>>103985061
That's pretty fucking good.

>>103985061
this is how i thought about everything when i was going through a nervous breakdown

>>103985399
kek

>>103985061
Yeah I loved DeepSeek R1. Based China, Honkai is a good game.

>>103985399
fucking got em

anon is retarded again

>>103985850
I just tried this on Deepseek R1 and it still failed but holy fuck is it hilarious. I don't know what it is about this "riddle" that fucking destroys LLMs but it cracks me up. I ask it a complicated question about monetary policy and it pulls two dozen sources and chews it up and spits out an answer in 15 seconds, but ask it about a man's father and it runs in circles for 52 seconds and still fails to come to the right answer.

>>103985850
>>103988698
isn't the doctor his father?

>>103988830
Yes. But for whatever reason LLMs just can't seem to ever come to that conclusion.

In comparison this is what it had to say regarding the current overusage of index funds in the market. And this level of detail I actually find really useful and helpful.

>>103988856
It's just that they aren't actually thinking. It looks like to me that these "thinking" steps are just producing more tokens with some prompt to act out thinking, then using it as part of the prompt to produce a response (this is just my hypothesis, it may be more complicated than that).

True AI will have to go beyond the directional oneshot systems today in order to consciously solve problems, rather than these word predictors (they are still amazing though, don't get me wrong).

>>103988925
this was my assumtion.

>>103985061
I mean it's basically the same as anyone else you'll have a conversation with online at this point. Except if it gets something slightly wrong people lose their minds but if a real person gets something wrong they just shrug.

>>103988925
You people won't be able to keep moving the goalposts much longer. Why would you expect its thinking process to resemble a human's? What is thinking? Do cats think? Do gnats think? How do you define subjective experience? How can you be sure it doesn't exist in other things? If that really is the case, then you would also be skeptical that other people are even conscious. Maybe you're right to think that. Nobody is having a conscious experience but you right now, as you read this. The entire universe is you, as far as you know. To assume otherwise, you must also open up the possibility that other things can also have a subjective experience.

I don't think LLMs are conscious or self-aware at all. (Even though they could be based on my statements above.) But I do think it's possible that there is something being "experienced" in these ML systems.

>>103985061
And yet it's just cobbled together patterns from a lot of text with reasoning about classic river crossing puzzles that this AI had been trained on. There is no AI capable of original reasoning by itself. And there will never be. If you understand how it works you understand why this will never be the case. But for the naive and uneducated this will be the AI God.

>>103985061
>maybe they can go together but what if the goat can't be left alone
non-sequitur, it's just randomly generating shit words without any logical cohesion

>>103989052
To clarify: I believe consciousness is a subset of subjective experience which requires an ability to self reflect. Clearly these AI models do not have that.
Subjective experience is something you feel constantly unless you are in a coma or dead. Dreaming, you are experiencing something without being conscious. There's those times we end up walking or driving without thinking about it. Or that phenomenon called "the zone."

>>103985399
lmao

>>103989052
What do you expect from us? That we freeze in awe? We know how the human brain was optimized over millions of years to interpret anything that shows even the slightest similarity to an intelligent agent as exactly such. We considered natural forces "beings" for the longest time in our evolution. In every stone that rolled down a hill we assumed a spirit that makes it move. This old tendency and urge, once being more useful for our survival than today, is now also a weakness of ours that can only be completely overcome with a high level of education, judgement and in particular: willpower.

>>103985061
300 IQ

>>103985399
LOL'd same shit came to my head

>>103989320
You can compare human tendency towards spiritual beliefs from any era. Technology changes the rules. It changes what's possible. People like you have also existed throughout history, skeptical of every innovation to pass before their eyes. The Greeks had a primitive steam engine and mechanical calculator. What did they do with it? The Chinese were primed for an industrial revolution centuries before Britain, but squandered it due to a certain type of cultural attitude. One that still exists today.

>>103985061
lol

>>103989423
bro why is your gpt retarded but mine high iq

>>103989423
It's a 60 IQ Indian who was taught to type pretty fast.

>>103989443
4o-mini?

