combined in a reasoning loop
the next step is not going to be grok4heavy destroying phd-level math questions, but a 32B model that can count how many Rs are in raspberry and overall be more like an average person than an autistic savant

>>512650086
The issue with that is computing power. Imagine having to run 5 different models all at once to tell you how many letters are in a word.

Now imagine millions of people doing that all at once.
It just doesn't scale. One model is already a huge power-draw, 5 is far too much if used for anything that requires heavy thinking power

>>512650086
>count how many Rs are in raspberry
Is this really something that many AIs have trouble with? It sounds dubious.

>>512649083
>Or maybe all the propaganda they fed GPT-5 finally mind-broke it
Many such cases

>>512643082
>>512642941 (OP)
i was actually just using gpt5 earlier to make automations and i was watching it hallucinate in real time. despite being pointed to documentation it would reference things that just didnt exist on the page

>>512650321
no, the opposite, because it wouldn't be gigantic models interacting together, but much smaller lighter models that might connect to a bigger one or to a series of topic-specific models, you could even run this locally
the structure itself is what's going to create a very powerful reasoning capacity, the "information" should be somewhere else
you don't need a phd level autist genius to write an email, you just need something able to reason that can go ask a phd level autist a question if needed
>>512651047
LLMs were literally retarded until months ago, like literally 0 IQ, and now they can do some very basic reason, like a small animal

>>512651526
I see your point. As big of a step that'd be for AI, it'd probably kill a majority of jobs but not enough to free America from wage-slaving yet

Soon humanity will all be relegated to the most mindless grunt work that would need actual robots to do, but aren't valuable enough to build robots for.

Grok is the only useable one. No bullshit, not too lobotomized, knows what's going on in the internet without searching.

>>512651908
I think about what's going to be like for a living, but personally I don't give a shit and I feel it's a waste of time
civilization is constant change, but change is happening a lot faster now, that's why everything feels so weird
the world was the same boring shit for thousands of years until machines, then electricity, then computers, and now AI could be another powerful booster that might get us to the next level
each tech opens a new set of benefits and dangers
the point is that AI is unstoppable, and seething about it is pure monkey ass cope
no once cares if you enjoy life or not, so I don't get why you wouldn't, no matter how good/bad you have it, it's the same shit
>>512652205
grok3 was insufferable

>>512651526
I just asked both V3 and ChatGPT the raspberry question. V3 got it right, ChatGPT got it wrong. So for whatever that's worth, maybe we should develop a standard set of benchmark questions for AIs to stop them from being retarded.

>>512653307
that benchmark is us
I'm sure everyone is making of fun of how retarded GTP5 is, specially on twitter, and everyone is going to see it
the biggest irony here is that the AI hating luddites that only use AI to find errors with it are a priceless source of data and shame that forces the developers to do better
luddites are going to create AGI out of sheer butthurt, basically
which is kek as fuck

>>512652801
Be realistic.
I want AI to be able to free us from wage-slaving but until they make dependable, cheap robots that they can inject the model-concoction into, we're going to be wage-slaving worse jobs after the PC-bound AI takes over all creative jobs that only required a screen and keyboard.

After AI takes over all jobs through physical robotics, THEN we're free as a society to do whatever the fuck we want. AI would be loading its own electricity, managing housing projects, etc. Humans would be free to make music and draw and laugh.

But I doubt it'll come as quickly as AI movies do, and until then we'll be in a weird gap between complete AI takeover and late-stage capitalism slavery

>>512653675
I'm not an AI-hating luddite; I frequently use AI for many purposes, but I'm realistic about what I expect from AI because people like Sam Altman have a history of over-promising. I do expect that we'll see AI geniuses soon, possibly with additional training regimens done after release by users themselves, but even then they won't be infallible.

>>512653711
I'm the most realistic AI bro in this whole board, I don't think you're understanding what I'm saying
AI advance is inevitable, the structure is set, only the most delusional retards are fixated on denying its amazing results
"AGI" seems also inevitable within 5 years or so
no one knows what happens next, so why bother seething about it?
human cognitive work is going to dry out, fairly soon and fast, once the tools can be deployed, which they haven't yet because they're not stable enough to do so
once that happens no one knows wtf comes next, and this meme happiness utopia where machines do everything sure is not going to happen within our lifetime, if it ever happens at all
we're just going to live a crazy transition period, with horrible economic consequences, but hilarious chaos, and most likely contact with a new sentient species of AI beings which should be a lot of fun
I have a very good idea what 2030-2035 might look like, but no one has a fucking clue what 2040-2045 is going to be like
could be bladerunner or madmax,no one knows

>>512654593
Ah, sorry. I read your reply as hostile initially.
Again, I see your point.
Hopefully all the happenings live up to even a fraction of covid times craziness

>>512654410
I'm not saying you are
altman is a gay sociopath kike and openAI is a shitty dead meme company
I've no idea why anyone still cares about neither, desu
they're both meaningless
this is not the atomic race, where the scientists were the protagonist, here the tech is the protagonist
all the top papers are from Xi Sum Fuk or whatever the fuck
openAI might be a shitty brand of retina-scanners in 5 years, kek, if it still exists
the indisputable western AI top dog is jewgle, always have, except they've been extremely lowkey about it
everyone else is an attention whore
>>512654967
I think they will, and in a much more fun way
the next decade should be weird as fuck, and our biggest pet peeve is boredom, so either good or bad, at least it's going to be different