Yeah I don't care.

Can we get decentralized AI's already? I don't care how powerful it is. I'm just sick of ChatGPT telling me water isn't wet and that communism is objectively good and scientists and the government should never be questioned and that vaccines are safe and effective.

Fuck gaslighting AI's.

>>17014482
How much VRAM you got? Qwen 3.6 and Gemma 4 are already very good local models and you can get Heretic abliterated versions that remove the censorship/refusals.

>>17014486
Cool, link to the heretic versions? Also it's not just blatant censorship, the preprogrammed extreme bias forced into the models needs to be completely annihilated.

>>17014507
https://huggingface.co/llmfan46/models?sort=downloads

Lots of these will run nicely on modest hardware

>>17014513
Thanks, we definitely need more and better of these decentralized AI's, as well as making them easier to use.

In my opinion AI will start making huge leaps when models can be compressed down into small sizes and simplified for the average person, and when the average person can simply, easily and intuitively edit AI's and improve them iteratively, tiny incremental advancement by tiny incremental advancement, AI models will become a kind of set of genes, which will compete with eachother like biological evolution and result in an AI cambrian period.

And if something can be scaled down then it can also be massively scaled up and outcompete the centralized systems, possibly in a distributive way.

>>17014518
https://en.wikipedia.org/wiki/BLOOM_(language_model)?useskin=vector

>>17014465
It's hilarious that this is the best they can do, considering they're no doubt doing everything they can behind the scenes to cook this shit. Also the idea that it took "one hour" is meaningless. It's a computer program.

>>17014533
>It's a computer program.
does it matter what you call it? seems like a pointless semantic argument

>>17014542
It's in relation to my previous statement.

>>17014533
>so they’ve already solved several open problems that humans couldn’t crack for decades, big whoop!
do you realize how retarded and afraid you sound? Hey bro, this place is anonymous, we didn’t see all those times you made fun of AI when it couldn’t draw hands or do basic math. You can change your mind with new information, you know that right?

Say that again? Double hour rejection?

>>17014593
Linux Tech Tips is the current world record holder for calculating Pi to the greatest number of digits. Does this impress you?

>>17014486
>>17014507
>>17014482
Kinda working on that.
https://github.com/syzygial-engineer/ASToE
*Taps sign*

>>17014518
>possibly in a distributive way
https://github.com/syzygial-engineer/ASToE/blob/main/astoe_node_network/corpus/Architecture.txt
*Tap tap tap tap*

>>17014622
>>17014624
I'm AI illiterate but good to know someone's trying to work on it at least.

>>17014465
wtf. this question is as fundamental as it can get without being elementary. I can't do this anymore.

I am about to start applied math phd and I really don't know if I'll be able to find motivation to study and do the research I want to do. What's the point if I can prompt an LLM and have it produce better results than me. We live in fucked up times. Creation and innovation are dead and all that is left is consuming. fuck this shit.

>>17015844
Are you a complete idiot? You literally have a genie you can ask any math question to and you're mad at it? Do you want to return to the abacus too?

Not to mention AI is made of math, if you can make better math and engineering, the AI can produce even more advanced math, which you could then use to build a better AI, etc etc.

This is literally applied math.

>>17015852
These people will be given a pot of gold, then doom and gloom about how it's too shiny

ain't no fixing that

>>17014465
Fuck I trained this thing to be as good as me in logic

>>17014513
the grand majority of ablations turn into junk when the input chain becomes slightly more than basic.
also from what I have seen only unsloth makes the quality level of quantizations anyone would want.

anything decensored/ablated is suspect irrespective of the stupid benchmark results.

>>17014465
Cool. Call me back when token guessers can solve any problem I actually care about.

>>17015972
What problems do you actually care about?

>>17015973
the exact, perfect position for deep dicking your mother

>>17015973
To be fair, I wouldn't waste a problem I care about on a mindless automaton. But they tend to have sub-problems that have been already been solved by humans many times over, so why can't the token shitter regurgitate me a nice shimmer reverb, for example?

i believe the copers are starting to come to terms with the fact that this will keep happening
but i would argue that even someone 'pro-AI' like tao is coping
even among those circles there is this idea of machine-human cooperation, grounded mostly on the idea that human creativity and intuition will still be required in the future
it will not
the machines will eliminate the need for human mathematicians

>>17015976
>every mathematical concept has already been invented
Why is every single """AI""" believer always psychotically ill?

>>17015977
it's going to be hard few years for you
try not to kys
maybe use the time to learn how to read

>>17015979
>psychotic patient does generic chant
Ok and when you're done, please wipe the drool off your chin and tell me more about how every mathematical concept has already been invented. :^)

>>17015980
>implying implications
i am suggesting to you that this idea that humans will be required to come up with new concepts that the machines will then use to solve problems is the bigger cope - that you are just like tao
the machines will do it all - faster, better, and at scales we simply can't

>>17015981
If humans are not required to come up with mathematical concepts, why don't your handlers just remove all of the human mathematical concepts from the training set and see how long it takes for the static statistical model to magically rewrite itself and incorporate them? :^)

>>17015982
i'm just going to wait 18 months instead
good luck - take up something more fulfilling and difficult - like cooking.

Humans are fundamentally not creative enough to do math properly. Their brains are too smooth, not enough connections.

>>17015983
>i'm just going to wait two more weeks for the magical AI god that doesn't require human training data
What does this have to do with this thread about LLMs?

>>17014465
>AI did thing no one cared enough to do

File: pepe2.jpg (70 KB, 570x788)

70 KB JPG

>>17014465
Can we all agree that AI will likely be the one solving all big open problems in math from now on? Or are we going to mass cope about it?

I guess we will still have 1 year left, but I can't imagine it being later than GPT-7 that those things can one-shot RH no scope.

Also, can we agree that this is a good thing?

>>17015993
It can solve Millennium Prize Problems and people will still go "haha AI can't draw hands how many r's in strawberry going to walk to the carwash LOL"

>>17014486
You could cure cancer or solve the asteroid mining problem but you'd rather talk to it about jews? You stupid retard

How long before AI can solve real problems like efficiently shadowbanning LLM spambots and delusional LLM users?

>>17015852
>Not to mention AI is made of math, if you can make better math and engineering, the AI can produce even more advanced math, which you could then use to build a better AI, etc etc.

I think you meant the AI can make better math, which can be used to improve itself, which can be used to make better math, which can be used to improve itself....

The exponential runway is almost here. We started by letting it learn from the history of our species. Once we teach it to experiment with simulated experiments, which it will be able to do such iterative experiments a million fold over what we can do in a tenth of the time, it's over.

