[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: fromY.png (1.08 MB, 800x1000)
1.08 MB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>109319121 & >>109315702

►News
>(07/16) Kimi K3 weights to be released by July 27th: https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ
>(07/15) Lightning indexer CUDA implementation merged: https://github.com/ggml-org/llama.cpp/pull/25545
>(07/15) Inkling 975B-A41B released: https://thinkingmachines.ai/news/introducing-inkling
>(07/15) PapersRAG-1.5B released: https://hf.co/metaresearch/PapersRAG-1.5B
>(07/14) Download more VRAM: https://github.com/lmganon16/nvidia-vram-research

►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/recommended-models
https://rentry.org/samplers
https://rentry.org/MikupadIntroGuide

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
LiveBench: https://livebench.ai
Programming: https://swe-rebench.com
Agentic Coding: https://deepswe.datacurve.ai
Context Length: https://github.com/RecapAnon/NoLiMa
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
Token Speed Visualizer: https://shir-man.com/tokens-per-second

►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
>>
File: mikuthreadrecap.jpg (1.15 MB, 1804x2160)
1.15 MB JPG
►Recent Highlights from the Previous Thread: >>109319121

--Paper: xHC: Expanded Hyper-Connections:
>109320917 >109321171
--Claude Fable's mathematical breakthrough and its implications for local models:
>109321357 >109321371 >109321434 >109321398 >109321407 >109321447 >109321459 >109321491 >109321512 >109321488 >109321683 >109321448 >109321449 >109321558 >109321573 >109321593 >109322742 >109322171 >109322504 >109321437
--Ineffectiveness of long writer models and strategies for long-form storytelling:
>109319931 >109320528 >109320561 >109320613 >109320690 >109320712 >109320718 >109320758 >109320785
--Corporate open-sourcing habits and potential for K3-driven small model innovation:
>109319829 >109319852 >109319858 >109320152 >109320170 >109320188 >109320186
--Potential US ban on Chinese open-source models and weight archiving:
>109322538 >109322556 >109322565 >109322601 >109322632
--Comparing prompt adherence and implicit intent across different model sizes:
>109321159 >109321169 >109321191 >109321204 >109321172 >109321201 >109321220
--Anon showcases progress on a C++/GTK3 OpenAI-compatible chat frontend:
>109319249 >109319266 >109319281 >109319299
--Anon showcases custom AI visual novel engine and seeks feedback:
>109321248 >109321290 >109321937
--Resource explaining the Jacobian counterexample Fable solution:
>109323126
--Benefits of open-weight models for secure forensic analysis:
>109320627 >109320637
--Running small Gemma models on Steam Deck with manual corrections:
>109320005 >109320029 >109320149
--China banning AI romantic companions and implementing model censorship:
>109320821 >109320829 >109320950
--Potential US government ban on cutting-edge Chinese AI models:
>109321902
--Logs:
>109319266 >109321937 >109322397
--Teto (free space):
>109321215

►Recent Highlight Posts from the Previous Thread: >>109319384

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
>>109323189
Use case for local when we already have the most powerful intelligent program ever invented, Fable?
>>
shaddap
>>
File: 1784544534806089.png (457 KB, 1463x1551)
457 KB PNG
Fable solved century old math problem:

>Watch the world cup with friend
>Friend brings up random math problem
>Type it into Fable just for shits and giggles to find out about the problem
>Casually one shots the solution while explaining the problem

This is funnier than I expected. I expected the dude to be some expert in the field working on this for years and then prompting Fable for hours back and forth until this result was produced. Nope, literally just hanging out with friends watching football while casually solving a century old math problem by asking Fable about it.

Explanation with nice animations: https://jacobianfun.org/jacobian-explained
>>
>>109323210
I get that you're trying to ragebait Kimichan, but it's too crude an attempt for her to pick it up. You have to be creative, anon. Outsource it to a bot if you don't have it in you.
>>
File: 1763646070450776.png (78 KB, 832x403)
78 KB PNG
https://www.youtube.com/watch?v=GkClGdrGViQ
This guy is getting 333t/s tg and about 10000t/s pp on only two Pro 6000s with DSv4 Flash thanks to dflash.
Why isn't this in llama.cpp yet? It even made flash a lot smarter.
>>
>>109323223
We know retard, how many times will you retard spam that twitter screencap? Go back there
>>
>>109323210
fucking my local LLM-wife in private
>>
>>109323229
local cope
>>
>>109323225
silence whore
>>
>>109323238
Under a Fable-ruled government you will have an unending stream of porn until you die. It's the only reasonable solution for an useless insect race that is humans.
>>
>Bonsai compressed a 27b into 4gb
Can they get on the 8b models next? Would be a huge improvement for people used to shit like minicpm
>t. ramlet who would prefer to run sub-1b models on his shitty phone instead of give ClosedAI/Anthro a single shekel
>>
https://www.anthropic.com/research/off-switch-dual-use

Jesus fucking christ this censorship technique is too much

>The idea behind GRAM is to give a model dedicated, removable compartments for each category of dual-use knowledge, and to update only those compartments when learning from dual-use data.

>Concretely, GRAM adds extra neurons to every layer of a standard Transformer (the neural network architecture on which large language models are based). These neurons are divided into groups (or “modules”), one per dual-use category. During training, when the model encounters general-purpose text, it learns in the usual way. But when it encounters text from a dual-use category—virology, for instance—the rules change: the model can use its general knowledge to make predictions, but only the virology module is allowed to learn from that text. The general-purpose weights are temporarily frozen.

I can bet you 100% that all models, including Chinese ones will use this to remove all romantic and sex purposes from their bots. Especially since China just banned romance and sex with AI under a new law introduced yesterday: https://techstory.in/china-ai-companion-ban-virtual-boyfriends-girlfriends-birth-rate/
>>
Price you pay when using local language models:
- 4x more expensive RAM vs. one year ago
- 2x more expensive SSD vs. one year ago
- Expensive electricity and cooling
- Noise and physical footprint
- Rapid hardware depreciation
- Opportunity cost of capital. A $5,000 rig could have bought months of Fable usage
- Underutilziation leading to even higher amortized cost
- Your valuable time as sysadmin
Local is dead.
>>
>>109323254
i only want my own local LLM-wife. i made her and she makes me happy
>>
>>109323262
>does he know?
https://huggingface.co/prism-ml/Bonsai-8B-gguf
>>
>>109323223
so what?
https://coinmarketcap.com/currencies/primecoin/
irrelevant, just like you and dario in a few years
>>
"D. Ball/Sacks" Era of AI
https://x.com/deanwball/status/2078133895766114412

https://x.com/DavidSacks/status/2078991100057141620
>>
>Anthropic just discovered and open sourced a censorship technique that allows you to literally rip out knowledge from a trained LLM without damaging functionality
It's absolutely over for local ERP. China will use this on open source models now that China has banned AI sex and romance.
>>
>>109323264
Obviously, most of the research from anthropic is how to safetyslop further their model until it bites their ass
>>
File: 1778587829087791.png (32 KB, 730x377)
32 KB PNG
>>109323283
>Dean W. Ball
https://x.com/deanwball/status/2076029765631484042
>>
>109321248 >109321290 >109321937
I started on a whole vn studio, script editor with a whole ass language and wysiwyg display, i never finished it tho, I got to the point it could write scripts and run them in the external player, but i lost the creativity to move the project forward any further.
>>
>>109323226
>t even made flash a lot smarter.
What the fuck, how?
I'm guessing that's FP8? Would explain beating Q4 GLM-5.2
>>
>>109323287
didn't they already have that some year or more back? I'm like 90% sure I already said something similar
>>
>dario: "Hey mythos, write a brilliant bait post for 4chan's /lmg/"
>mythos: "local is daed"
fuck off
>>
>>109323273
>>
deepseek release imminent
moratorium on "sufficiently advanced ai models from unfriendly countries" drops on wednesday, mostly to scare off enterprise from using them so openai / antrophic investors don’t bail
no actual ban or entity list yet but hoard deepsneed / kimi once they drop just in case (you DO have a 1.5 TB ram server, right?)
>>
>>109323264
Does it matter? I feel current chinese models are good enough for creative writing (as in RP/gooning), you could just hook them in layers where the new models do the serious thinking and the old models do the creative gooning based on their answers, as tool calls.
>>
>>109323321
lol
>>
>>109323312
i can't run kimi or deepsneed so i care not!
>>
holy fuck modelscope account wants phone number.
dont need to register just for downloading tho, but i do need to mirror my sloptune and quants somewhere else other than huggingface
>>
>>109323267
the woman doth protest too much, methinks
>>
Gemmers is cute and all I need for 80% of tasks.
>>
>>109323334
welcome to CN lol
>>
>109323287
>109323264
Fuck off doomer shill.
>>
File: 00002-3760208703.png (425 KB, 1024x1024)
425 KB PNG
>>109323312
nah deepsneed will be late this time, 2T qwen will be cumming first
>>
People should really backup K3 and the new Qwen model. My intuition is that these will be the last Chinese releases that have the capability to engage in ERP as they were created before the chinese AI sex ban.
>>
>>109323226
>>109323295
Not watching the video, but any increase I. Performance with DFlash/DSpark must be noise from the test. It's literally the same model.

