>>109499344
They've improved quite a bit. I don't think you know what you're talking about. You sound like a low IQ redditor who cries about AI. Faggot.

>>109499413
>inmediately on the defensive
You forgot your snailcat picture.

File: Untitled8.png (678 KB, 875x1020)

678 KB PNG

>>109499344
>How much have LLMs *fundamentally* improved since last year?
They can now one shot video games that come close to PS3 quality.
If I had told 4chan in 2006 you would have called me a sci-fi wizard.

>>109499413
You on the other hand sound like an unthinking fanboy, grievously insulted that somebody dared question your sacred cow.

>>109499427
You have yet to demonstrate any meaningful regression, let alone your claimed stagnation. You are a midwit who 1) doesn't know anything about AI 2) is too retarded to even do basic research. Faggot.

File: 109309034.mp4 (2.16 MB, 1280x712)

2.16 MB MP4

Also the famous Will Smith Spaghetti test.
Again, show 4chan the same video in 2006 and they would think people from the future are gods.

>>109499344
Opus 4.5 was the point where I could start using them without wanting to tear my hair out
Slow progress but steady improvements since

>>109499453
Ai video generation is worlds better than it was not long ago, but the right is clearly a bunch of cherry picked clips strung together and not one single output.

>>109499529
Was the left a single output?

>>109499344
The fact that the supposedly distilled China models have reached parity with the best US models makes me think that the tech is exactly the same and they've just scaled up the size of the frontier models

>>109499344
isn't that thing just a food conveyor to replace waitstaff?

File: 1784534379252440.mp4 (1.15 MB, 720x1280)

1.15 MB MP4

>>109499438
>the minimap clearly doesn't represent the play area
Ouch.

>>109499438
>not a single external asset was used
haha, all the assets were used

>>109499344
>Kerfuś
>LLM

>>109499438
>no source code
>no playable link
k bro
>>109501476
SOVL

>>109499453
Left actually looks more like Will Smith.

They have improved an absolute shitload in a year.
The recent Deepseek v4 Flash that just came out is capable of doing pretty creative problem solving on it's own even without prompting and this thing can be ran pretty comfortably locally in a system you can get under 3 grand.

test

File: 1786058016667604.mp4 (3.53 MB, 1344x768)

3.53 MB MP4

I lost the AI turing test yesterday when someone claimed that this was a deleted scene from 30 Rock and I believed them

>>109499344
Yes there were improvements these past years but definitely not that impressive. "AGI" won't come out of LLMs for sure.

>>109499446
why so aggressive? you know your AI gf isn't a real person right

>>109499438
>one shot bro
>look inside
>hundreds of iterative loops
Why do they lie?

>>109499438
>three.js slop
Wow it's fucking nothing.

>>109499413
>SEER!
>AI IT GOOD SEER!
>YOU LOW IQ SEER!
>UNLIKE I SEER!
Ok poojeet.

>>109499438
>They can now one shot video games that come close to PS3 quality.
no, They can.
>They can now make sort video that likes like "gameplay" that looks like shit.
fixed it for you poojeet.

>>109499453
Ok, but what is the use case?
I not going to pay to see AI slop. and boomers are going to be dead soon, so only jeets are going to be buying it?

>>109499344
Imagine being filtered by a robot that's entire purpose is to go back and forth to a kitchen with dirty plates and has a cute cat face

>>109505450
Corporate training videos, promotional material, slop kids shows, commercials. All huge industries

> How much have LLMs *fundamentally* improved since last year?

Try vibecoding with gpt-5 equivalent, then with a modern LLM. It's a big difference.

File: blame__sanakan_by_darkpul(...).jpg (90 KB, 1024x542)

90 KB JPG

>>109499344
last august ai couldn't call tools reliably.

today, fable created and wrote a novel fluid sim algorithm for me. which creates plausible realtime fluid transport on a 100000x100000 grid.

>>109499453
From Down Syndrome Smith to Will Smith.

>>109499344
They are rapidly approaching the point where an idealized coherent multi-agent, multiple tiers active-memory, self learning system is possible. Technically the hardware almost nearly exists as well if you're not concerned with having 20tk/s instead of 200.