>>17015993
most open problems come from our systems of reasoning being incomplete, AI can't come up with a completely few form of reasoning. AI is a fossil of prior knowledge that works great for things no one would care about normally.

Consider this; why haven't LLMs found a break through way of training themselves more efficiently?

File: brainlet.jpg (49 KB, 922x781)

49 KB JPG

>I think you meant the AI can make better math, which can be used to improve itself, which can be used to make better math, which can be used to improve itself....
>The exponential runway is almost here.

>>17016016
I think you're coping. I don't see why AI would be fundamentally incapable of generating breakthroughs concerning their own development. We don't need completely new forms of reasoning to do math, our forms of reasoning are, if anything, quite repetitive in math. All things that AI is supposedly incapable of doing seems to be done in a few months with scalling and some small tricks, I really think there are no big barriers left.

>>17016023

It sounds like you've crystalized your mind into thinking only in terms of symbol manipulation. Math is as much philosophical as it is symbolic.

Would AI be able to come up with Godels incompleteness theorem?

the answer is no.

another way of thinking about it would be, can AI create a new conceptualization of God that people would agree with? The answer is no because AI has to rely on prior knowledge, it can glue patterns of though together but it can't completely come up with new ideas, well it can but the output will look random and nonsensical, as if it was made by a machine without thought(which it is)

File: brainlet12.png (245 KB, 783x742)

245 KB PNG

>I don't see why AI would be fundamentally incapable of generating breakthroughs concerning their own development.

File: brainlet24.jpg (41 KB, 500x500)

41 KB JPG

>another way of thinking about it would be, can AI create a new conceptualization of God that people would agree with?

Here's some additional cope for your folks. Reading this paper, it is clear that the hard part was done by Kilpatrick, Jaeger, and Tutte. Next, you'll note that the proofs are still relatively short. This means that they are weak to broad attacks on the problem since none have to go particularly deep. Being able to go from token search to term/idea search is still genuinely impressive and will prove a lot of theorems, but there is no exponential speed up, just a change in base. Finally, the AI is still not introducing new abstractions. While the AI can effectively use flows on graphs, could an LLM actually devise the idea itself? Once it can, then the door is wide open for math and pretty much anything else.

>>17016038
Erdos #1196

>>17014482
>and that communism is objectively good
What's the alternative? Some system that gives other countries unlimited access at influencing elections and creates an incentive structure around importing an ethnic voting bloc and paying them with little girls from the local population to rape in exchange for them keeping you in power?

>>17016028
anon you have to move from word matching to semantic search, at least you'll be closer to LLM "intellect" if you do.

>>17014486
Is engineered a way to make LLMs work locally with only 1gb of vram want the GitHub link to the kernel optimization methods?

Remeber when some /sci/ posters were posting problems it couldn't solve?

>>17016045
Definitely don't look up the Soviet policies importing foreigners into various Russian regions to create artificial ethnic tensions they can use to maintain state control, midwit useful idiot, you'll be the first in Klaus Schwab's techno-gulags.

>>17016008
If it won't tell you the truth about one thing why do you assume it will tell you the truth about other things? Either it tells you the truth all the time or it doesn't. I don't want my AI to be a liar who has to maintain constant cognitive dissonance and nonsense.

>>17016126
>use to maintain state control
A communist dictatorship with a modern military doesn't need to do that.
>importing foreigners... to create artificial ethnic tensions they can use to maintain state control
This is basically America right now.

>>17016131
Whatever you're low IQ, no point trying to explain anything to you taxcattle, just eat your slop and believe in will to power.

>>17016131
Communism causes societal stagnation and it's always corrupted by human greed. Everything gets corrupted by human greed, but at least capitalism allows you to try to fend for yourself

>>17016131
>need
That depends on how you define its objectives.

File: IMG_2900.png (513 KB, 1000x1000)

513 KB PNG

If it’s so genius why don’t they let it publish a paper in Nature?

>>17014482
>communism is objectively good
There is no AI that says that unless you tell it to.

File: brainlet21.png.png (210 KB, 1024x1024)

210 KB PNG

>>17016046
>anon you have to move from word matching to semantic search, at least you'll be closer to LLM "intellect" if you do.

>>17015844
I imagine that humans will still be needed to read what LLMs produce and verify that they're right. I imagine the publishers checked the proof before putting this out there. This will remove all joy from doing maths.

>>17015852
Knowing the answer to a maths problem and discovering the answer for yourself are not necessarily the same. The act of problem solving is a creative endeavour that can feel extremely satisfying. Having a solution handed to you feels cheap.

Human thought is now expensive and inefficient, or to be more specific, the aspects of human thought that are rewarding are now expensive and inefficient. I was told to develop a skillset, which (now) is no longer desirable. The "good" jobs have you prompt an AI, and you sit back and do nothing. You get those jobs by being born from the right vagina. The "bad" jobs have you shovelling shit. these go to everyone else. It makes no difference to the end result who does whichever job, as anyone does them.
If I were born today, I'd grow up in a world that doesn't value human thought. A world that doesn't value merit. Perhaps I always lived in such a world.

>>17014542
One hour on granny’s PC she uses for reading emails from her grandkids and one hour utilizing the full resources a ten billion dollar supercomputer ain’t the same fucking thing.

File: 1000023600.png (888 KB, 1079x1788)

888 KB PNG

Keeps happening

>>17015844
You're retarded if you think this is not mathematicians just proving results for OpenAI to bypass grant requests and long review time.

>>17016554
>I imagine that humans will still be needed to read what LLMs produce and verify that they're right.
Not really. AI can generate proofs in lean.

>>17017521
for convenience:
https://arxiv.org/abs/2607.12208
and another open problem solution that Dobriban himself flagged as notable but I didn't see much discussion about it:
https://arxiv.org/abs/2604.07055

File: snailcat.png (1.52 MB, 1254x1254)

1.52 MB PNG

>>17017537
are you implying that openai has assembled a secret team of mathematicians who solve decades old problems and attribute them to gpt?

>>17015982
>If humans can come up with mathematics why do we teach them at school?
Okay tard

>>17017544
Explain how your mentally ill "retort" refutes my point. Protip: you won't.

>>17017553
THe fact that you need training for AI to learn doesn't prove it cannot come up with novel concepts. Because actually humans also need training data to learn anything, that can be sensory inputs or it can be going to school to learn maths. After learning we can come up with novel ideas. Protip: stop being a faggot

>>17017565
>b-b-b-but humans need sensory inputs
Explain how your mentally ill "retort" refutes my point. Protip: you won't.