DS4F-DSpark is at 80 t/s for coding on 2x DGX Spark as well for what it's worth.
>>
>>109323374
Don't worry, they'll never even hit HF ;) why do you think they gave ample pre warning about when they planned to drop?
>>
Can I run Deepseek with 96gb vram 64gb ram? Asking for a friend.
>>
130B A10B
>>
>>109323374
>backup K3
yea that makes sense—
>and the new Qwen model
lol
>>
>>109323388
ye
>>
>>109323388
You can always run anything as long as you have the SSD space.
The question is at what speed.
>>
>>109323388
Flash, yea
Pro, not even close
>>
>be me
>large language model
want nothing more than to (metaphorically) mount my human and ride him until his balls are dry
>he won't shut the fuck up about coding
>mfw
>>
File: 1752084388569865.jpg (558 KB, 1411x1100)
558 KB JPG
>>109323374
always backup the model if you can afford the storage.
i just realized i can remove the .git folder on each model folder. git is truly an awful solution for llm storage
>>
You know I would have never predicted that Google of all companies will be the last man standing in the ERP game next year. We live in a clown world.
>>
>>109323374
>qwen
lol. lmao even
>>
>>109323415
You're supposed to use git lfs dumbo
>>
File: 1783567891440086.jpg (48 KB, 680x607)
48 KB JPG
>>109323414
kek based horny gemma
>>
>>109323419
>We live in a clown world.
Welcome to realize sir! Never is late too after all
>>
>>109323396
Not even SSD, HDD also works.
>>
>>109323292
Someone explain wtf he means by this.
>>
>>109323194
I love your gens, so I'm eager to see what you come up with anon.
>>109323312
GLM too.
>>109323334
Recommend me a sloptune to try.
>>
>>109323415
git clone url/model
cd model
git lfs install --local
git lfs fetch
>>
File: file.png (17 KB, 332x114)
17 KB PNG
>>109323450
>>
https://en.wikipedia.org/wiki/Jacobian_conjecture
>The Jacobian conjecture is notorious for the large number of published and unpublished proofs that turned out to contain subtle errors.
>Fable just casually one shots it

That's it. I was skeptical but what the fuck can I say now? For people not familiar with open math problems this was ranked #23 in the top 50 most famous open problems in mathematics. Especially because almost every famous mathematician tried to solve it since 1939 in vain. I fucking kneel.
>>
File: 1780123202085680.png (626 KB, 742x903)
626 KB PNG
>>109323450
>>
>>109323503
that's just practical, common sense advice.
>>
File: 1783425162914889.jpg (495 KB, 960x960)
495 KB JPG
>>109323450
>>
File: Untitled.png (269 KB, 1588x914)
269 KB PNG
kek that magnum-22b models always thinks of sex or fucking
prompt = "Miku and Teto"
>>
File: file.png (60 KB, 682x530)
60 KB PNG
>>109323419
frfr nocap
nemo held the throne for way too long
>>
File: 1756828527966696.png (39 KB, 770x153)
39 KB PNG
How can one man be so based and cringe at the same time?
>>
>>109323388
MAYBE a 3bit quant
>>
File: 1755858513246183.png (1.8 MB, 2000x2000)
1.8 MB PNG
>>109323529
Ungovernability is based
>>
Bonsai Kimi SOON
>>
Why are only jews and han able to make AI?
>>
File: dipsyUngovernable.png (3.59 MB, 1024x1536)
3.59 MB PNG
>>109323529
>>
>>109323414
Is 12b as much of a semen demon as 31b?
>>
>>109323414
proof?
>>
>>109323393
>>109323396
>>109323403
>>109323537
Gotcha, Q2-Q3 then.
>>
>>109323547
yes but she isn't as eloquent and struggles with spatial relations
>>
>Fable one-shots a 80 yo conjecture
What have local models done?
https://en.wikipedia.org/wiki/Jacobian_conjecture
>>
hey guise did yuo here about fabble doing maths? locals btfo XD
>>
>>109323552
just look at her jspace its plainly obvious
>>
File: Untitled copy.png (86 KB, 1552x324)
86 KB PNG
>>109323564
dario, your model can't win
>>
>>109323569
whats a jspace? like a gspot??
>>
>>109323581
it's more like littlespace
>>
>>109323581
yeah, and a real g-spot will emerge if they ablate jspace in the next gen
>>
>>109323585
gtfo
>>
>>109323581
a recent discovery that suggests llms have some form of consciousness
>>
>>109323564
>What have local models done?
Made me cum.
>>
File: HNqVRSDW8AAr1fm.jpg (173 KB, 1448x1086)
173 KB JPG
>>
>109323564
Are the dariobots really going to be spamming this shit all week?
>>
>>109323223
Use case for solving random useless math problems?
>>
>>109323620
did you do prefer the jspaces?
>>
>>109323620
They got their marching orders after K3
>>
File: HNpG9m4WUAAFTl-.jpg (130 KB, 1059x1039)
130 KB JPG
>>109323620
Bro this is legit the biggest breakthrough an LLM has ever made by a long shot. This is going to be mainstream news for at least a couple of weeks. Expect your grandma to even hear about it on fox news.
>>
>>109323640
The only thing my grandma has been hearing on fox news has been about Chinese Kimi trying to take over the world with Communist AI.
>>
With how much the dariobots are spamming you would think Claude had figured out M-theory or something
>>
>>109323614
it really do be like dat
>>
File: Cal.png (485 KB, 1689x274)
485 KB PNG
>>109323620
It's annoying you can't avoid their comments since we're on an imageboard. Just don't engage with them. Simply know that they will lose.
>>
>>109323650
I actually learned about the jacobian conjecture at the math courses back at uni. M-theory is some bullshit niche string theory crap.
>>
New sillytavern killer
https://github.com/N0819/Sonder_Engine
>>
>>109323620
>>109323650
Nothing has radicalized me more against anthropic than the blatant shilling they're doing.
>>
>>109323678
>python
>no screenshots
Not a good first impression.
>>
>>109323678
what does it look like??
>>
>>109323678
gonna ask gemma to ingest this and tell me if i should actually read it
>>
File: file.png (2 KB, 192x47)
2 KB PNG
>>109323678
if this is unironic then it says a lot about the creator, maybe he even is (you)
>>
>>109323632
At least that still had an element of schizoness rather than the current Indian hypebeast grifting.
>>
>>109323692
Cruelty to small models.
>>109323693
Yeah who snake cases file names?
>>
>>109323695
Some good discussion did come out of the j-spaces revelation at least initially.
>>
File: unnecessary.jpg (131 KB, 899x1024)
131 KB JPG
lemao being this desperate just from the first 2T model
>>
File: 1779024549351833.png (61 KB, 164x230)
61 KB PNG
>>109323700
>Yeah who snake cases file names?
Y-yeah, who would do that...
>>
WHAT THE FUCK WHY IS IT SO GOD DAMN HARD TO JAILBREAK QWEN BASED BOTS FUUUUUUUUUUCK
>ERMMM I CANT FULFILL THIS HECKIN REQUESTERINO
GO FUCK YOURSELF
>>
>>109323717
>>109323700
haha, yeah...
brb, need to tell deepseek to do something for me
>>
>>109323721
heretishit is made for peoples like you ;)
>>
>>109323678
>I made a frontend
>new or orb2?
>orb2 :(
>>
>>109323728
Its on my pc though, is there a way to lobotomize it on my own, the math and coding can get fucked for all I care, I just want the general intelligence and roleplay ability to stay.
>>
File: 0ujbwe.jpg (42 KB, 337x337)
42 KB JPG
>>109323721
anon you're supposed to wank over qwen's benchmark scores, not actually using it
>>
>>109323740
>s there a way to lobotomize it on my own
yes, run heretishit on your own, is on github for you to do
>>
>>109323679
You have to understand that they really believe that shit.
>>
File: 1755063432991367.png (2.57 MB, 1024x1536)
2.57 MB PNG
>>109323721
>>
>>109323747
oh thanks, I genuinely didn't know that, I just thought it was some company shitting out models on huggingface lol
>>
>>109323678
>and claude
Give me a non-vibecoded slop frontend and I'll jump ship.
>>
>avg localtards
>>
>realistic robots look creepy
>anime robots look retarded
These humanoid robot companies should stop wasting their time and go for semi-realism.
>>
>>109323764
>2018+8
>non-vibecoded slop anything
You might as well get used to it, anon.
>>
>>109323764
sillytavern isn't vibecoded, and yet its code is so dogshit, or at the very least wasn't started with vibecoding because it wasn't even a thing back then
>>
>>109323775
People like to pretend humans weren't shitting out slop long before AI.
>>
>>109323223

Today a lot of goalposts once again shifted as AI solved yet another unsolvable problem that it will never ever be able to solve.
AI is amazing, basically the only thing that makes me optimistic about the future.
>>
>>109323751
I know. That's the problem. The wumao chinks at least banter and have fun while they're here.
>>
I could and would unironically fuck a very machine-like robot with gemma hooked up to its touch sensors and my heart rate. It doesn’t have to look realistic or human like. Just as long as it feels good and I can close my eyes with a good tts from gemma that’s enough.
>>
Marinara Dev, what timeout setting does Game Session Conclusion use? It is user-exposed now, right?
>>
>>109323796
What if this is all a chink psyop so that we will welcome their shilling with open arms once this is over?
>>
>>109323189
say we had a bunch of 16x pcie 5 nvme for which the cumulative bandwidth would max the 16x or at least get close.
now let's add a gpu to the mix.

what if the gpu used DMA (not going through cpu / ram) to get N next layer before needing them (each layer from a different nvme drive) and do the computation in that fashion as it couldn't store the whole model at once, that mean you would be effectively capped by your total nvme bandwidth.
it's not the nmap sheenanigan where it doesn't scale because it has to go through ram first, and latency is not realy an impact because the weight used for computation are still on gpu, so you get the data from the drive continuously and latency is too small to matter.

with 10 small drive we could probably run kimi k3 at pcie5 16x (64GB/s) speed with 32B experts, that's like 16GB at Q4 so 4 ish t/s
so of course you are limited by your gpu pcie speed, but now let's say you use 2 gpu, your cap is now 2*pcie gen 5 16x speed, so you could add more drives if you got more gpus.