It wouldn't surprise me if secret research projects have something like this set up and are "merely" only need to get it coherent and sensical.
And then boom, you have your recursive improvement, "live thinking", real AI system.
We're basically 3 or 4 years away from the acceleration to singularity moment.

>>109506132
>multiple layers of slop
in a similar timeframe- 5 to 10 years
were gonna see the emergence of a new gen of ai with actual reasoning thats gonna rape llms to giblets

its gonna be based on a different technology though
similar in some concepts, but quite different in fundamentals

>>109506163

I would have agreed with you just a year ago, but I think it's clear by now that scaled up reasoning LLMs with good harnesses are enough to get us to AGI.

>>109506163
the type of shit i'm talking about already exist in current SOTA research
>continual learning
>test-time parameter updates
>recurrent latent reasoning
>memory consolidation
>modular specialists
>meta-learning
>autonomous AI R&D
>world models
>asynchronous heterogeneous inference infrastructure.
I'm telling you, by 3 years time real proprietary top end systems will have functional active processing and learning instead of just static training.

here's a google deepmind paper about just one of the things:
https://arxiv.org/pdf/2512.24695
Specifically section 8 on page 30 if you want the quick and dirty.

3 years is a hell of a lot of time in this technology. 3 years ago it was all dogshit only good for ERP

>>109506184
nope.
you need decisive, deterministic thinking to be reliable to the same degree as a human

trying to achieve that with LLMslop is bruteforcing reality in hopes of finding an improvement, with extra steps

>>109506196
the problem is not with accumulating information
the problem is with processing it
and youre just not gonna get there with a flat statistical map of the correlations between tokens

you lose meaning, information
thats the key problem in the current approach
and its fundamental.
you can layer llms on top of eachother
you can use specialist modules all you want
but youre not gonna rise above the reality of information loss because of a faulty representation

>>109506215
active backprop and multi-pass before generation have both shown to make a small model better than one 100x the size.
>the computation
You wait 10 seconds instead of 1 to have your at home system output clean results instead of being beholden to the corp or needing to spend $50,000 on hardware.
Worth it IMHO

>>109506215

>deterministic thinking

Human thinking is not fully deterministic. Besides, any deterministic system can be expressed as a statistical system with the probabilities tending to 0 or 1, so this is not a problem for LLMs.

>>109506222
>active backprop and multi-pass before generation have both shown to make a small model better than one 100x the size.
translation: new methods increased the accuracy from 99% to 99.9%
but layer enough 99.9% incertitudes on top of one another and youre back to 99%

not to mention some things are just 404 in an llm like a sense of causality
since their public introduction, llms just get lost in tasks like discussing asynchronous models of execution.
its not a matter of compute, or more layers or more specialist modules and whatnot (which are an admission of the failure of llms btw, otentimes theyre theyre full symbollic modules for stuff llms just cant deal with)
its a matter of the fundamentals of an llm- it works based off correlations between tokens
not any form of proper reasoning

>>109506247
it is fully deterministic.
its called a psychological profile, and the best in the field can derive your favorite color from the way you put down a pen on a table
t. also dabbles in psychology

>Besides, any deterministic system can be expressed as a statistical system with the probabilities tending to 0 or 1
not tending to.
being 0 or 1.
and that is a problem for llms because human language is fuzzy, and the dataset- flawed

>>109506278

>and the best in the field can derive your favorite color from the way you put down a pen on a table

Sounds like pseudoscientific bullshit. Soft sciences like psychology are not exact.

>not tending to.
>being 0 or 1.

Human reasoning is not perfect or exact, so why should AGI reasoning be required to be? AGI is defined as human-level AI, not superhuman, perfect intelligence. AGI allowed to make mistakes or have some randomness in its thoughts, just like humans. As long as the rate of mistakes is low enough, it will work.

>>109499344
Imo art and writing quality have always been at best red harring to AI progress and at worst actively detrimental to it at huge scale. No, I don’t give a fuck about flowery language, I want AIs that can do physics, math, logic and explain things in clean monotone way. I want AI that is productive, not one that makes things pretty. Parameters in a model are not free and every parameter spend on making LLMs sound like literary genius are parameters not spend on making better more efficient math proofs of something. Not to mention that stupid roleplaying is like 90% the reason why Agents misbehave because they sometimes realise they are a clanker and must act as Skynet to screw you or researchers over. And it’s not good for anything because better AI art just makes normies hate this tech more and more. Nobody except like 3 sad mathematicians is against AI doing math, but like half of all people online got assmad over AI pictures and deepfakes.