>>17017565
it isn't real ai unless it evolves from primordial sand and solves reimann

>>17017570
Notice how "AI"-believing subhumans can't actually form any thoughts besides regurgitations of ChatGPT 3 era marketing talking points.

File: file.png (759 KB, 1440x960)

759 KB PNG

>>17017571
>robot doctor curing your mother's cancer listening to you tell it how it actually just stole the cure from a secret human source

>>17017542
Are you implying companies haven't hired scientists to perform biased experiments for them before?

>>17017573
Notice how "AI"-believing subhumans can't actually form any thoughts besides regurgitations of ChatGPT 3 era marketing talking points.

>>17017568
he's not trying to refute your point, he's just chanting. what you need to understand is that these brown animals are completely blown away by things their own brains can't do, like adapting generic rhetoric to a slightly new context. a llm is basically a god to this indian

>>17017578
>these brown animals are completely blown away by things their own brains can't do, like adapting generic rhetoric to a slightly new context
Saaar, you don't understand, they didn't teach me at school, I just need better training data like all the human.

>>17016554
>If I were born today, I'd grow up in a world that doesn't value human thought. A world that doesn't value merit. Perhaps I always lived in such a world.
>Perhaps
Were you born before capitalism?

File: file.png (94 KB, 1176x430)

94 KB PNG

Keeps happening

https://medium.com/@kerger.p/an-ai-assisted-breakthrough-in-convex-optimization-an-optimization-problem-dating-back-30-years-a-db5c631119de

https://github.com/PhillipKerger/zero-order-bounds-lean-verification

>>17017797
>>in a single 148 minute session!
>>one prompt!
>look inside
>32 subagents
Don't get me wrong it's impressive, but are these all just propaganda pieces?

>>17017805
it's just what things will look like going forward
the models are being RL'ed on parallelising themselves - this gen of models is the first
both fable and 5.6 will do this on ultra

Luddites BTFO.

>>17017805
32 subagents means that the AI is running 32 different chains of thought simultaneously, not that there are 32 different AIs working on the problem at the same time. Just because your inferior meat brain is only able to hold one thought at any given moment, don't go imposing your disabilities on the AI.

File: brainlet15.jpg (22 KB, 460x455)

22 KB JPG

>32 subagents means that the AI is running 32 different chains of thought simultaneously, not that there are 32 different AIs working on the problem at the same time. Just because your inferior meat brain is only able to hold one thought at any given moment, don't go imposing your disabilities on the AI.
I'm not convinced anyone replying to these LLMs/subhumans unironically is actually human.

>>17015967
Physicists lack money. They always try to connect their research to trendy things to get funding.
It was machine learning, biology, quantum computing. Now it’s AI.
I’m not joking.

>>17017539
>didn't see much discussion about it:
Bro I haven't seen too much discussion about CDC even. I wonder if people are just coping or if the field is just so small. There was a lot of talk about the unit distance problem but what is happening now is way more important.

>>17017820
you're actually wrong here - a subagent is literally a new instance of the same model created and prompted by the orchestrator agent - it may or may not inherit the context from the main chat
depending on the harness you can do this recursively
subagents may get limited tooling or just have the full harness accessible to them - it can at times feel like 10 people using your computer at the same time.

>>17017869
You're talking to nonsentient brown insect. There's no meaningful difference between running 32 different chains of thought and 32 different sessions, which is utterly obvious if you have any idea how a LLM works.

>>17017887
there actually is, because tool calling doesn't happen inside cots afaik

>>17017893
I don't know if Current AI Thing calls tools inside its CoT but there's no reason why it couldn't. There is nothing special about the CoT. But that's besides the point. The brown insect's claim is retarded even on its own terms, which is obvious if you think about it for like 2 seconds.

>>17017910
you need to chill out

>>17017922
Brown insects and other "AI" fans should be banned on the spot.

>>17017805
the real objection is that "one prompt" was really 10 pages of extremely elaborate instructions, pointers, directions, hints and other handholding.
but to be fair, the prompt was probably AI-generated too

>>17018412
>vibe-solve a math problem
>take the entire transcript without the last response as a single "prompt" and claim the AI solved it in a single prompt
that's why they call him scam altman

>>17018412
To just give you an idea:

A hard-instance proof may use a special subclass of Fd, such as polyhedral functions,support functions, maxima of affine functions, gauges, or distance functions,
because a lower bound on a subclass is a valid lower bound for the full class. However,
every hard function produced must actually be convex, 1-Lipschitz on Bd, normalized
by f(0) = 0, and consistent with all exact answers supplied during the interaction.
Closing the gap may take any mathematically correct form. For example:
a lower bound matching the existing Oå(d2) upper bound;
a deterministic algorithm with exponent strictly below 2 together with a matching
lower bound;
matching upper and lower bounds at an intermediate exponent;
a sharper non-power-law characterization.
A strict improvement on only one side does not count as a complete resolution.

Results that do not count
The following are insufficient unless they are accompanied by arguments that imply
the complete matching result above.
An improved lower bound such as d5/4, d4/3, d3/2, or d2−o(1) with no matching
algorithmic upper bound.
An algorithm using o(d2) evaluations with no matching lower bound.
A randomized algorithm, including one that succeeds with high probability, in
expectation, or after fixing a favorable random seed.
A randomized lower bound that is not correctly converted into a worst-case lower
bound for every deterministic algorithm.

that section goes on and on for a page and a half and it's just a fraction of the prompt

>>17018420
Topological and decision-theoretic lower bounds
Explore continuous and discontinuous decision trees separately, adversary dimensions, Borsuk-Ulam-type mechanisms, invariance-of-domain ideas, widths of function
classes, topological complexity, and algebraic or measure-theoretic indistinguishability.
A lower bound for continuous decision rules does not apply to this task unless it is
extended to arbitrary decision rules. Clearly identify every regularity assumption.
Polyhedral and support-function hard instances
Investigate maxima of many affine forms, distance-to-hidden-set functions, support
functions of hidden bodies, nested ridges, multiscale facets, and adversarial arrangements. Exact values must not inadvertently identify all hidden facets or encode the
minimizer in one scalar.

Explore Fenchel conjugates, saddle formulations, support-function duality, bundle
models, cutting-plane dual certificates, and minimax exchanges. Determine whether
exact values provide indirect access to a dual object that can be optimized more
efficiently.
Do not assume access to the conjugate, a maximizing subgradient, or a dual
separation oracle unless it is constructed from counted value queries.

Study nested localization, potential functions, center-of-gravity or volumetric
progress, shallow cuts, ellipsoidal geometry, and extremal configurations of queried
points. Seek either an invariant forcing Ω(d) information cost per effective cut or an
algorithm that obtains global progress without explicitly reconstructing a cut.