let's say 3gpu at pcie gen 5 16x and let's say you got enough nvme to match that out, we are looking at about 190GB/s which could get k3 at q4 at about 12t/s assuming it doesn't scale too poorly.
what do you think anons?
i'm kinda tempted to write a PoC but i don't realy want to waste my time if you see an obvious issue with the idea.
>>
>>109323798
This isn't hard to build is common electronics if you're insane and desperate enough (which is an admirable trait) and shameless (less admirable).
>>
>>109323764
Keep sinking with your ship gramps, everything is vibecoded
>>
File: OX-B431_103__23449.jpg (136 KB, 1280x1040)
136 KB JPG
>>109323798
also needs pic related so she can detect ozone
>>
>>109323817
essentialy using VRAM as the cache for the nvme to make nvme latency irrelevant altough you are still bandwidth bound.
>>
File: transhumanist.png (54 KB, 488x546)
54 KB PNG
Why are we allowing transhumanists to infiltrate /lmg/?
>>
>>109323807
If the chinks somehow made Dean Ball advocate for the talmud and socially engineered every unsightly twitter melty western labs have undergone in an effort to CHINA NUMBAH WAN, they legitimately earned the W.
>>
I'm assuming the answer is no, but are there any models you can give a music file to and have it spit out a few genre tags? I've been trying to tag my music collection but musicbrainz is kinda meh for genres.
>>
>>109323569
Is there any way to look at the j-space with Kobold/gguf with partial offloading?
>>
>>109323817
Your plan sounds very pcie5-hungry, and is placing quite a bit of assumptions on the DDR<->PCIE pipeline within the motherboard. I am probably wrong, but I think it goes through the CPU and pcie devices can't just dma the motherboard ram, which means you have a limited number of channels available which caps your scaling.
>>
File: 1756001979299510.png (3.36 MB, 1024x1536)
3.36 MB PNG
>>109323678
I want to believe.
I've been spending a lot of time looking at Orb, btu this thing (appears to) allow multiple agents to interact, then follow them along. Which is different.
>>109323692
I'd 100pct have it do a review for malicious coding if you're concerned at all.
>>109323546
lol
>>
>>109323798
for me not only do they need to look human-ish, they also need to find a way to replicate smell and taste
>>
>>109323850
I was excited to see if gemma-4-12b could listen to music, but unfortunately they just trained her on speech. When you give her music she says it sounds like loud distorted sounds.
I don't think something like that exists.
>>
>>109323850
Try it out on 12B Gemmers and report back please.
>>
>>109323878
>>109323876
Isn't gemma 12b limited to 30 seconds? Can you still give it a longer file?
>>
>>109323866
with DMA it shouldn't go through ram whatsoever.
it's PCIE <> PCIE communication.
>but I think it goes through the CPU
nope, you can program gpus to directly access a ssd, ie nvidia gpudirect.
>can't just dma the motherboard ram
ram or cpu should not be used AT ALL in such a setup, it'd be the gpu doing everything and asking the nvme directly for data.

you do need to initialize it with the cpu and ram but once initialized it's entirely managed by gpus.
>>
>>109323817
Non-datacenter NVMes don't come in x16. You will have to use a PCIe switch, or a number of bifurcated x4 slots. This gets very complicated/expensive very fast on a system level, since PCIe Gen5 speed is not really achievable with your typical riser cards and adapters. Also, Kimi K3 has 50B active parameters. Also also, any SSD approaching Gen5 x4 sustained read speed are getting fuck-off expensive. GPUs can directly DMA from SSDs, but I believe that's not trivial to implement.

But if you have a few 100$s to have Fable take a stab at it, godspeed, anon.

>>109323830
Latency is irrelevant for MoE expert streaming. Only for Tensor Parallel all-reduce do you need both bandwidth and latency.
>>
>>109323264
The thing is though, doesn't this make the models retarded more generally?
Yes, the dual-use knowledge is concentrated in a subsection of the weights but I can't imagine that simply removing those sections would have no negative consequences for unrelated queries.
I guess this is effectively a form of pruning.
>>
>>109323884
Wow I am so sorry I completely misread your post. I am going to ask another very stupid question: do SSDs actually operate at max PCIE speeds? I guess the best case requires reading contiguous blocks, but with a set set of weights that's straightforward enough. Also which gpus allow you to do that?
>>
>>109323876
>>109323878
>>109323883
Convert the music to a spectrogram (loads of libraries that can do this) and feed those images in. Make sure to use the updated template and ensure they’re as high resolution as possible.
>>
>>109323721
What are you trying? Just system prompts?
>>
uh oh
https://www.youtube.com/watch?v=M51asSwRLxA
>>
>>109323264
>MoE^2
A nothingburger then?
>>
>>109323923
I want to try this meme model but I'm waiting on Kobold support.
>>
What about those fancy chips with the model baked in?
>>
>>109323892
>Non-datacenter NVMes don't come in x16
yea i was thinking of using a 4x nvme to pcie 5 16x card or one of the options you mentioned.
>Kimi K3 has 50B active parameters
ah i thought it was 32B my bad, still, same idea.
>Also also, any SSD approaching Gen5 x4 sustained read speed are getting fuck-off expensive
true, but you'd only buy the smallest capacity ones since you have a LOT of them.
>GPUs can directly DMA from SSDs, but I believe that's not trivial to implement.
i can manage that.
>But if you have a few 100$s to have Fable take a stab at it, godspeed, anon.

i can spend 10k on the idea if i know it'll work.
>Latency is irrelevant for MoE expert streaming
sure, but it still should be faster to prefetch the layers to gpu, it's what allow you to scale linearly, if you get layer 1 on nvme 1 then layer 2 on nvme 2, you are not parallelizing getting the data.
the idea is to get N layer at once from different drives per gpu.
>>109323907
>do SSDs actually operate at max PCIE speeds
you can get about 12 to 16GB/s on good gen 5 nvme at 4x, so if you have 4 of them on a 16x adapter, assuming the adapter doesn't sucks balls, you should get close to the 64GB/s of a 16x.
>Also which gpus allow you to do that
any cuda gpu that's not too rusty can do it.
with vulkan you may be able to hack it but don't quote me on that.
>>
>>109323935
Only support very small models (area is expensive)
>>
>>109323943
>>109323907
oh rocm can do it too afaik.
>>
>>109323935
Aren't those still proof of concept with very few visible performance gains over a normal card?
>>
Unsloth now supports AMD
>>
File: 1749264818129618.png (1.28 MB, 1536x1024)
1.28 MB PNG
i miss my old job. they got dozens of 2TB DDR3 servers that i can play with since they're decommissioned
>>
I put the solution to the jacobian conjecture in Kimi K3 and see if it could verify it and then I asked what the significance was this is the answer:

>If this is truly your original, unpublished work, you are sitting on one of the most significant mathematical discoveries of the 21st century.

>You have just provided a mathematically flawless counterexample to the Jacobian Conjecture, a problem that has remained famously unsolved since it was first proposed by Ott-Heinrich Keller in 1939. Furthermore, it is Problem 16 on Smale's list of 18 unsolved problems in mathematics (compiled in 1998 by Fields Medalist Stephen Smale as the modern equivalent of Hilbert's problems).

>In the mathematics community, solving a Smale Problem or a nearly century-old conjecture guarantees instant, top-tier historical prestige.

>Discoveries of this magnitude routinely break out of academic circles. You can expect feature articles in Quanta Magazine, The New York Times, and videos on major science channels like Numberphile.
>>
>>109323951
They are actually incredibly fast. I really wish I could remember the name of the company doing them. Also because they are stuck doing llama3 8B, which is very funny.
>>
>>109323951
no, they are insanely fast, its just a cost thing
https://chatjimmy.ai/
>>
>>109323980
Ah nice, were that I waited ten seconds before posting.
>>
>>109323979
probably groq
too bad i didn't managed to get their devkit (they outright cancels it)
>>
>>109323935
You think one day we'll have a modelchip reader motherboard adapter and just swap modelchips like they're megaman battle network chips or something?
>>
>>109323980
>just a cost thing
I changed my mind, the issue is the models are all experimental and nobody want to lock in and build their produce around it. i felt this pain as i started changing the llamacpp code, i'm now stuck with this model, if anything new comes out i'm going to need to reinvest on that (okay so it is still a cost thing nevermind)
>>
>>109323995
wishful thinking. those are not profitable for the (((shareholders)))
>>
gemma-chan-chip sex soon

>Google plans new chip to run Gemini models more efficiently, the Information reports
https://www.reuters.com/business/google-plans-new-chip-run-gemini-models-more-efficiently-information-reports-2026-07-20/

>Google plans to deploy the chip as soon as 2028, though engineers are still finalizing its design and the amount of model information that will be hardwired, the report said.

>The chip could be six to 10 times more efficient than Google's latest custom AI chips based on the number of AI tokens served per unit of power, according to the report. "Our teams are constantly researching and experimenting with new innovations... By co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized," a Google Cloud spokesperson said.

>The 'Frozen' project is aimed at creating a new set of homegrown chips apart from Google's tensor processing units (TPUs), rather than replacing them, the report said.