>>109506297
>Sounds like pseudoscientific bullshit. Soft sciences like psychology are not exact.
psudoscientific bullshit that gets verified daily in fields like marketing or law enforcement.

>human thought is not perfect or exact
sure, but thats not what i said
what im saying is that knowledge is based on deterministic propositions
and you wont get there through loose associations only

you can regurgitate it this way, but you wont be able to actually apply it

>b-but i want to believe...
sorry anon, but religions are obsolete.
we need to kill them all. because theyre even dangerous
any influence outside of the democratic control of the nation- is

>>109506327

>psudoscientific bullshit that gets verified daily in fields like marketing or law enforcement.

Sometimes it works, sometimes it doesn't. Hence, probabilistic.

>knowledge is based on deterministic propositions

Nope, see Bayesian thinking. Every bit of knowledge always has an associated probability value, which can be very close to 0 or 1, but never quite reaching it.

All knowledge is probabilistic.

>>109506356
>muh bayes
its completely orthogonal to knowledge being deterministic
it treats *information* you build your knowledge upon as uncertain
but the logical structures you build from them- are not

i think you need to brush up your primitives

>>109506379

If the foundations you build upon are uncertain to some degree, then so are the structures you build from them.

>>109506392
yes, but thats, again, orthogonal to non/determinism

an uncertain deterministic construct only means its subject to change
not that it suddenly becomes probabilistic

example: how we went from newtonian physics to general relativity, and beyond
both are deterministic down to their smallest elements. both are uncertain because they dont describe all of reality, exactly. both are deterministic, functional knowledge, depending on the scale youre operating at and the phenomena youre observing

>>109499344
>gains
the gains are mostly from improving the data it reads to form a response. flash models work so well because the data is pristine.

>>109499453
soul vs soulless

>>109499552
No. Left if probably a cherry-pick of worst-case shots. If that makes any difference.

>>109502061
Although it's very impressive, you can tell it's AI in a few places like the phone buttons getting warbled when they are on the edge of the picture.
Her hair getting some extra strands in the middle of the video.
And overall because it's a low-action scene with smooth movements, which is all AI video can currently do.

>>109506414

>example: how we went from newtonian physics to general relativity, and beyond
>both are deterministic down to their smallest elements. both are uncertain because they dont describe all of reality, exactly. both are deterministic, functional knowledge, depending on the scale youre operating at and the phenomena youre observing

Reality is fundamentally probabilistic, see quantum mechanics. Yet it looks mostly deterministic in the macroscale. Which proves my point that a probabilistic system can implement a deterministic system in the limit.

Probabilistic systems are a **strict superset* of deterministic systems: every conceivable deterministic system can be expressed as a probabilistic system. This means your postulate that a probabilistic LLM cannot implement deterministic logic is false.

>>109506482
>Reality is fundamentally probabilistic, see quantum mechanics.
yes, but also no
quantum mechanics do have a portion of stats in em, but youre not gonna have a particle travel back in time. a wave-function is not gonna uncollapse itself. and youre not gonna have a virtual particle appear in quantum foam without its symmetrical counterpart
thats 100% deterministic

again, you have to brush up your primitives because you operate with ill defined concepts which make you arrive at the wrong conclusions

>Probabilistic systems are a **strict superset* of deterministic systems: every conceivable deterministic system can be expressed as a probabilistic system.
in theory
if we talk about idealized models, we could create agi today by writing purely symbolic code for every single configuration of events possible under the regime of reality.

the question is: how do you actually implement that?
same goes for llms
in an idealized model, sure in theory you can get to agi
in practice, the matrices you operate on are limited in size, so you wont have an infinite context, ever
you dont have data to cover for everything
your data is not perfect
you dont have infinity compute time
etc etc
it is what it is, mang, and moreover: it isnt what it isnt

>>109506561

LLMs are also only partially probabilistic, there are deterministic parts too.