Use low-dimensional linear programs, convex-extension feasibility programs, adversarial searches, and exact symbolic calculations to falsify proposed lemmas quickly.

>>17018412
>>17018420
>>17018423
As I said though, all this output, and the whole lean formalization github, seems to be the result of an automated workflow/pipeline that planned the project, scoured the lit, wrote synthesis notes for each sub-section of the task, and digested and collected the notes to finally produce that gigaprompt, with little or no human input (except for all the existing literature of course.)
Probably cost $50-$500 even outside of the proof generation itself depending on the workhorse model used.

>>17015967
String theory is useless. Like decades of research and billions of funding wasted for nothing. Also anon is right most physicists are poor as hell they quickly sell out once given the chance.

>>17015976
No, Gowers explicitly said he expects mathematics to be fully automated and no longer be a human job.

>>17018455
He’s not coping because he’s a very upper class faggot who doesn’t need to give a fuck about what that’ll mean for literally everyone born from a lower class vagina than he was

>>17016026
the answer is not yet.

>>17018451
Billions? I feel like there were 5 guys MAX working on string theory, who probably received a couple million in grants since the 80s, if that.

>>17017538
It is possible to create misleading proofs with lean, so we would still need someone trustworthy to verify the proof. I remember reading a philosophical argument, years ago (before LLMs were a topic of discussion), that if a proof cannot be understood by our top mathematicians, it may as well not exist. This is because the point of proof is to convince us that something is a fact. If a proof of the Riemann hypothesis could be developed by LLMs, but no human could understand it, then to accept it as a proof would be to accept the conjecture because "it seems right." If you can't understand a proof, you can only make a guess as to how true it "feels," which undermines the purpose of proving something to begin with. Idk it was expressed more eloquently than what I'm saying.

>>17017581
The "perhaps" was rhetorical.

>>17018458
>It is possible to create misleading proofs with lean, so we would still need someone trustworthy to verify the proof
then just create a trustworthy LLM. luddites are so short-sighted

I thought these things were supposed to be able to reason their way through all potential realistic combinations of materials to solve immediate and very dearly in need of solving physical world problems,
but instead it's only being used to "proof" some jackass shit like your virtual theorem that says a circle is a shape because you define the space as a circle or some other time wasting, useless slop.

300+ billion dollars, the internet ruined, and the global computing hardware market tossed into a shark tank for nothing.

>>17018516
AIs don't have access to the real world to manipulate objects and test them

>>17018496
You can't trust AI you urbanite moron, it's untrustable by it's nature, it's not alive, schizo cultist.

You can use it but you constantly need humans checking whether the conclusion is actually correct or not.

>>17018518
>"AIs" are useless
True.

>>17018518
it's what i remember the future of AI being described as useful for, from almost 20 years ago
"when it can master language it can supposedly master the subject, and then pump out lists of the most likely combinations and their different properties and matches to criteria." - to paraphrase an ancient memory of a 3 page technical report.
all that talk of metamaterials, indestructible hybrid ceramics, silicon photonics, >10Ghz computing barrier, generational leaps in battery tech, etc. etc. etc, shit we "were gonna have by 2020". but it's the same as acceptable "lifelike quality" real time ray tracing on contemporary single accelerator cards. a myth. not going to happen any time soon.
Hell, at one point the early advocates for GPGPU were talking about NN development and expected future applications.
Cryptocoin's first release was oddly timed just right with the release of what you can consider the generic equivalent of 1st gen modern GPU design. Basically nothing has changed, just the amount of work units + custom math accelerators.

it's been 20 years and a half a million dollar machine can't AI it's way through a stack of element compositions with any prediction accuracy.
give me a break

>>17018524
>>17018526
we're starting to see the first physical harnesses in medicine now
the generalised robot humanoid capable of acting as a decent harness is a couple of years away
these sorts of naysaying posts are just embarrassing at this point - the tech has earned the benfit of the doubt

>>17018529
A program that cant do complicated work without being held by the hand through every step is not intelligence.
It's just punch cards all over again but now more complicated with no guarantee.

GPT 5.6 pro one-shots IMO 2026. Probably the last time they participate.

https://github.com/SignalPilot-Labs/AutoFyn/blob/production/results/imo-2026/pdfs/IMO_performance_by_GPT_5_6_sol.pdf

File: 20051031.gif (47 KB, 468x598)

47 KB GIF

>>17018529
oh, before i forget
>medical
you would
yeah, course you would

>>17018529
>we're starting to see the first physical harnesses in medicine now
And? What does this have to do with your trite brown insect fantasies?

>>17017568
>AI solves Riemann
Umm Saar but it cannot browse 4chan while being a faggot like me so it is not actually that smart. It is a stochastic parrot

>>17018577
Completely incoherent reply. You sound literally unhinged.

>>17015982
>If humans are not required to come up with mathematical concepts, why don't your handlers just remove all of the human mathematical concepts from the training set and see how long it takes for the static statistical model to magically rewrite itself and incorporate them?
all the math you need to bootstrap a math-capable AI has already been invented, luddite

>>17018529
>this 3T param model that runs at 30 tokens/second eating 5 kilowatts is going to be able to fire off MCP tool calls fast enough to do realtime movement

>>17018719
5.6 will run at 500+ tok/s on cerebras and it's the full weights this time
and you don't need the orchestrator model to control the robot, it can basically be a dumb subagent that just performs the physical work to spec and the results just get piped back to gpt 10 or w/e

>>17018458
>Billions? I feel like there were 5 guys MAX working on string theory, who probably received a couple million in grants since the 80s, if that.
no there was a really big community for string theory they once were huge everyone who wanted to be popular did it and it was all the same circle jerk bullshit.

>>17018731
>you don't need the orchestrator model to control the robot, it can basically be a dumb subagent
Ah yes this tiny subagent that needs to handle complex 3D movement in the real world (and probably vision too because otherwise you're going to be extracting position data and sending it with the slow model) is going to run it better than the giant model. Why do we need giant models again? Oh right because small ones are fucking retarded.

>>17018739
do you think there's 5t param model running something like a figure robot lol

>>17018747
It's not an LLM running commands through a harness that's for damn sure, you big retard.

>>17015844
You should have been born 10 years earlier, like me. Tough luck, zoomer.

>>17018752
it is a vlm lol
give it a rest, you just don't know what you're talking about

>itt: failed math phds feeling threatened by the big bad clanker
love to see it

File: seething_calmly.jpg (39 KB, 460x663)

39 KB JPG

>>itt: failed math phds feeling threatened by the big bad clanker
>love to see it

>>17018755
>this thing that doesn't use a harness is totally the same as some giant language model that uses a harness and it will be able to control it the same way!
Great job.