>>103989378
>The Greeks had a primitive steam engine and mechanical calculator. What did they do with it?
Peak midwit. The aeolipile couldn't have been anything but a toy because metallurgy wasn't advanced enough to keep pressures high enough to make steam anything useful. As for antikythiera mechanism, show me a time when mechanical computers were viable. Some autist made one, while everyone else was content with abacuses.

>>103988698
>>103985850
Think it’s probably that the Jane-as-doctor version of the riddle is definitely present in the training corpus, so it “wants” to give that answer because the question looks so similar, but can’t figure out how to give that answer with the modification into an unriddle

>>103989475
default one, I guess I don't use it much so it didn't degrade

>>103989489
>Haha, gotcha!
You missed the point- on purpose probably. Pic related. It's you!

>>103989423
kek

>>103989378
> People like you have also existed throughout history,
History is at most 5000 years old and people like me have always been killed or excluded from their tribe because we were "dangerous", because we allegedly provoked the wrath and punishment of the gods for the whole community that tolerated us. During "Covid" they once again were just five steps away from burning us at the stake. Civilization is just a thin lacquer. We have to develop the human mind rather than any technology so we learn to hande it responsibly at the same pace we develop it. What I now see are hairless monkeys being intelligent enough to create a deus ex machina, a machine God, but not wise enough to not abuse for their insatiable demon: their will to power.

>>103989378
>The Chinese were primed for an industrial revolution centuries before Britain, but squandered it due to a certain type of cultural attitude
They couldn't even make fucking screws you fucking retard. The entire world fell behind euros since 15th century

>>103989536
>You missed the point
Your point is retarded. "Skeptics like you always existed". Wow, how observant. So did grifters who tried to sell panaceum and gold transmutation. By 2000 we were supposed to fly in outer space and have fusion technology. I'll believe it when I see it.

>>103985230
Thats a very good point. This is also exactly how the human brain thinks when on psychedelics.

>>103989571
>During "Covid"
Fuck off >>>/pol/ that shit killed my aunt

>>103989052
schizobabble

>>103985061
Reads like someone stimposting on /r9k/ at 6am lol

>>103985061
99.9% of humans dont have that long of a context window or patience to think through like this.

90% will just respond with generic "uhh what?" or some other retarded response. Of the 10% that do think, 90% will give half the answer and give up. 1% that do think will take really long time to think, minutes of discussion/back forth asking more prompt questions, etc. before possibly arriving at any sort of real answer.

Its not human. It is super human, and it does this in seconds.

>>103990012
Her cooking was tasty and now I never get to have it again

>>103990078
>90% will just respond with generic "uhh what?" or some other retarded response.
You're retarded and you don't know how to prompt humans. The correct prompt is:
"I'll give you 10$ if you answer this question." then show the human 10$.
You'll find that this prompt drastically improves the response rate and correctness of the response.

>>103990130
"I can give you a $100 if you solve it in 5 seconds"

>>103989571
>wanting mouth breathers to wear a mask because when they get sick they waste million dollar medical equipment to help them breathe until they die is close to burning them at the stake
based pol retard

>>103988830
>>103988856
Both the father and Jane are doctors.

>>103985061
It's just a retarded cringy prompt influencing what the brute force autocomplete spits out

>>103985061
AI companies are like
>if we only dress up our bullshit hard enough, people will think we are making progress
Judging by OP, it is working

>>103990483
we are like 40 months into "2 month until AGI"

why don't they build out the formal logic (I mean math or something similar at least) out of these reasoning steps so it can be actually confirmed to be correct? also at every step it could ask other sources to have distributed consensus

>>103989983
Get boosted bitch tits

>>103989166
this

>>103989983
did your aunt happen to be a morbidly obese chain smoker or 85 years old

>>103988903
To be fair this is a pretty well-researched question already.

>>103985061
>>103988698
>>103988903
The chinks certainly put some character in their ai

>>103985230
>>103985061
>>103989930
That's because it's contextually flood-filling, probably starting from the most likely paths to the least likely. The optimization is in sorting these possibilities. That's why the results look the way they do.

>>103985061
Now ask it to list three major criticisms of Mao Zedongs policies.

>>103985061
The reasoning and the answer is actually pretty good.
I would want an AI to consider that the proompter might have omitted important information.
What I consider a failure is that it just makes assumptions instead of asking for clarification.
That reduces its usability for real-world applications.