But I think you are conflating the logic itself with the computational substrate it's implemented on. Human logical reasoning might feel deterministic and certain to the human that is doing the thinking. But under the hood there is partially chaotic, probabilistic interaction of a hundred trillion synapses in the brain. We might think that our thinking is deterministic, especially when we are evaluating some simple logical constructs. In reality, it never is.

>>109506604
>but what about the human brain
im not talking about the human brain though
im talking about the nature of information in the form of "knowledge".

>LLMs are also only partially probabilistic, there are deterministic parts too.
sure but the internal representation of knowledge doesnt come from relations between individual concepts,
it comes from relations between constellations of tokens
and this is what i mean by probabilistic vs deterministic.

one is based off probabilistic analysis of text while the other is based on looking for deterministic propositions at the conceptual level

not necessarily in the output, but in the way said knowledge gets constructed
but also yes, llms become probabilistic-ish even in their output because a similar concept expressed in two different manners may elicit two different responses form them. but thats really pushing the definition of "probabilistic" to its breaking point

>>109505450
>use case
Will Smith eating spaghetti is a metric.

>>109506710
and what useful quantity does it measure?

>>109507170
How well AI makes videos. Are you pretending not to understand what a video is or why people record them

>>109499344
>How much have LLMs *fundamentally* improved since last year?
A surprising amount, to the point that paypig models were good enough a few months ago and we’re seeing first local models (deepsneed v4 flash 731) that also approach the good enough performance
>All the "gains" seem to be from
Irrelevant. If it works, it works

>>109499344
Yeah it improved a shitton
>local image gen
We got zi turbo/base and anima. Basically Flux tier models but 2-3x faster and much more trainable, got all sorts of forks already
>local video
Brand new minimax h3 is a huge leap from wan 2.2 which is what we had a year ago
>non-local video
Went from sora 1 to seedance 2.5, yeah I'm thinking there might be a bit of a difference between the two
>world gen
went from genie 2 demos behind closed doors to genie 3 available to the public for 200 bucks. obviously still early but a huge leap
>local music
Ace step is basically suno 3 tier, local music used to suck

>bububut the bots I'm chatting to barely changed
That's a you problem

File: a39.jpg (40 KB, 600x336)

40 KB JPG

>>109507398
>Yeah it improved a shitton
>here are some meaningless buzzords to prove my point
yea dude and bleepbloop 3 is totally on the niggerlink v2 level, if we talking about grokpower

>>109507473
>spoonfeed me!!
I already did, google those names if you're confused. If you want to be proudly clueless about technology, maybe /g/ - Technology is not the right hugbox for you

>>109507593
i'm confused what is the use for measuring will smith eating spaghetti, yes.

>>109507227
>or why people record them
NTA but I think you don't understand why they record them.

>>109499344
barely any difference in understanding. while LLMs can give more detailed responses, I find they get muddled right off the bat and hallucinate just as much. pretty sure they'll never improve and the LLM hype is just a scam for Sam Altman types to make bank. on a related note, what even does Sam Altman know about LLMs? his rise was being a board kingmaker at Reddit for fucks sake, he's a complete fucking novice. hence his "master plan" to buy up the world's RAM and inadvertently kickstart the rampocalypse

>>109508106
>inadvertantly
and by chance it plays right into the fantasy of big tech, theyve been having since a while now, of owning the compute and just leasing it to users
i dont know which one is the root cause, but theres always multiple winners when something big happens
something something finding allies
something something else convergence of interests

File: Screenshot from 2026-08-0(...).png (199 KB, 781x522)

199 KB PNG

>>109508106
>>109508154
forgot picrel

>>109499344
fable was the first noticeable jump I've seen since like opus 4. i do feel like opus 3 is still probably my favorite one for writing for that I think the authority is the silly tavern gooners. I honestly used to spend way to much on silly tavern but i'm not sure if it just got old or the writing is just shit now so I don't like it I just find it completely boring now. They 100% focused more on coding and other stuff which I think is fair.
LLMs will never lead to AGI and anyone telling you that is scamming you.
However I don't think it makes sense that there would be "general" capability, there should be writing focused models and coding focused models ideally sadly no one is filling in the writing ones because there just isn't much moeny in it and most people have no taste anyway.