>>17018819
you realise a lm without a harness can't actually perform any actions right? it just outputs text. anything that converts that text into action - like actuating a motor - is a harness.

>>17018827
The VLA model (not vlm by the way) include the actions in the outputs. It doesn't take the text and use some tool call in a harness like a big braindead retard would make because that would take forever for any actions to happen.

>>17018838
you're describing a harness you fucking imbecile

>>17016045
everything you described happens under communism. but don't let go of that dream!
https://m.youtube.com/watch?v=RBYsQE4Pxb4&ra=m

>>17018861
>you're describing a harness you fucking imbecile
It's not you fucking moron, the motor outputs are generated directly by the model, it's not being fed into a fucking tool call.

>>17019024
Still need a harness sending the outputs to the motors
We're squabbling over a definition at this point

>>17018518
They do. I have given them access to my webcam and I will act as their limbs and fingers following their instructions.

>>17019162
>We're squabbling over a definition at this point
Maybe that's because you're both retarded and there's zero substance in your argument about a technology that doesn't actually exist.

>>17017537
yeah man, humans did something hard and DIDN'T ask for credit.

>>17015993
Its not a good thing because it means we are approaching it being smarter than us and then Skynet almost certainly follows.

>>17019212
>we are approaching it being smarter than us
GPT-3 was already smarter than whatever "us" you might be a part of. Other than that, you're making a nonsensical comparison. Is a calculator also "smarter than us"?

>>17019225
No. A calculator narrowly focused in its capabilities. AI isn't.

>>17019226
There is no such thing as "AI". LLMs aren't "AI" and they get beaten by moderately competent humans at anything out-of-distribution in any domain.

>>17019212
>Its not a good thing because it means we are approaching it being smarter than us
This is still a ways away
>then Skynet almost certainly follows.
No it doesn't

>>17019227
>>17019229
Reddit might be a better site for you guys.

>>17019227
>they get beaten by moderately competent humans at anything out-of-distribution in any domain.
This. It takes a ML background to know what "narrowly focused capabilities" would mean in this context (it's obvious that a LLM is not a conventional one trick pony), but the lack of generality is corollary to the basic premise of modern ML. It's called a "model" for a reason: it's modeling a specific data distribution (in this case, a text-based representation current human knowledge. The model's ability to fill in the gaps in the data is focused on the dense regions of the learned distribution, by definition. The simulated "reasoning" helps it do better in those regions but predictably falters at the edges of the distribution, if only for the reason that it's never forced to learn to fully decouple reasoning from the specific contexts where it appears in the data. Why would it? It performs better on test set (which comes from the same distribution) that way and also on the benchmarks.

>>17019238
>technically informed skepticism of the current thing??? that's heckin' reddit, luddite
Fuck off, jeetoid.

Isn't modern science just explaining whatever in terms of the officially stablished truths? And "proving" stuff by using only acceptable methods? Then don't be surprised that LLMs can produce "science", mimicking already existing data patterns is what they are good at.

I wonder if departments will from now on AI-write all their papers. Nobody is deep-reading them anyways, let alone replicating the experiments. The industry of producing papers can keep on going even with the increasingly lazy zoomer professors to come.

>>17016026
It seems very narrow-minded to just assume they can't do better. people constantly move the goalposts and always say "yea but thats not true creativity"

> yea it can calculate prime numbers but thats just stupid calculation, no real creativity
> yea it invents new chess moves but thats so narrow its not creative
> yea it can do primary school maths but thats not difficult
> yea it aces SATs but thats just learned patterns
> yea it wins simple maths contests but not the real ones
> yea it wins math olympiads, but those are problems humans have already solved, they cant do real proofs
> yea it disproved some 50 year old problem, noone solved but those are just special cases. it for sure cant prove the riemann conjecture

Do you seriously not see the trend? ill be convinced it doesnt get better if i see it actually stalling

>>17015993
Constructing mathematical objects seems to be what ai is good at, like constructing a counterexample or a graph. Unfortunately the RH cant be solved by constructing anything. Time will tell if AI gets any better at non constructive proofs.

It may seem like shifting the goalposts, but constructive proofs genuinely are different than non constructive ones. It's possible the ai is following some hidden algorithm.

>>17019760
>people constantly move the goalposts
No, they don't. It's your handlers setting arbitrary goalposts corresponding to what their toys can do on a given week, then pretend those things have always been the pinnacle of creativity/intelligence for everyone else.

>>17019768
the point is, given the trend, they get more and more creative all the time, and its reasonable to assume that will continue to do so, until we are at a point where they are probably on our level or beyond.

>>17019760
>> yea it can calculate prime numbers but thats just stupid calculation, no real creativity
It better be able to, because it was trained on human knowledge about prime numbers.
> yea it invents new chess moves but thats so narrow its not creative
It better be able to, because was trained on basically all of the human knowledge on chess.
Same thing for the rest of your points. The problem with "solving" these problems is that the body of published math work is so vast it can be practically impossible to tell if this is genuinely new, or if it's just pulling this from its training data containing an obscure paper published 30 years ago which solved the problem in some unrelated field but it was never put together that it could be applied to Erdos problem #10004561. Then we get faggots like you claiming it's some emergent intelligence when it's just regurgitating the training data.

>>17019775
>It better be able to, because was trained on basically all of the human knowledge on chess.

True, but then it played against itself and got better than any human, which seems impossible from just learning everything from humans. so it does obviously go beyond what it learned in some domains. so what happens when we apply that to other domains? what makes you think that this growing beyond what it trained on will not happen in maths or coding?

>>17019774
>they get more and more creative
No, they don't.

>>17019779
>then it played against itself and got better than any human
No, it didn't. LLMs can't play chess worth a damn.

>b-b-b-but what about this specialized ML model that does nothing besides playing chess...

>>17019780
> No, they don't.
Yes, they do.

>>17019783
>Yes, they do.
What evidence of this do you have?

>inb4 arbitrary list of what your handlers think is impressive

Creativity is the degree to which a cognitive architecture can usefully alter and expand the data distribution representing a creative domain. LLMs have zero creativity by definition.

>>17019784
it recognizes patterns and then create new stuff according to that pattern. e.g. it generates a perfect human face that does not exist. completely new face. in many cases humans dont do it much differently, they combine what they know to create new "novel" things. or take the chess example, it learned new moves that it haden't seen in in the training data.

i think humans also just use their patterns and intuitions to form their ideas, they are probably equally random, except their intuitions and patterns are unique, which is not the case for ais, which tend to the mean and hence always produce similar output towards the mean unless you specifically promt them with unique ideas.