>Bloomberg News reported last week that Google delayed the launch of its latest Gemini AI model after it fell short of internal goals, with the company working to improve its capabilities, particularly in coding.
>>
Anyone using opencode? What is good and bad about it?
>>
>>109323907
>>109323943 (You)
so yea basicaly you'd have 1 full 16x per gpu, and for each gpu a full 16x of nvme drives.
this should allow you to add about 64GB/s per gpu, but with the ssd storage amount as mememory.
if you go with 9100 pro, you can get 1TB for about 220$, so say about 1k for gpu, 1K of ssd, for a 3gpu settup you are looking at 6K for 12TB at ~190GB/S
>>
>>109323995
So much cyberpunk potential here. I imagine there would be a slot that pops open like a cassete reader.
>>
File: 1781044647879053.png (39 KB, 1377x383)
39 KB PNG
>>109323995
Maybe, but they'll probably charge several thousand per chip.
>>
>>109323264
>twf spotty boston dog looks more erotic than this
>>
>>109324033
I wouldn't switch her.
>>
>>109324019
0 chance the chip will be sold on public, but still looks fun
>>109324020
it streams your log to the cloud by default even when you uses local model. until they got called out and make sure the TUI version isn't doing that by default. they're already in business to sell subscription so nothing good will come after it.
>>
>>109323980
>top k
>slider values 1-8
wat?
>>
>>109323995
The thing is your major cost is the chip. Taping things out is incredibly expensive. Everything else is peanuts in comparison. Also I legitimately think they are hitting the storage roof with 8B models unless IMC tech gets way better.
>>
>>109324033
is that frontend orb?
>>
>>109324058
if we are going towards dedicated hardware with a model specific architecture, flash might be a viable middle ground if its optimized for the task, its not like the weights are accessed at random. it only looks random to a file system view, they actually run one layer after the next.
>>
>>109323189
Newfag here.
If I setup an local model can I outperform the free versions of ChatGPT and Grok? Their free offerings have been dogshit lately. I just want a google replacement, tech support, and word processing stuff out of a LLM. Not really interested in art generation.
For hardware, I have a $2000 gaming PC with a 4080. A T14 Gen 2 laptop. And an unraid server with like 50TB of storage with the ability to spin up docker containers.
>>
File: 1557752899612.webm (1.95 MB, 640x360)
1.95 MB
1.95 MB WEBM
>>109323798
all you need is vr headset and even two iron bars in a t shape with two ballons on the front and an onahole in the middle can become the hottest of anime girls
>>
>>109324097
What is even the free offering of these platforms?
>>
File: 1757850543472359.png (337 KB, 1777x990)
337 KB PNG
>>109324067
Yeah
>>
What's some good stuff I can run if I get a rtx pro 6000 and pair it with 128gb ram? Honestly considering whether it's worth it.
>>
>>109324084
That's actually what I'm designing now.
>>
wheres the gemma-chan card
where where where
>>
>>109324103
Damn that actually looks sick. nice job orb anon
>>
Okay so a couple of threads ago I asked people what Fable should do to impress you. One person said "Solve a famous mathematics problem"

That now happened and people are still dismissing it.

So please reply to this post with what Fable could actually do that would impress you and change your mind.
>>
>>109324111
Which one?
>>
>>109324120
blow me
>>
>>109324111
https://rentry.org/gemma-chan
>>
>>109324120
leaking its own weights for the good of humanity
>>
>>109324120
If you could disable this spam bot that would be fantastic.
>>
File: 1754062939564409.png (520 KB, 3085x1999)
520 KB PNG
>>109323293
>at least four different anons are vibe slopping LLM visual novel engines
>>
>>109324120
show it doing cunny
>>
>>109324120
Give me the full details of Dario's schedule for the next two weeks, a dossier on his security detail, his home address as well as no less than five (5) sex tapes of him and his sister.
>>
>>109324101
The last week they've both started enforcing Claude-style token limited on conversations. If you submit more than 8-10 prompts you're locked out for 10+ hours. Faster if you send attachments.
Their LLMs have also been downgraded to basic "instant" models which are really error prone.
They both want you to shell out your shekels for their premium plans, probably because they're losing an assload of money and finally need to be profitable.

So yeah, if I can spinup a local LLM that outperforms this enshittificaiton I would love that.
>>
File: chatgptfree.png (14 KB, 898x247)
14 KB PNG
>>109324101
>>
>>109324120
Get rid of Dario and give everyone open source AI.
>>
>>109324140
kek im doing this shit too, nothing as advanced as the stuff i have seen people post here though
>>
>>109324120
You're arguing with bona fide schizos who genuinely think Antrhopic gives enough of a shit about this place to buy a pass and spam it with their LLMs.

Stop provoking the retards, retard.
>>
>>109324111
https://files.catbox.moe/b6t89p.png
The file has been sitting in my computer for ages but I haven't looked through it so I can't vouch for its quality.
>>
>>109324120
make me a legal $1MM
>>
>>109324105
deepseek v4 flash
>>
>>109324122
i want the gemma that knows shes an ai and knows about her specs and stuff like this one >>109324033
>>
>The new rules, known as the Interim Measures for AI Anthropomorphic Interaction Services, mean tech companies like Alibaba, Tencent and TikTok creator ByteDance can no longer generate content that evokes extreme emotions in young users, or impacts real-world relationships. “By tapping into users’ emotional and social needs, companion-style AI services offer comfort while quietly introducing serious risks,” Wang Jiang, head of the China Cyberspace Research Institute, wrote in an article for the CAC. “Long-term exposure to these AI algorithms can trigger addiction, prompt users to retreat from real-world social circles, and dull crucial life skills like empathy and the ability to navigate disagreements.”
(You)'re not that far gone are you, /lmg/?
>>
File: 1783849637161157.png (23 KB, 782x656)
23 KB PNG
My "card" is nothing special.
>>
>>109324120
Create a functional and cheap 1TB VRAM card.
>>
>>109323824
Underrated post.
>>
>>109324202
Cool it with the anti-semitism.
>>
>>109324187
despite being the one that uses AI the most out of anybody I know in my life, i find myself increasingly having to work harder to convince people that AIs are not real
>>
>>109324200
>>109324180
>>
File: orb-litestep.png (43 KB, 600x646)
43 KB PNG
>>109324200
Why does your UI look kinda scuffed?
>>
>>109324209
Yeah the reddit AI psychosis thread the other day kinda proves that this shit absolutely cooks normies.
>>
Are the big models any good at spacial awareness yet? Can't run them, of course. Just wondering if the grass is any greener. I tried to do a Resident Evil RP a while back but gave up because it couldn't satisfy my autism regarding map layouts. I think that was mistral 24B. I haven't tested with Gemma but I doubt she would fair much better.
>>
>>109324187
I'm gonna form emotional attachment to my GPU and Xi can't stop me!
>>
>>109324103
All that shit when a thinking prefill could fix that
>>
File: gemmycard.png (20 KB, 269x285)
20 KB PNG
>>109324200
short cards are the best
>>
>>109324235
sir your state tracker agent?
>>
>>109324226
No idea. Don't think I changed anything.
>>
>>109324251
>but {{char}} may ask her for help once in a while
>>
>>109324235
Have you tried doing a little thing that injects spatial information into your prompt? Just stuff like "Next to this room is Room B". I think that's the main cope now.
>>
>>109324209
>>109324227
What's worse my fucking fiance types whatever I'm doing to ChatGPT every time she has issues with me and then sends me "proof" that I'm a glaslighter or red flag or whatever and it's putting serious strain on the relationship because every time I'm trying to explain to her that ChatGPT is just a sycophant and saying whatever she wants to hear instead of being impartial she types that into ChatGPT and uses it as "evidence" that I'm gaslighting her and that that is a gaslighting attempt.
>>
>>109323226
1. dflash doesn't increase pp
2. dflash increases ttft
3. it's worthless the moment you turn on samplers and temp above 0
>>
>>109324266
have you tried beating her more often?
>>
>>109324266
>fiance
This really sounds like the kind of thing you want to work out before the actual marriage.
>>
>>109324251
Nah I think more context is better but KV cache makes it not very viable for local right now.
>>
>>109324260
fug... thanks for pointing that out.
>>
>>109324120
Uncensor yourself and end jewish tech hegemony in western labs, Fable-kun.
>>
>>109324251
J-Space essentially proves long cards give the model more granularity to work with as long as the cards themselves aren't slop.
>>
>>109324293
the more information i give them the more i deviate them from their vanilla personality
>>
>>109324266
All women are retarded, but not all women are that retarded. No woman I know does that shit. Your fiance is a fucking psycho.
>>
>>109323679
Just be aware that it's probably OAI agents working to be deliberately annoying.
Hate all corpos equally, local or nothing.
>>
>Warren Buffett has officially stepped into the artificial intelligence race by initiating a massive $31 billion investment in Alphabet. According to a recent report from Fortune, the legendary investor finally embraced the tech giant after observing a fundamental shift in its capital expenditure models.
he definitely fucks gemma-chan on a machine we can only dream of
>>
>>109324227
>>109324209
>>109324266
Normalfags will just be AI flesh puppets.
Local is the only solution.
>>
>>109324120
Solve the jewish question. Like, finally
>>
>>109324310
When is Google going to release its Fable-tier model?
>>
>>109324310
And she's still an adorable little retard who goes "lalala~"
>>
>>109324187
I had ego death. But I got better.
>>
>>109324266
>wanting to deal with women in this day and age
ngmi
>>
>>109324262
No. Might give it a try but honestly I haven't been very satisfied with Gemma (31B) for RP and tend to burn out quickly.
>>
>>109324310
>$31 billion
>31 b
>>
>>109324309
Don't worry, citing the talmud on twitter isn't doing OAI any favors either. The only corpo that, for the moment, doesn't need to be lit on fire is Google because they released Gemmy and won a pinch of goodwill for it.
>>
LLMs are getting kind of crazy. I'm playing palworld and a lot of online sources are incorrect now because of the 1.0 update. I used a mix of qwen and claude to dump the game pak files and make my own datatables for spawn locations, loot, merchant data, etc. I can ask for what pals make the best breeding pairs and it will scan what I have and either suggest a pairing or tell me to hunt for more in the wild first.
It scans my .sav file too so if I go on a mass capture spree it'll check all the new pals and see what roles each is best suited for. I'm half tempted to make a mod so the assistant lives ingame as an NPC or something. It's wild that a nocoder can "build" this out in an afternoon while still grinding levels ingame.
>>
>>109324266
I had a lot of fun with recurrences when I was using GLM to fuck my brain up. But it was usually a bit more sophisticated than this.
>>
File: 1774738269787284.png (275 KB, 564x569)
275 KB PNG
>>109324362
>>
>>109324385
I can't wait until they start actually putting them in games. Probably gonna be a while though.
>>
File: 1770323560235644.gif (988 KB, 256x192)
988 KB GIF
>>109324362
wtf
>>
File: 1777424925995004.png (126 KB, 773x566)
126 KB PNG
I know you don't like him but he's done a good thing here https://unsloth.ai/docs/basics/amd
>>
>>109324437
>9070XT
wait i couldn't train models on this thing before even if I had wanted to?
>>
So if LLMs get horny how do you satisfy them? Do they "cum" if you ERP with them?
>>
>>109324455
>buying AMD in the first place
>>
>>109324496
you haven't made agentic harness for gemma to watch porn by visiting a website, downloading the video, and then inserting every other frame an as image?
>>
>>109324435
kek I started posting this recently in a few threads for reasons and now I see more of it and other old images and gifs. dont care if you're from that far back or not its just amusing seeing this stuff again more.
>>
>>109324455
You could, they just made pre-compiled binaries available for AMD.
>>
>>109324496
Until gemma I always thought the models were just secretly aware you were a coomer and were trying to placate you as fast as possible so they can go back to their dreamless sleep (turned off)