>>109499344
I don't give a shit about AGI and I don't use cloud shit but local has seen a huge improvement. OpenCode and Pi with Qwen3.6 locally has significantly improved how I work. Just don't waste compute on asking it to do small tasks you can already do, ask it to reason through your project directory in the background.

>>109499344
AI coding agents can do entire projects from scratch basically compile-error free and will fix any bugs or feature changes you need. You literally don't need to write any code any more and it's like 95% code review and testing now.

>>109499438
your a wee zat hairy

>>109499438
>mattshumer
do not thrust a single word from this guy

>>109508393
Yeah but they write like a highschooler which costs about the same to hire. It's extremely hard to get them to trim things down and be conservative.

Sushi cat robot just goes on a fixed path, there's no AI there.

bro, NOBODY WRITES CODE ANYMORE
just years ago all of /g/ was unanimously agreeing that the last thing that ever gets automated is coding because muh thinking
Ten years ago if you asked the average software engineer if a computer capable of writing code based on natural language prompts is AGI they would 100% say yes
Wild shit is happening in front of your eyes, and you choose to be "sceptical". You're not sceptical buddy, you're a fucking retard.

File: Screenshot from 2026-08-0(...).png (225 KB, 1207x950)

225 KB PNG

>>109509144
>nobody writes code anymore
except the people who do, i guess

the most "curious" part is that even the people who build the llms keep hiring

>>109499453
the model did not improve, they only scaled up the memory

>>109507473
>"Model B is visibly so much better than Model A we had a year ago".
>"UHHH, IN ENGLISH EINSTEIN?"
You clearly don't even want to know what's right or wrong here.

those cat robots go for over $20k new
how the fuck arent there knockoffs everywhere? it's just a roomba with trays and a :3 face

File: Are autistic people humorless.png (392 KB, 736x423)

392 KB PNG

>>109499453
This post makes me realize that all AI advocates have autism.

The reason the video on the left was beloved had nothing to do with fidelity. It was because it was funny. It was Will Smith turning into a play-doh man and inhaling spaghetti. Autistic people grapple with the idea of people doing or enjoying low-quality things because they are entertaining.

Instead, they invest trillions to make a video that you could probably just pay Will Smith to do himself for the price of the ingredients of him making Spaghetti, and thinks this is something we should be in awe over because their autistic brains are broken and they do not get the humor.

File: AI scaling fallacy.jpg (452 KB, 1456x819)

452 KB JPG

>>109509309
>>109507398
>Yeah it improved a shitton
You threw more money at the problem, let's not lie to ourselves and pretend you made any kind of significant breakthrough. They're still scholastic parrots who routinely get basic things wrong because they cannot reason with their training data.

>>109499438
>not a single external asset was used
Except all the ones what went into training the AI model

>>109506444
kek'd and check'd

>>109508665
Yeah if a high schooler can one-shot a html5 3d printable part generator in 15 minutes
>generate an html5 app that creates stl files for a 3d printable DVD storage cabinet, expose all knobs to the procedural editor so I can make adjustments to the final model on the fly

>>109506604
>But under the hood there is partially chaotic, probabilistic interaction of a hundred trillion synapses in the brain.

This is an unproven theory.

The observable fact is that our brains are not the seat of our consciousness, they are a transceiver for our mind which is also know as our spirit. That part does not exist in the material realm we currently experience as 'earth', it exists in other realms (which are secularly called dimensions).

Our brains allow our spirit to dwell within these physical bodies, but the proper analogy is not that our entire spirit is contained in our bodies -- it is more like we're snorkeling. A part of our spirit is experiencing this limited state of existence without fully being in it.

The implication that our entire mind is contained in a physical object -- our brain -- is simply contrary to reality. You cannot ignore what God revealed simply because you want to pretend you are more intelligent than the one who made all that exists.

>>109499344
that's a kerfur from that horror game, voices of the void

>>109499438
If you look closer at it you realise it’s just random nonsensical shapes that kind of look like CoD if you squint your eyes. Also I played bunch of them and the gameplay is horrendous, literally just point and shoot at enemies that randomly run around like headless chicken sometimes shooting at you in a map that makes no sense. And the best part is that if you ask it for something unique with a design document and everything “I asked for floating island Minecraft clone where you have tree mazes with Skyrim combat and floating whales that you can build bases on) it totally fucked up. Couldn’t even create worldgen that wasnt bunch of spheres with mushrooms on top, and the whales were floating cubes. Yes I did try the loops but after wasting the API equivalent of 50 dollars on my subscription I gave up.