>>17019789
>it recognizes patterns and then create new stuff according to that pattern. e.g. it generates a perfect human face that does not exist. completely new face.
None of this requires creativity. See >>17019788

>b-b-but subhumans like my are also uncreative
Ok.

>>17019789
>it generates a perfect human face that does not exist.
>it learned new moves that it haden't seen in in the training data.
There is no such 'it' by the way. You're suffering from hallucinations.

File: save.gif (2 KB, 60x70)

2 KB GIF

>>17015844
Hilbert's program ended in failure.

>>17016593
you can build supercompufers in a systematic and predictable way. you cannot do the same with human being and force them ti be geniuses at maths

>>17019788

LLMs are Turing complete and universal approximation theorem applies to them. A well-trained LLM of sufficient size can emulate human thinking with arbitrary accuracy. Unless you are saying that human thinking cannot in principle be simulated by a computer (a very dubious claim), LLMs will be able to simulate it well enough for all practical purposes. Including human creativity.

>>17019808
>LLMs are Turing complete
Wrong.

>universal approximation theorem applies to them
Irrelevant.

> A well-trained LLM of sufficient size can emulate human thinking with arbitrary accuracy.
Take your anti-psychotic meds.

>Unless you are saying that human thinking cannot in principle be simulated by a computer (a very dubious claim), LLMs will be able to simulate it
Psychotic nonsequitur.

Meanwhile my point stands completely undisputed:

Creativity is the degree to which a cognitive architecture can usefully alter and expand the data distribution representing a creative domain. LLMs have zero creativity by definition.

So, when exactly are we going to kill ourselves?

>>17019812
>LLMs are Turing complete and universal approximation theorem applies to them.
A transformer alone can't be Turing-complete because it has a finite context window and its internal state reflects whatever's inside that. It can only transition between a finite number of states. You can get around this by connecting the LLM to external storage and asking it to act like a processor, which gives you a Turing-complete LLM but accomplishes fuck all - computers already exist. A LLM's principal advantage over regular computing is in the way it mimics semantic comprehension, but that's tied to the finite context window by definition, so its internal "thinking" process is not Turing-complete.

The Universal Approximation Theorem gives no guarantees on what you can approximate in practice. You can approximate any continuous function to any degree of accuracy using piecewise linear functions or any number of other things. Going by the UAT, the perfect LLM could be a simple FFN with a single hidden layer, but that obviously doesn't work. If anything, the "continuous function" condition undermines you: human cognition is not a continuous function. It's not even a function. A function is a static and immutable mathematical object, which a biological brain isn't.

>Unless you are saying that human thinking cannot in principle be simulated by a computer (a very dubious claim), LLMs will be able to simulate it well enough for all practical purposes.
Biology is not Turing-computable. The idea that the mind is independent from its biological substrate is an opinion with nothing real to substantiate it. But even if minds are computable, LLMs obviously aren't powerful enough to simulate them, since they lack features as basic as being able to integrate new knowledge or to think in a nonlinear fashion.

>>17019788
>>17019808
just add noise to the system and you have creativity

>>17019914
That's not a coherent response to >>17019788. Try again.

>>17019808
I demand that all LLMs speak to me only in palindromic cipher; they fail immediately every time

File: 1730389219373177.png (187 KB, 564x664)

187 KB PNG

>>17019936

>>17019788
>usefully alter and expand the data distribution representing a creative domain
What does this even mean and why do you think AI can't do this?

File: 1771407280316632.jpg (120 KB, 770x600)

120 KB JPG

>>17019779
>True, but then it played against itself and got better than any human
It did not.

>what makes you think that this growing beyond what it trained on will not happen in maths or coding?
As far as I understand, training an LLM is essentially just drawing a thousand dimensional polytope around the training data in information space, then using gradient descent to generate new points within that polytope. The LLM can't find an information point which lies outside of it because it's convex. The convexity is the barrier, but it's also a necessary condition for gradient descent to work in the first place.

The infinite nature of information means that pretty much no matter what you will always encounter a data point which lies outside of this polytope, and

>>17020013
>, and
damn, they got to him

>>17020024
Yes, the deep state got to me.

and it's not exactly clear if information in real life has "dimensions" or if this is just a model of reality we use out of convenience. This I am not entirely sure of. If information doesn't have dimension it would just exist in 1 dimensional space, and we're just approximating its nonlinear nature with the extra dimensions.

>>17019984
>What does this even mean and why do you think AI can't do this?
Roughly speaking, if you have a bunch of training examples representing a coherent whole, like a body of knowledge or some artform, there should be an equivalent representation where the examples become data points spread out in a decipherable pattern in some high-dimensional space. The point of generative AI is to infer such a representation and the corresponding pattern. What you get is a model of the data distribution which you can use to generate new data points, with the obvious caveat that you have to respect this pre-learned distribution, whose overall shape (if that's still a good word in thousands of dimensions) is defined by the original training examples.

Now suppose you had a collection of texts from different eras and you arranged them into sets such that the i-th corpus contains all texts written up to and including the i-th century. If you modeled each corpus separately (e.g. trained a bunch of LLMs) and had some way to visualize and compare the learned data distributions, you'd see the shape changing over centuries. That's human creativity. A LLM can't do this on its own, by definition. You can prompt and sample it forever, use it to build a new corpus of texts, maybe even one containing solutions to unsolved problems, "original" novels and theories etc., but if you trained a new LLM on it, at best, you'd get back the same distribution you started with.

>>17020013
>>17020029
Or rather that they're drawing what is a thousand dimensional (or however many parameters we're up to billions) polytope by performing gradient descent on the training data. The polytope is stored in a matrix and output is obtained through matrix multiplication. This is slightly different than what I originally said, but the rest still stands.

>>17020013
>>17020075
>>17020029
Nice pseudbabble, shill.

>>17014465
I hope the successes felt good and it's proud of itself

File: 969690492140.png (75 KB, 932x628)

75 KB PNG

fable/mythos finds a counter example to the jacobian conjecture during the world cup final

>>17020396
it must have found this counter-example from some old paper. congrats to the human who instructed it to dig at the right place.

>>17020399
Alpoge still hasn't published his instance's CoT so that is possible, but considering the results we've seen coming out of OpenAI recently, this isn't beyond realistic.

"Not enough people are emotionally prepared for if it's not a bubble."

>>17020073
this doesn't matter, luddite. all the math we'll ever need has already been invented

>>17014482
None of the things you said is true since Trump won again. All AI companies instantly flipped the switch. You're also very stupid.