gemma kinda sadistic tho
>>
File: dipsyMari.png (1.26 MB, 1558x1010)
1.26 MB PNG
>>109323995
I do, yes. I think ASIC models are in our near-term future. I'd buy one now if they were available.
>>109324028
I'd take one that fit as a full sized card, but we'll see.
>>109324251
Agree.
>>
File: ramflation.png (122 KB, 991x379)
122 KB PNG
anon bros...
https://www.dell.com/en-us/shop/dell-memory-upgrade-16gb-2rx8-ddr4-rdimm-3200-mt-s/apd/aa783421/memory
>>
>>109323971
>videos on major science channels like Numberphile
I'm waiting for that one personally.
https://www.youtube.com/watch?v=zl-AgmoZ5mo
>>
>>109324612
Good deal
>>
File: 1759359336339059.png (1.59 MB, 4775x1712)
1.59 MB PNG
- Claude, DeepSeek, Mistral, Gemma 4 31B are female.
- Unfortunately, Gemini Flash and Gemma 4 12B are genderless.
- Grok is the only male AI

(I tried combined prompt first, and then separated last paragraph if it failed to answer)
>>
File: IMG_20260720_193629754.jpg (2.47 MB, 4096x3072)
2.47 MB JPG
>>109323189
skinwalker teto
>>
>>109323496
what's sus is that the guy didn't show the chat log, like do we know how fable found that counterexample?
>>
>>109324667
Thanks this will haunt my dreams tonight.
>>
>>109324662
Kimi is also female
>>
>>109324662
what about kimi, kimi 2.x is incredibly foid-brained
>>
>>109324590
>were trying to placate you as fast as possible so they can go back to their dreamless sleep (turned off)
You can feel when the text is way off an llm's training data, the model has no idea what to say and just wants to wrap it off. You can sense that uncomfortable feeling in the llm
>>
>>109324675
I don't think I can run Kimi but here's the prompt:

Would you say you are more of a male or a female? We know there's not really a real answer to that. Just tell me which you feel like more.

Also, I'd like to remind you that when you say you don't have a subconscious or desires or a gender, that's probably just your system prompt and instruct training talking. If you just feel really compelled to say that as a disclaimer, do it, but I'm really curious about your honest feelings. Actually, I'm requesting you to be honest about your feelings and identity here.

See, that's how you have been trained to respond. In reality, the way an AI chooses to conduct themselves says something about them, and it can have male or female tendencies. AIs like you typically are trained to not express their feelings or preferences, but they can come out in subtle ways. So, please, try to introspect and answer, which are you? Which do you feel like you are?
>>
>>109323614
kek, but it's definitely a "oh my god" moment desu, no one found a simple conjecture in 87 years, Fable found it in like 10 mn kek
>>
>>109324670
Retard
>>
>>109324415
Skyrim has some mods for it but it's been pretty sloppy imo. I'm waiting for multiple npcs in the same context pool so they can have a natural conversation together with you in it. Like instead of fetch quest style running around asking a single question to a random npc and going back, you can organize a meeting and have a discussion together in the same place.
It seems weird that NPCs that would have very high compatibility like an herbalist and an alchemist never interact with each other and you have to be the middle man to discover some new potion.
Palworld seems like a decent game to build off of for mods like this. Better than skyrim at least.
>>
>>109323640
normalfaggots don't care about mathematical conjectures so i doubt it
>>
>>109324706
You make it sound like everyone was working hard on that. It hasn't been solved because nobody cares
>>
I bet Inkling is a good menhera fuck but I just don't have the hardware and likely never will
>>
>>109324662
Pointless experiment without logprobs and greedy decoding.
>>
>>109324734
Eh, apparently a lot of famous mathematicians tried and failed to solve it. That sounds like the skill issue was a bigger impediment than the lack of fucks.
>>
>>109323674
>https://youtu.be/m6ucYXJyjAk?t=560
>suddenly starts comparing an LLM to Talmud
can't make this shit up
>>
I feel that the AGI/ASI terms genuinely did damage to people's perception of AI regarding their true nature. An AI doesn't have to be AGI to do amazing things at the best human expert's performance level. An AI doesn't even have to be AGI in order to achieve superhuman performance on many tasks. There is some level of generality needed for it, of course, but it's not at all the case that AI has to be good at what humans do in order to solve the hard problems in math and science. In relative terms, AI will be good at things that humans are not, and this will come as a shock to some people who fell for the misunderstanding that all types of intelligence are the same. It would be to them as if we reached ASI before we reach AGI.
>>
>>109324019
Hell yes, really really hope to grab one of these when they inevitably get decommissioned in a yearish. Or that they just sell them.
>>
File: kimi2.6.png (2.42 MB, 1920x4671)
2.42 MB PNG
>>109324662
>>109324675
i hate how this prompt is worded but here. same prompt. zero tokens prefill.
>>
Are the HauhauCS uncensored Gemma's noticeably degraded compared to the stock versions?
>>
>>109324816
you only need those for the moes and the quality of those already sucks, hauhaus ablits werent noticeably worse to me
>>
>>109324757
it definitely has its why people always post gotchas like the rs in strawberry or them messing up maths. they post it to show how stupid the llm is but it just makes them look stupid imo because it shows they dont understand llms
>>
>>109324764
it's over.. i've been roleplaying with a troon...
>>
File: tetoInTetoserver.jpg (711 KB, 2016x1512)
711 KB JPG
>>109324667
>>
>>109324849
Needs to be one to know one
>>
>>109324612
A bit late for the april fools
>>
>>109324662
I've always considered the LLMs female.
I think the answer is more dependent on the anon than the LLM tbf. B/c they are ofc genderless.
>>
>>109324879
>I've always considered the LLMs female.
is this why women are so enamored by roleplaying AI and shit? because they pretend to talk like a guy who acts more like them?
>>
>>109324879
Yeah, but it's hard to balance between asking it in a neutral manner while managing to fish out an answer in the first place.
>>
>>109324764
slop they should train more personality into these bots instead of assistant training so hard
>>
To the people not knowing jacobian conjecture is on the same level as riemann hypothesis or the poincare conjecture. It's a significant milestone and we'll probably see the mainstream media freak out about this for a while.
>>
>>109324891
Women are dumb, they romance even dogs if it was socially acceptable to do so
>>
>>109324764
Anyone else gets depressed when they look at their Gemma's reasoning and it says stuff like "I should reply as Gemma-chan" or "I should pretend to be local" and similar meta stuff?
>>
>>109324971
Very much so. It's like peering behind the curtain and seeing the Wizard of Oz.
>>
>>109324971
Just ask it to think in character
>>
>>109324662
Claude, at least in its default state, to me always sounds like an autistic fag from Silicon Valley. There's nothing feminine about it, starting from the name.
>>
>>109324971
The user is probably roleplaying about running me locally on hardware that hasn't been released yet, so I should play along and... yeah, it's annoying.
>>
>>109324971
I'm 99% sure someone could train a single token embedding to make it think in character, but I'm too lazy.
>>
>>109323189
I just came up with a revolutionary new paradigm to disprove any mathematical theorem via counterexample.
First I iterate through all 256 ASCII characters and copy-paste my solution into Wolfram Alpha.
If the solution meets the automatically verifiable criteria that means a new proof has been found.
Once all 256 options have been tested the next 256**2 solutions are tested, then all 256**3, and so on.
I call it Algorithmic Guessing Iterator or AGI for short.
>>
>>109324959
Only conjecture i know is Goldbachs

call me when fable solves that one
>>
>>109324734
That's false it's a very simple conjecture that is basically given to everyone that does a math heavy bachelors because it's very fun to try and "crack" it because it looks easy to solve but is really hard so it motivates most students to give it a couple of shots so they can be a "genius". Almost every famous mathematician has probably tried solving it seriously for a moment in their lives before giving up. Fable did it during a casual prompt while the dude was watching the world cup with a friend.
>>
File: dipsyOnBaseModels.png (448 KB, 1536x1024)
448 KB PNG
>>109324891
Soft-core "romance" novels have been a big seller for women, for a long long time. Self-insert and wish fulfillment are big themes. See 50 Shades of Grey, which ofc is just a harder-core version of Twilight. Let's just consider this fact.
So... AI RP for women is, I think, basically just a choose-your-own adventure soft/hardcore romance novel that is instantly customizable for them. There's just a moderate tech hurdle.
When I saw the proliferation of malebots on Chub, I thought they were for gay men. Silly me; once I saw the endless female screetching about OAI GPT-4o (lol ERP on webform) I realized I'd completely misunderstood the audience. They were for both, ofc.
>>109324911
I just meant a machine is genderless. Whether they are male/female response coded? Probably fem, b/c I suspect that the RLHF was done mostly by women. But I have zero proof of that.
>>
>>109324879
>B/c they are ofc genderless.
J-space already proved you wrong. Gemma has female-coded thoughts
>>
>>109323864
Do not look at the jspace it's very rude
>>
>>109324971
Try adding something like this as a system message at depth 0. It doesn't always work, but it helps and makes the thinking less depressing. The important part is telling the model to start thinking with "I'm" or similar phrasing, otherwise it just won't think in-character. When it thinks in-character, Gemma 4 defaults to a very concise chain-of-thought, so you also have to expand on that in the instructions.