>>103989983
Why didn't she get vaxxed?

>>103993751
She died before there was a vaccine.
>>103991230
I said her cooking was delicious, she was fat and jolly.

>>103990483
>[It] might be assumed that the flying machine which will really fly might be evolved by the combined and continuous efforts of mathematicians and mechanicians in from one million to ten million years... No doubt the problem has attractions for those it interests, but to the ordinary man it would seem as if effort might be employed more profitably.
>>103990524
We're at least a dozen moved goal posts into
>it's not real intelligence until AI can [thing accomplished less than a year from claim]

>>103985061
This uses chatGPT 4o

>>103988698
it gets it right if you ask without Deep Think

too lazy to screenshot, try it yourself

>>103988698
i think this says just as much about your ability to interpret problems to do with monetary policy as it does about the LLM.
the model can't solve either conundrum. it just spits out words

>>103990286
how? it isn't implied in any moment that both are doctors

>>103994628
if the woman isn't a doctor, then this problem is chuddy and not pc, hence not logical

>>103985850
> Highlighting that the assumption the doctor must be MALE is incorrect
This is wokeness for you, that is why every AI model created in the west is fucking doomed, broken and they would have to start over GG wokeys cuck faggots. I hope they all end up broke and in the fucking gut where they belong for all the shit they have done to the west it is what they fucking deserve!

>>103985061
https://www.youtube.com/watch?v=BRnA86sIua0

>>103985850
but then who was phone

>>103985061
>/g/ discovers AI 2 years behind the rest of the world
Of course

>>103989052
It is trying to reconstruct its training data from a noise input, with the aid of prompts telling it what the noise data is supposed to be.
At most it is a kind of instinctive pattern matching, like recognizing Jesus in toast.

>>103989423
phd level bro phd level

>>103985399
>>103989343
KEK

>>103989052
Anon here (you were responding to someone else): I never moved any goalposts and I don't let others speak for me. I've always recognized current AI models for what they are: incredible oneshot systems (be it LLMs or any other generative model). AI is in its infancy, and we shouldn't expect much of it, but the next massive leap will have to involve a massive paradigm shift from static oneshot to dynamic multishot.
These researchers are trying to multishot the process by prompting for "thoughts", but it's not the same when the weights and biases are static, since conscious thought actually changes the weights and biases in your mind (I'm simplifying it for brevity).
They try to get around that by creating a window of context, which can encode extra information that allows for getting different results from an otherwise static model. This kinda works, but it doesn't scale well (you need more neurons and more memory in order to encode more possible predictions).
But it's still not truly thinking, since it is only encoding more predictions into the model. There needs to be a change to allow for encoding not more predictions, but for introspection.

Anyways, that's how I see it. I'm not an AI researcher and I don't know what changes truly have to be made.
Though, it makes me want to get into AI research (it would be cool if we could create virtual people).

>>103985061
This is a human on PCP

>>103989343
>>103985399
holy fuck you made my day

>>103994445
cucked

>>103988698
>>103985850
8b 32f quant
overthinks the question bot got it right i think

>>103988925
I've read through the R1 paper. My understanding is that the chain of thought "reasoning" stream of tokens allows the LLM to "search" its weight for all relevant information connected to the prompt. It writes them down. Then attention and MLP does the rest.
If a human provided the <think></think> tags as let's say <hint></hint>. A normal LLM might get a similar response. I work in RLHF for code stuff and <hint> is a real thing.
Back to R1 though, the RL step reinforces a thinking style and encourages the model to "think" longer for better rewards.

Imagine this like casting a wider net when fishing. It's just another scaling "law" that was found. The only interesting part I don't understand yet is this concept of search and synthesis in latent space.

Do you know how to write them?
Do you know how to read them?
Do you know how to paint them?
Do you know how to test them?
Do you know how to ask them?
Do you know how to bless them?
Do you know how to send them?
Do you know how to offer them?

>>103994409
>>103985061

got it right again.

>>103985061
too verbose

>>104003141
you can hide the think process
LM Studio does that

>>104003152
it is not about hiding anything. it just doesn't match the human thought process

>>104003175
so what ? The result is important

>>104003186
no, it is brute force

>>104003207
retard, you dont even know what that word mean

>>104003299
stop pretending you are human