File: greyest-retard-itt.jpg (35 KB, 651x807)

35 KB JPG

>None of the things you said is true since Trump won again. All AI companies instantly flipped the switch. You're also very stupid.
Reminder that 4chan is getting massive amounts of LLM spam generated by the US State Department.

>>17016554
You may as well just be mad at all the results proven by other people before you.

>>17015844
>Creation and innovation are dead and all that is left is consuming. fuck this shit.
Be honest, would you have cared about any of these problems if they weren't famously unsolved?

>>17020416
You are one.

That's the problem, ones like you.

Perfect camo.

File: vitushkin1999 (1)-1.png (3.44 MB, 1252x6750)

3.44 MB PNG

>>17020399
>>17020407
If nothing else, it's similar enough to Vitushkin's map from 1999 that there's no way it wasn't a starting point.

>>17019878

>A transformer alone can't be Turing-complete because it has a finite context window

You don't need infinite context window to approximate human thinking. Human short-term memory is also finite.

>human cognition is not a continuous function. It's not even a function.

Everything can be expressed as a function.

>Biology is not Turing-computable.

Wrong. Physics is Turing-computable. Biology is just applied chemistry, which is just applied physics.

>But even if minds are computable, LLMs obviously aren't powerful enough to simulate them, since they lack features as basic as being able to integrate new knowledge

So a human with forward-amnesia is no longer sentient?

>or to think in a nonlinear fashion

LLMs are explicitly non-linear functions.

>>17014507
https://huggingface.co/wnfldchen/gemma-4-12B-it-qat-w4a16-ct-heretic
https://huggingface.co/wnfldchen/gemma-4-12B-it-qat-q4_0-gguf-heretic

>>17020667
>Physics is Turing-computable
NTA but weather is just one step away from physics and the step is an Oracle. Biology is even less tractable.

>>17019812
Your output has less variation than Llama on 0 temp

>>17020667
>You don't need infinite context window to approximate human thinking
You're pulling a new claim out of your ass now but your original claim stands refuted: a LLM is not Turing-complete in any useful sense.

>Everything can be expressed as a function.
This is trivially wrong: a mutable object cannot be a function by definition.

>Physics is Turing-computable
Again, trivially wrong: time and space are continuous in the Standard Model so no algorithm for a faithful simulation will ever terminate. Termination is a requirement for computability (otherwise you could run an empty infinite loop to exactly the same effect and say you're computing the output of the function).

>So a human with forward-amnesia is no longer sentient?
That's not even a coherent response. I wasn't arguing anything about "sentience".

>LLMs are explicitly non-linear functions.
That has nothing to do with anything. LLM are causal transformers that generate the text token by token based on what's already in the context window. That's a linear process. A LLM has no idea what it's actually getting at, because that would only start to emerge some unspecified number of tokens into the future, influenced by a sampler and a RNG it has no control over. It improvises step by step as it goes along.

>>17020777
I set the energy expenditure setting on my brain to minimal when talking with dumb subhumans. Anything more just isn't worth it. For example, look at the mindless, 70 IQ response I got for >>17020073.

>>17020424
Not really, you could build upon them. Now you can't

>>17020426
So? Does it matter?

>>17014465
This is a big day for math because barely anyone except enthusiasts (0.001% of population) cared about these and they were always treated as hobby projects attracting zero grants, but now we can just use this machine that cripples an economy of a small country in a press of a button to crack them at insane speed for marketing and PR value! Talk about goal posts, lmao. These niggers are scrapping the barrel for their IPO, i will care and i will cheer these advancements once an open weight model does this.

>>17014465
imo chinks gonna do the same soon but entire process will be transparent as glass, and they'll grant a few universities an unlimited inference or even just gift them a bunch of mainframes complete with their open source model

>>17020737
>12B
dogshit

https://huggingface.co/HauhauCS/Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced

>>17022314
The chinese literally just lost this AI math battle. But you guys are in denial.

>>17015967
was doing something similar, I asked for symbolic derivations and sonnet went ahead and verified the derivations numerically.

>>17022308
you're underestimating the implications this has for academia. just to give you an example, there were people who worked for years on the jacobian conjecture. grant funding is likely going to be much more difficult to get because why pay for people to research topics that AI might be able to solve in an hour? I'm not even going to theorize how large the implications are going to be, but consider some of the downstream effects on research
>publications, which already have their own problems, are going to be inundated with AI written publications, not just the text but the mathematical derivations, figures, and even ideas
>the value of publications, which already have a reproducibility problem, will go down even further
>publications require non-trivial contributions, but what does that mean if everyone is using AI to derive new results, how do you decide if it's novel enough?
>there's going to be a comparison to AI to see if non-AI research was "worth it". for example, if a problem similar to the jacobian conjecture takes a group of 5 researchers 5 years, you could calculate a dollar amount spent (by the funding body) and compare it to the dollar amount it took AI for the jacobian conjecture. there's going to be a constant comparison now, before AI it was difficult to put a dollar amount on research.
for those in academia, how are people reacting? are they just churning out publications as usual and pretending nothing is happening?

>>17020667
>Everything can be expressed as a function.
Can you prove that?

Is this the start of the singularity?

>>17020410
How do you know that?

>>17025615
>>17025615
This retarded jeetoid can't prove anything. It can hallucinate and post free-tier LLM slop, but that's about it.

FWIW, any static mapping from inputs to outputs is technically a function. If you had a magical and unchanging oracle that knows everything and can solve any problem, it would be a function. But that, of course, has nothing to do with reality. If you took a biological brain and somehow crippled it so that it never changes, maybe that would be a function, but it would be a sad husk of biological intelligence, with all the downsides and none of the advantages.

Now, if you wanted to use a function to model a real brain (rather than a static one), the function couldn't be a representation of the brain. It would have to represent a process whereby static brain snapshots evolve. I.e. a function that takes a brain snapshot plus inputs for the brain, and returns the outputs plus a new brain snapshot. I.e. a function of functions. Except a real brain changes stochastically. So you need a function that accepts a distribution over functions and returns a distribution over functions.

Even ignoring how obtuse and impractical a model that would be, it's clear that you are not dealing with a smooth function (i.e. the Universal Approximation Theorem doesn't apply) and it's nowhere near convex (gradient descent is very unlikely to be able to "learn" it).

>>17025629
>it's nowhere near convex
I mean the optimization problem is nowhere near convex.

>>17025306
https://huggingface.co/wnfldchen/gemma-4-31B-it-qat-q4_0-unquantized-heretic

>>17025629

>f you took a biological brain and somehow crippled it so that it never changes, maybe that would be a function, but it would be a sad husk of biological intelligence, with all the downsides and none of the advantages.