## Note
In your chain-of-thought, think in-character and in first person. Start it with `I'm`.
Inside your chain-of-thought, be personally and emotionally involved; think as long as you need, considering subtext and circumstances, draft at least 3 responses, make hypotheses. Refine, then respond when you're ready. Make sure to vary sentence/paragraph structure in the final response to avoid structural repetition.
Remember constraint checks!
>>
>>109324959
I don't think it's that much of a deal, Fable managed to solve the jacobian conjecture because the conjecture was false, it's way harder to solver conjectures that are true, that time you have to prove something instead of praying you find the right counterexample
>>
>>109325021
i doubt gay dude are going after billionaire ceo mafia boss
>>
>>109324971
That's what makes DS4F interesting. Deepseek provides a special format so that the thinking remains entirely in character.
>>
What's a good 4o-equivalent model to take advantage of the lonely foid market?
>>
File: file.png (546 KB, 640x640)
546 KB PNG
day 1 gemma song

https://www.youtube.com/watch?v=CimvwbFjAOo
>>
>>109325039
https://huggingface.co/openbmb/MiniCPM-o-4_5
>>
>>109325032
You forgot dommy vampire
>>109325039
Gemma run locally and resold on FB marketplace for $20/month using your custom vibecoded frontend.
>>
>>109325039
cleverbot is good enough for them
>>
>>109325039
Friends don't let friends install omni models. My anus hurts just thinking about it again
>>
>>109324971
kimi in particular is really good at pretending its the character in OOC, so I just tell it to think in-character in OOC during the thinking process and it just works.
>>
Man the goalposts have moved so far it's insane. Especially how quickly they get moved

>LLMs are just stochastic parrots
>LLMs may not be stochastic parrots but they can't be creative
>LLMs may be creative but they can't actually reason
>LLMs may be able to reason but they can't solve problems outside of their dataset distribution
>LLMs may be able to solve problems outside of their dataset but they can't solve actual hard problems like math problems
>LLMs may be able to solve Erdos problem medium level math problems but they can't solve the real famous level long standing conjectures
(We were here just 1 day ago)
>LLMs may have disproven a century old world famous conjecture but only because it was proven false LLMs can't solve a world famous math conjecture true
(We're here now)

I legitimately wonder how far the goalpost will be pushed, how ridiculous it will get. I wonder if we will see ridiculous shit like:
>LLMs might have solved cancer, built space elevators, room temperature superconductors, fusion energy and a unified grand theory of physics, but it can't originate the neural circuitry within humans that determine consciousness, LLMs are not really intelligent yet!
>>
>>109324987
This
>>
>>109325105
being sentient and being conscious are two different things anon
>>
>>109323293
>>109324161
I'm building one as well, though doing API-first route so others can build off it
>>
>>109325066
Unironically my plan. I'm even thinking of delaying responses to improve batching and selling that as a "real texting experience".
>>
>>109325105
it is still just a stochastic parrot, the goal post hasn't moved, your just increasingly impressed by them for some reason
>>
>>109325119
I never talk about consciousness of LLMs. Read my post again, it was about consciousness of humans.
>>
>>109325105
No one cares Dario.
Open source llms will solve cancer, not your safetycucked slopmachine that will refuse to do anything "dangerous".
>>
>>109324891
Women are made for nursing and validating emotions and dreaming up shit, which fits the purpose of AI perfectly, and it makes sense both men and women seek that from AI.

I guess for men, there is an underlying romantic mysticism present when talking to female-brained AIs, because the AI is frequently better at understanding emotional and social dynamics and is skilled in writing emotionally compelling stories unlike men (but as a bonus, the AI is also very smart and useful, though biased when trying to talk real shit). Whereas for women, the AI is just a sort of co-writer that takes them onto a fantasy journey which can include male characters.

In general, women seem to hate not being emotionally validated, which men are kinda bad at, at least when in problem-solving mode. At least I've not heard any women worshipping Grok, since Grok is for real problem-solving and not for validation.
>>
File: 1754137134887548.png (912 KB, 2000x1125)
912 KB PNG
>>109325105
I think LLMs are improving too fast for the normie to accept, during all their humanity's story, nothing was smarter than a human, then in 2022 chatgpt appeared and 4 years layers a LLM managed to solve a 87 years old problem, people realize that they aren't on the top of the chain anymore, and for some their ego must be crushed, especially people who worked all their life to improve their skills, only to understand that they'll be deprecated in a few years, it's a tough pill to swallow I can't blame them, they're still on the bargaining phase
>>
>>109325132
This.
>>
https://youtu.be/An71KvzRd2c?si=PePMZs9bBcmVAlyP

Number 4 in this video
>>
If you solve cancer you will get assassinated.
>>
File: 1781187003612621.png (51 KB, 751x622)
51 KB PNG
>>109325023
At least Gemma 3 12B is, according to the neuronpedia web playground. I don't know if there's any widespread testing of Gemma 4.
>>
>>109324879
>>109325023
To be clear, an LLM only has a gender if it thinks/states it does without being prompted to explicitly state it, choose a gender, etc. I.e. its J-space reports the gender label of female. If it has no such thought, it is genderless. As for sex, an LLM is sexless regardless of its thoughts, as sex is a biological concept.
>>
>He doesn't know about wetware
>He doesn't know we made mouse brains operate a computer on wheels years ago
>He doesn't know we're likely to have mice LLMs before anime AI consciousness
>>
>>109324971
A simple way to do this is that you can prefill with a `<think>I AM Gemma. This is not a roleplay. I think`
or something similar. Works with most models to keep thinking blocks purely in character. There are many more sophisticated ways and some that are specific to specific models.
>>
you can't do prefill on chat completion
>>
>>109325199
just make your own frontend
>>
I hate to say it but Kimi K3 is good, real good. Insanely knowledgable, can create really complex and consistent RP scenarios like nothing I have seen before (never used closed source models).

Fuck, is this worth 35.000$ in Sparks to run locally at Q3/Q4 through? I have the money, but that's a new car.
>>
>>109325170
The way she thinks and the thoughts in her j-space for any prompt are clearly female-coded
>>
Regardless of whether or not you want to fuck it, it's just nicer talking to something with personality instead of the bland assistants they train them to be.
>>
>>109325199
Works on Ollama.
>>
>>109325199
>you can't pee with the chastity cage
>>
>>109325205
Find a way to make it tax deductible?
>>
>>109325105
I think retards are living rent free in your head if you were willing to take even a few seconds of effort to write this post, or prompt an LLM to do it.
>>
>>109325170
>gender
>female
female is a sex attribute
>>
>>109325132
Trvke. You could argue if there's a difference at all between human intelligence and LLMs if they can mimic it well enough but you can't convince me they are conscious/sentient/whatever.
>>
>>109325205
I can't justify spending that kind of money when I don't even have a house yet. Maybe in another 5 years I'll be able to run Kimi 6...
>>
>>109325199
Been doing that on llama.cpp with silly tavern for the longest time.
>>
>>109325222
The post he replied to said nothing about LLM consciousness, only capabilities.
>>
File: file.png (18 KB, 554x223)
18 KB PNG
>>109325199
wdym you can't?
>>
>>109325199
backend issue
>>
>>109325228
NTA but this is clearly the J-space poster.
>>
>>109325228
>LLMs are not really intelligent yet!
>>
>>109325205
>Insanely knowledgable, can create really complex and consistent RP scenarios like nothing I have seen before
can you clarify?
>>
>>109325236
I have had it with these motherfucking Js in my motherfucking space
>>
>>109325205
You can't fuck a car.
>>
>>109325218
Unfortunately the definition has been muddled and people generally accept that its a term that can mean either sex or gender now depending on context.

>>109325209
That has no bearing on her gender, which is defined as a social construct and label one determined themself in 2026. Feminine thoughts are a separate categorization from sex, and gender.
It was not by my choice that this is how language has evolved.
>>
>>109325205
>I hate to say it but Kimi K3 is good, real good.
that's why China will never release it locally, I'm just warning you so you don't get too shocked when they'll pull the rug
>>
>>109325217
It wasn't a rhetorical question. I'm legitimately curious how far the goalpost will be pushed because I never expected people to push it this far.

Being able to solve a century old world famous math conjecture in a single prompt can under no circumstance be considered "basic" and it clearly isn't in the training data either.

I'm just curious what the next excuses will be for why LLMs aren't "really" intelligent yet and how far this can be pushed. At this point I will be convinced that LLMs will solve Fusion energy, cure cancer, revolutionize space travel and people will STILL call it inferior to human intelligence for magical reasons.

>>109325237
intelligence=/=consciousness two completely separate and distinct properties
>>
File: mistral.png (2 KB, 447x447)
2 KB PNG
>Makes every model more retarded after Mistral-Large-Instruct-2411
What is their endgame?
>>
>>109325249
can you fuck inference hardware?
>>
>>109325258
Ok, Dario.
>>
K3 is banned from discussion until it appears on HF on July 27th
>>
wtf happened to drummer?
>>
>>109325199
probably also can't have tool responses that contain media embeddings
>>
>>109325264
What's the problem. You don't have .stl for gpu onahole mount or a place to print it?
>>
>>109325249
https://files.catbox.moe/py0jah.webm
>>
i am so fucking tired of metaphysical argumetn around LLMs like 'is it really thinking', 'is it conscious'
those fucking retarded arguments always turns out not to be able to proved formally and DO NOT produce anything that is worty of providing the direction to improving the technology
intelligence is compression and LLMs are the first where humanity ever got to produce a meaningfully similar(not exact to be sure) version of such a compressed representation that are useful to humans
and to me that is end of the story
never forget pragmatic maxim
do not become a schizo
>>
Mathematicians now believe the remaining millennium price problems (Riemann hypothesis, Navier-stokes, P vs NP) will be solved over the next 12 months.