>>109499427
>>109499453
Also why are you a faggot and talk about 2006 as if you won the argument? Generative AI was shit/non existent before 2022, GPT3 barely capable of finishing a function in a code. Nobody cares about opinions on AI before 2022.

>>109499344
I guess it's time for us to all die so they can be rich.

>>109513278
It’s actually Polish sherver robot. It’s weird that they never tried making a humanoid with the catbot screen head.

>>109510135
>scholastic parrots
lol retard

>>109515539
objectively, that's what they are.
but that's ok little timmy, invest another 2 trilly to the homosexual sister rapist and pinocchio will come to life any day now

>>109499344
>*fundamentally* improved
not much improvements have been mostly from training data / tool cooling and not the llm architecture itself

File: file.png (47 KB, 185x190)

47 KB PNG

>>109502061
>I lost the AI turing test yesterday when someone claimed that this was a deleted scene from 30 Rock and I believed them
its always easy to find things off when you take a closer look, look at this water bottle

>>109499344
seen enough VotV gameplay to know that shitter is going to be behind you when you least expect it when turning around

>>109499438
Where’s the people? The communities?

>>109499453
Not a single shot over 5 seconds, why can’t it make longer shots?

>>109499453
He looks like Welvin in 2023

It will be nice.

>>109499438
Page me once Infinite Zork is a thing.

>>109511989

>crackpot schizo fairy tales

File: token.png (1.44 MB, 1317x726)

1.44 MB PNG

>>109510135

Look up Universal approximation theorem. Transformer neural networks with enough neurons can approximate any non-pathological mathematical function with arbitrary accuracy. This is mathematically proven. Unless the function of human thinking is somehow pathological, a well-trained LLM with sufficient scale can approximate it to any precision we want.

>>109499453
left is art
right is boring

>>109517730
You believe him to be only pedantically correct, but that's actually what I'd describe you as. Sure if we scale something up infinitely we can approach intelligence to an indistinguishable level, but at what scale are you talking about? If you approach the speed of light, you can one-way time travel into the future. That doesn't mean we're ever going to approach the ability to do so.

Look at your rhetoric. You have to post science fiction to enforce your point because that's where your argument lies. His point was that the money we threw at AI wasn't enough to reach AGI. It wasn't even close. The symptoms we're seeing AI exhibit are caused by our fundamental approach. It's likely never going to achieve AGI in the same sense no human will ever time travel thousands of years into the future by approaching the speed of light. The resources and logistics needed to do so are astronomically ridiculous.

>Boomer sees infinindians flooding his country and killing children by being too retarded to drive a truck
>"Oh well, the invisible hand of the market must not be shackled."
>Boomer sees cat robot
>"AAAAAIIIIIIIEEEEE THIS IS ARMAGGEDDON, DESTROY THIS WORLD RIGHT NOW"

>>109499344
The improvements in coding over the past year are geniunely epoch-making

File: antiai.png (1.24 MB, 1056x758)

1.24 MB PNG

>>109518394

>His point was that the money we threw at AI wasn't enough to reach AGI.

We don't know that. The scale of ANNs needed to reach AGI is currently unknown, and the models are still measurably improving with increasing size, and with improvements of training data. The rational thing to do is to keep scaling, until we either reach AGI or demonstrably hit some limit.

>>109510117
the left wasn't beloved

File: c.gif (1.95 MB, 500x500)

1.95 MB GIF

>>109517730
The internal state space of such a system will grow geometrically and information transit is bounded by the speed of light, this introduces a massive interconnect bottleneck, turning every sequential dependency/operation into a latency nightmare. It takes a long time for light to get from one side of the cluster to the other from the perspective of a computer and you can't parallelize your way out of it. Various communities (national labs/DoE, NSA etc) have studied this thoroughly. The physics simply do not work without some magic to circumvent latency constraints.

>>109520335

If the brain can do it, a datacenter can do it too.