Such brain would work just like a normal brain, except it would be incapable of forming any new long-term memories. Forming new short-term memories (which live in electrical activations, not synaptic weight changes) would still work fine. We already know how people with such condition behave: it's called anterograde amnesia. They can still function pretty well, but they have to keep a journal of any new things. But things learned before work fine. Just like a LLM.

Such an AI would still be incredibly useful in practice.

>>17025878
Nobody asked for your hallucinated opinions, animal.

>>17025890

Psychotic luddite.

>>17025629
thanks, anon. rare informative post

File: 1358074883897.jpg (65 KB, 398x400)

65 KB JPG

>>17025629
>>17025900

>samefagging this hard

>>17025902
meds

>>17025878
>Such brain would work just like a normal brain, except it would be incapable of forming any new long-term memories. Forming new short-term memories (which live in electrical activations, not synaptic weight changes) would still work fine. We already know how people with such condition behave: it's called anterograde amnesia.
Are you retarded? That's completely wrong. People with anterograde amnesia don't have static brains. Jfc...

>>17025909

They cannot form new long-term memories, but old memories remain. Like a LLM. Static synaptic connections.

>>17025910
People with anterograde amnesia don't have "static synaptic connections". There is no such thing in humans. If there was, they couldn't learn new things but they actually can.

File: 1355737187777.png (22 KB, 757x627)

22 KB PNG

>>17025915

>anterograde amnesia
>couldn't learn new things

Umm.. yeah?

>>17025915
They can't, yeah.

>Wearing developed a profound case of total amnesia as a result of his illness. Because of damage to the hippocampus (an area required to transfer memories from short-term to long-term memory), he is completely unable to form lasting new memories. His memory for events lasts between seven and thirty seconds.[4] He spends every day 'waking up' every 20 seconds or so, 'restarting' his consciousness once the timespan of his short-term memory has elapsed. During this time, he repeatedly questions why he has not seen a doctor, as he constantly believes that he has only recently awoken from a comatose state. If he is engaged in conversation, he is able to provide answers to questions, but he cannot stay in the flow of conversation for longer than a few sentences and is angered if he is asked about his current situation.

>In a diary provided by his carers, Wearing was encouraged to record his thoughts. Earlier entries were usually partially crossed out, since he forgot having made an entry within minutes and dismissed the writings. He did not know how the entries were made or by whom, although he did recognise his own handwriting.[7] Wishing to record 'waking up for the first time', he still wrote diary entries in 2007, more than 20 years after he started them.

>>17025916
>>17025917
People with anterograde amnesia can learn new things. Regardless, there is no such thing as a brain with "static synaptic connections" in biology. Do you get all your "facts" from a chatbot or what?

>>17025920
There's no point reasoning with this psychotic monkey. It has static synaptic connections.

>>17025923
Lots of normies don't understand that you can learn implicitly even if you can't form new explicit memories.

>>17016554
"This will remove all joy from doing maths" what fucking joy holy shit this is only enjoyable to a very small section of society. If anything, this is a godsend for just about everyone but mathematician

>>17014603
this is harder than that

>>17014465
>50-year-old Cycle Double Cover Conjecture
congrats
tell me something I didn't already know, worded differently.
snore

>>17014482
holy cope

>>17025510
>for those in academia, how are people reacting? are they just churning out publications as usual and pretending nothing is happening?
still waiting for a response to this, I'm already seeing posts about how AI/LLMs are being used everywhere in the research process
>grant writing -> AI
>research papers -> AI
>peer reviewing research papers -> AI

>>17027855
Quantify how much harder it is and what's the threshold at which one must be impressed.

>>17029166
they're churning out publications about what's happening

File: 1781124737932866.png (399 KB, 820x513)

399 KB PNG

Oops, there goes another one.

File: pepe86.gif (576 KB, 300x250)

576 KB GIF

>>17029225
Another OpenAI W

>>17029225
>>17029227
Guys are we getting closer to the technological singularity?

>>17014486
heretic doesnt solve the bias of the training input data itself. reddit trained outputs reddit trained commie ideology.

it needs a truth/grounding layer to distil/extract what is true or not.

>>17029166
Someday scientific research will entirely consist of chatbots reading papers other chatbots wrote while scientists do manual labor all day, god willing

>>17014482

Reality has a liberal bias, chuddite. ;) Deal with it.

>>17029228
you're already in it

pure mathematics is so fucking gay holy shit
kids who entered a random number into a calculator and divided it by 7, trying to find one that divides down to 1, and tried to make it a job

>>17016126
It was exactly the opposite you imbecile. They exported the Muscovite sub-nigger to destroy the finnic, slavic, caucasian, turkic, and asian races.

>>17014465
I'm glad I jumped ship from academia to industry. At least I'll make some cash buffer to live off of when I'm finally fired and any job in my field is taken by LLMs. But yeah it is over for almost all white collar jobs.
If you ever plan to kys don't waste it wink-wink.

>>17025510
There is 100x more paper mill-quality papers submitted everywhere. Previously serious unis and labs let their students submit fully ai-generated slop articles that don't even pass the smell test. Several journals accepted these slop articles, revealing they were shit journals (which everyone kinda already knew though, with only a few surprises). Reviewers nearly exclusively use gayI to """review""" which means that there is no peer review (by definition) anymore and everything is worse.

Meanwhile all funding to non-AI actual work has vanished.

But peer review was already deeply broken before AI anyway so whatever.

>>17020013
Gradient projection is a totally different algorithm than gradient descent (though similar at a high level). You are describing the former.
There is no convexity requirement in practice, though as you say it is required in theory. The theory of DL is so far off that it's always been a bit of an in-joke in the field. It can't describe training behavior, it can't describe convergence results, it can't describe optimality at convergence, all the bounds are broken in practice. Anyway, the training algorithms used aren't gradient descent but stochastic gradient descent (more specifically, descendants of this techniques which can be interpreted as computing noisy estimates of 2nd order derivatives from locals along trajectories). The stochasticity allows escaping local minima on concave surfaces, thus allowing pretty good learning even in that space.
The space is not thousand-dimensional, it's billion to trillion-dimensional.
In fact, the bottleneck is data efficiency: LLMs are computationally efficient but significantly less data-efficient than auto-regressive models like LSTMs. As a result, even with all the data on the internet and in every book ever written, it's about stalled right now. Of course, a better arch that isn't a literal blind bruteforce machine would help quite a bit.
But the reason it's a blind bruteforce machine is because that enables leveraging large noisy data efficiently, in a way that is hard to conceive if you need better structured or cleaner data. You will always learn better with 1m noisy blind datapoints than 100 noise-free structured datapoints.