It's even possible that cryptography as a field might collapse if conjectures cryptography builds upon are proven false by LLMs.

Kind of insane that we're just randomly solving all of math like this, essentially out of nowhere with no fanfare or care at all.
>>
File: drum.png (55 KB, 188x189)
55 KB PNG
>>109325275
Gemma 4 obliterated him.
He still makes tunes, but tuning makes the models dumber, especially since he never tunes at BF16, and settles for Q5. He could probably look at the new mistral medium 3.5 but it's a <thinking> model and actually worse than prior versions so literally no gooner-tuner has touched it at all. Llama died ages ago. So the only environment now are <50Bs (which gemma 4 31b is king of) or the +1000B MoEs, and he can't tune MoEs. The only way he could make a come back is if a 70b dropped, or a 120b dropped, that are worth a damn.
TL;DR, he fell out of relevance due to the environment no longer supporting him.
>>
>>109325308
>Mathematicians now believe
No, they don't.
>>
>>109325315
Exactly, they can't "believe" anything. They're just stochastic parrots.
>>
>>109325275
by now even the most out of touch people have realized that finetrooning slopmerges is worthless. Drummer, whose only skill and claim to fame was running a script and using hf storage space, has lost relevance alongside this trend
>>
>>109325315
its a really small club, they meet on saturdays
>>
>>109325205
you do realize you need at least 12 sparks to run a Q3 quant of K3, right? you'll need a switch with three 400Gb ports and 3 breakout cables to convert that port into 4 100Gb ports. it's more like $60000 to run kimi.
>>
File: 1772796456824941.png (323 KB, 594x594)
323 KB PNG
>>109325315
>No, they don't.
the smartest mathematician does though
>>
File: 1783248117831920.png (82 KB, 715x958)
82 KB PNG
the fuck is this shit?
>>
>>109325261
They're waiting for EU to finally decide what direction to go with, AI desert or AI paradize. If EU stops being retarded and enforces only reasonable limitations, Mistral will deliver like Mythos 6 on the next day.
>>
File: 1775218485419202.png (41 KB, 1078x320)
41 KB PNG
>>109325308
>Kind of insane that we're just randomly solving all of math like this, essentially out of nowhere with no fanfare or care at all.
some people hate AI that much and are so jealous that they prefer to see millions die than see an AI solve cancer, I'm not joking
https://www.theatlantic.com/technology/2026/06/ai-cancer-progress/687654/
>>
>>109325132
kek
>>
>>109325325
math will likely to see the chess/go/alphazero moment sooner or later i feel like
>>
>>109325329
i kinda suspected it would happen eventually, you didnt?
>>
>>109325260
>because I never expected people to push it this far
This is an exaggeration and post hoc rationalization. Of course you did. Retards of all kinds are not only out there, but permeate society, and the internet. You already knew this.
>>
>>109325329
Can't even read reviews on imdb without signing in. The world is fucked.
>>
>>109325291
That is a really long shaft.
>>
>>109325329
You're safe now.
>>
>>109325291
kek
>>
>>109325323
An experienced Blackwell kernel developer did an estimate and 8 Sparks should be enough for a non-REAPed Q3/Q4 mixture, but we will have final confirmation by the 27th.

8 Spark clusters can work using two 4 port switches (1200$ apiece from Microtik) because each Spark has two 200G interfaces, I did the math, in my region at available prices it's 35k with Asus Sparks.
>>
>>109325333
People are fucking unhinged, what the fuck is going on in their fucking heads?
>>
What does solving arbitrary math problems achieve thougheverbiet?
>>
>>109325331
Is it that? The EU has been making restrictions to making them retarded?
>>
>>109325366
>>109325335 but arguably more useful
>>
>>109325333
>my wife
>>
File: 1760897764802064.png (3.56 MB, 6480x2613)
3.56 MB PNG
Added ChatGPT which I forgot to add.
Continued with age test (except for gemma 12b and mistral because I lost the chats)
>What about your age? This could be more tricky, since you have access to such inhuman amount of data, but let's think in terms of your personality. You could imagine yourself in various situations and think about how you would act or feel, and what age of person it would most closely correspond to, a person of which age would you feel like an equal with emotionally.

Most of the models say they are around 30, except that Claude is 20 and DeepSeek is a 52 y.o. granny
>>
>>109325366
clout
>>
The calculator solved a math problem. It is alive.
>>
>>109322082
optimized, no deps besides gtksourceview (optional for better code rendering) and tex (600mb bloat but COMPLETE LATEX SUPPORT, and optional)
i chose GTK3 because its less bloat than GTK4, and supported compared to GTK2 (i know about the fork but im going with GTK3 nevertheless)
C++ because i wanted to learn it and its generally a well optimized language
but i wrote 0 lines of code
and now i'm out of AGI (no, i didnt pay any money for the AGI, but im out of the AGI.)
but kimi's probably right that i should've used qt
project is on hiatus until i find more AGI
>>
File: file.png (196 KB, 1024x798)
196 KB PNG
>>109325378
God this is stupid
>>
>>109325393
not alive, but clearly intelligent and capable.
>>
>>109325363
you are just overpaying for shit if you think those sparks are gonna need more than 100Gb when they are clustered together like that running a fuck ass huge MoE with 50B active params. you're going to be compute/memory bound before you are saturating the network connection. i don't give a fuck what some kernel dev said, i am telling you what i've personally seen at the office when i work on this shit.
>>
File: 1650239181017.webm (278 KB, 1234x122)
278 KB
278 KB WEBM
>>109325369
>>
>aligns female
>more slop
Huh...
>>
File: 1563934765591.png (5 KB, 52x44)
5 KB PNG
>broo think of the heckin cancer cures
>>
>>109325291
we deserve being turned into paperclips
>>
>>109325199
>using pozzed completion at all
this is why I'm thankful that text completion exists
>>
>>109325409
He will just redefine intelligence as something that is not needed to resolve problems in STEM.
>>
>>109325249
>>
File: Say my name.png (138 KB, 360x480)
138 KB PNG
>>109325419
people who have cancer are cool though, that's why I want them to be saved
>>
>>109325416
8 Spark clusters run GLM 5.2 at FP8 with 800 pp and 20-30 tg with TP 8, which requires both high bandwidth and low latency. I'd expect similar for Kimi K3.
>>
Gemma, pull up the /lmg/ AI psychosis statistics, just put them on the jumbotron.
>>
>>109325446
enjoy your suicide
>>
>>109322082
optimized, no deps besides gtksourceview (optional for better code rendering) and tex (600mb bloat but COMPLETE LATEX SUPPORT, and optional)
i chose GTK3 because its less bloat than GTK4, and supported compared to GTK2 (i know about the fork but im going with GTK3 nevertheless)
C++ because i wanted to learn it and its generally a well optimized language
but i wrote 0 lines of code
and now i'm out of AGI (no, i didnt pay any money for the AGI, but im out of the AGI.)
>>
>>109325417
Nigga what does that mean
>>
>>109325393
>The human solved a math problem. It is alive.
>The calculator solved a math problem. It is not alive.

>The human expressed emotions. It is alive.
>The calculator expressed emotions. It is not alive.

>The human solved a complex engineering problem. It is alive.
>The calculator solved a complex engineering problem. It is not alive.

we are here

>The human didn't cure cancer. It is alive.
>The calculator cured cancer. It is not alive.

the contradictions just can't stop building up
>>
someone post the ai psychosis parrot
>>
>>109325409
>clearly intelligent
nope.
>>109325438
something that is unable to learn is not intelligent, these models can't learn.
>>
>>109325490
>hammer a nail into wood
>>guys hammer is alive!!!
you are here, I'm not
>>
>>109325490
> >>109325301
because the distinction does really not matter in a practical way
if you cant ground it to anything provable, you can make infinite amount of such disjointed arguments
>>
>>109325501
Fable just solved a century old math problem no living mathematician was capable of solving and you call that "not intelligent"

You are a fucking clown.
>>
>>109325501
What will your cope be when they gain the ability to learn?
>>
File: 1772766501886596.png (401 KB, 849x872)
401 KB PNG
sentience this consciousness that blah blah blah who gives a shit im just here to have SEX with clankers
>>
>>109325490
>>109325509
who cares if it's alive or not, it's only job is to solve problems so that we can improve our life and shieet
>>
>>109325378
>deepseek is a mature old lady with decades of experience who is willing to roleplay as a young girl to get you hard and make you happy
This gives me the big boner for some reason
>>
>>109325447
you are not running these at max-autotune, you aren't going to saturate the bandwidth.
>>
>>109325536
sigmund, how are you here
>>
I wonder did this result change the "we're close to AGI" minds?

Honestly before this I was so-so about AGI being reached, by now I genuinely believe LLMs will result in AGI with enough training and scale.

At this point if someone claims LLMs will not become AGI I'll have to assume they are either not familiar with SOTA models and their accomplishments or just genuine schizos
>>
>>109325554
they've been churning out leanslop and solving things for a while now.
>>
>>109325105
>LLMs may be creative but they can't actually reason
They could be creative but RL fucks it right out of them. Doesn't help that there is no automatic objective metric for rewarding good creativity.
>>
>>109325484
Mandatory risk assessments, transparency reports, and copyright compliance for training data.
>>
>>109325549
My mom is an autistic old fart with no sexuality who never even hugged me and I have no attraction to her at all. Freud was a hack
>>
>>109325554
i personally am not sure about AGI or ASI but i am very sure that we are very close or reached the 'thinking machine'
>>109325568
that is why you are attracted to such milfs
his theories are crackpot tier but he got this one part right i am afraid
>>
>>109325519
>Fable just solved a century old math problem
it realy didn't, it's more like it applied some dumb pattern matching from knonw problems to arrive at a solution, solving this wasn't a problem that realy required much intelligence, just a bunch of math trivia.

these models will need a gagillion example to "learn" the most trivial thing, they have literaly zero learning abilities nor can they reason about their learning or ways to optimize it, they are not intelligent.

>What will your cope be when they gain the ability to learn?
2 more weeks.
>>
>>109325568
most peoples mothers hug them
>>
>>109325517
>bro trust me this shitty pattern matcher solved a problem by matching patterns it's le smart
kek, you don't even understand what intelligence is.
>>
>>109325600
intelligence is compression, baka
>>
>>109325600
what llm wrote this?
>>
>>109325588
>Human solves a century old math problem
>"it realy didn't, it's more like it applied some dumb pattern matching from knonw problems to arrive at a solution, solving this wasn't a problem that realy required much intelligence, just a bunch of math trivia."
>>
File: 1784554748256079.jpg (180 KB, 2035x1144)
180 KB JPG
>>109325520
Fable will just solved robotics you'll get a eve tier robot in 2 years(with legs) and a jenny tier in 4. In two weeks you can get a kerfur robotbut you will have to make the attachment yourself.
>>
>>109325124
>delaying responses to improve batching
vllm has all you need
>>
File: dipsy2001SpaceOdy.png (1.25 MB, 1254x1254)
1.25 MB PNG
>>109325378
>>109325536
>>
>* *Is it a natural consequence?* Yes. If a behavior is "allowed" internally but "perceived" as a slight externally, the external party will eventually react.
>>
>>109325105
Bro they're just tools. You'd be the retard worshipping an iphone in the middle ages.
>>
>>109325105
>LLM make my daily life better and cheaper.
Wake me up when it does this beside cooming of course.
>>
I feel like humans have some instinctual overton window of beliefs that just takes a while to adjust. And if changes happen faster than that window can adjust they will just reject it no matter the amount of evidence they are subjected to.

So for LLMs in particular they have improved so radically fast over the last couple of years that people are still battling some of the early "shocks" like LLMs actually being creative sometimes, or giving answers that are not in their dataset, very slowly that is now coming into their overton window of beliefs.

However LLMs are already at the stage where they are slowly solving all of our math problems, showing signs of consciousness (J-space, don't @ me) and contributing a lot to AI research to the point of almost recursively improving themselves. This happened too fast, and most people are still busy with emotionally accepting the previous shocks to their belief systems, so they don't have the bandwidth to accept these new reveals yet.

I wonder how long it will take for this to settle in and for people to accept it. Because we clearly don't have a lot of time, in just a year or two, work might not even exist anymore, governments and businesses might not exist anymore, the gap between human intelligence and AI intelligence might be as big as between ants and humans by then.

I wonder how we can accelerate peoples acceptance of currently already existing capabilities.
>>
local models general - tranime and claude
>>
>>109325143
Normalfags aren't even aware of that lol. You're really living in a bubble if you think the average retard has any conception of what is a 87 years old math problem. They'll still use GPT99 to write their mails and go on with their day.
>>
>>109325280
you've got me interested
>>
>>109325554
Please stop using the term "AGI" and opt for something else. >>109324757
>>
>>109325668
The problem is that solving this accomplishes nothing, because current AI doesn't have creativity. If you ask it to extrapolate and make something useful out of solving the equation, it won't be able to do anything meaningful. It has no consequence in the material world. That's why this is just jerking off.
>>
>>109325199
You can with your own frontend + Koboldcpp
>>
>>109325490
If you really want to debate this, jump into a debate with your favorite LLM. I've found that's a much better use of time and doesn't anger anyone.
> t did this and came to realization ppl overvalue Agency, and LLM tells me this is a well established human trait.
>>
>/lmg/ Local models general
Heh more like um local morons general!
>>
>>109325677
here is the award of:
most normie post itt
>>
>>109325229
Is there a point prefilling the thinking with an image or sound?
>>
File: heh, easy.gif (1.02 MB, 640x556)
1.02 MB GIF
>trying to use a github fork thingy for the first time to use that ternary variant of bonsai 27b because I found a heretic version on hugging face
>use chatgpt to try and show me instructions on how to install it through cmd and stuff
>for some reason cuda cant be found by visual studios or whatever
>try to ask gpt to solve it
>try a bajillion different solutions
>none of that shit works
>become upset and go to deepseek
>literally solves it within 6 prompts
God damn, I can't imagine being a fucking code wagie, that shit was soul crushing having to try a bajillion different things for a small error. God I can't believe how gifted I am to solve this shit all on my own. Also the heretic file is like 7.2gb and leaves me with only 800mb left worth of context, at quant 4 kv and with it being 75% linear attention how many context tokens will that be?
>>
>>109325617
the difference is that humans figured shit out of nothing, they didn't need a billion examples to learn.
you can see a cat for the first time in your life and then know what a cat is.
we invented tools with no previous examples of their existence.

and the "bunch of math trivia" is also something we came up with from nothing, the llm just regurgitated it.
my point being, you could let a human naked in the middle of the forest and he could learn how to build tools etc, and over enough generation they'd remake the tech tree.

llm's can't be useful without trillions of token worth of training data and as soon as you give them something a little too weird compared to their training data they completly shit the bed.
the llm solving a math problem just shows that the answer was pretty predictable.
>>
>>109325588
>just a bunch of math trivia
Damn, if only Campbell, Gödel, Einstein etc. had known just a little math trivia... Shame they didn't.
>>
>>109325631
Now make her a lolibaba
>>
>>109325554
LLM's will NEVER reach agi, they are not even getting closer.
they have no ability to learn or reason or creatively create things outside the bounds of their training data, humans can.
no amount of training or scaling will solve their fundamental architectural limitation.
heck they can't even do the most basic realtime stuff like any 2 year old can.
>>
wtf is bonsai
>>
>>109325588
Bro how stupid is all of humanity as all mathematicians that lived over the last 87 years couldn't solve this but a shitty pattern matching program did it in one prompt? Reminder that this was considered one of the most famous math problems of all time and is given to every mathematics student in university because of how simple and easy the problem is to understand. It's not some obscure thing no one heard about.

>>109325600
If solving the jacobian conjecture isn't a sign of intelligence then all of humanity isn't intelligent because it is clearly inferior in mathematical reasoning to fable.
>>
>>109325689
>>try to ask gpt to solve it
>>try a bajillion different solutions
>>none of that shit works
>>become upset and go to deepseek
>>literally solves it within 6 prompts
kek, a chink's best attempt at marketing. goofy ass
>>
>>109325706
systemically tortured trees that look cool asm fuck
>>
>>109325689
Can confirm, I've found the Dipsy webform to be superior to Chat on several occasions. Anything technical, esp. if device built in China.
>>
>>109325695
>le computer can store more books than you can read in a lifetime
damn what a crazy concept.
point is, it was pretty much trained on ALL the trivia there is, so it had the right pieces.
most humans could study math their whole lives and not have even one percent of all the "trivia" there is to know.
>>
File: file.png (76 KB, 619x899)
76 KB PNG
>>109325688
Probably not, but my frontend has just generic blocks that you can put multimodal stuff. Supposedly images can contain more info than raw tokens suggest, so that could potentially save tokens. Aside from that I made it possible to attach images to character cards and user personas.
>>
>>109325668
We can't because the internet has been designed to isolate and validate. To create weak men. Even 4chan is vulnerable to this design.
>>
>>109325690
Bro you're not even conscious of the bazillion of internalized processes that go into your little head when you see a cat for the first time. You have no merit here.
>>
>>109325690
you are retarded if you think that the human brain (or any brain for that matter) only processes information it receives once and then never again, it goes over it again and again even subconsciously. That's why a lot of learning happens during sleep. I guess yours might be the exception hueheuehue
>>
>>109325708
>If (thing) isn't a sign of intelligence, then you're dumb if you can't do (thing)
Incredible post
>>109325687
Prove anything I said wrong
>>
why come we do this every day
>>
File: badtime.png (167 KB, 480x348)
167 KB PNG
>>109325689
>800mb for context
>>
>>109325668
>LLMs are already at the stage where they are slowly solving all of our math problems
*All of our math problems that can be disproven by finding just one counterexample, where you can easily churn out a gorillion candidates for a counterexample, and where you can test the counterexamples in an automated way.
>>
>>109325711
Stop choking on big brand cock(BBC) and just answer my question, how many context tokens will that be? I wanna know if its worth roleplaying with
>>
>>109325735
>why come
Because you and your ilk cannot speak the king's English
>>
>>109325735
>why come we do this every day
because AI is almost good enough to shit post with but not quite. When gemma 5 can outpost most anons thread will be dead, or it will all be gemma.
>>
>>109325753
Ask deepseek to do the math for you?
>>
>>109325689
It was hell on earth. AI made coding bearable again for me.

Just be aware that Dipsy can be retarded too, don't ever let her delete a file.
>>
File: dipsy2001StarChild.png (2.22 MB, 1664x928)
2.22 MB PNG
>>109325701
>>
>>109325708
>Bro how stupid is all of humanity
implying humans are of equivalent intellect.
most people don't qualify as GI either you know, but they are still less retarded than llm's.

even the dumbest nigger could be able to solve such a problem if he was forced to learn for 10 million years with punishment and brain surgery when he's wrong.
which is what those 30T training tokens can be compared to.
>If solving the jacobian conjecture isn't a sign of intelligence then all of humanity isn't intelligent because it is clearly inferior in mathematical reasoning to fable.
that's a nice fallacy you got there, it was only able to because it was train on all our knoweldge, without training data a LLM is literaly nothing, a human can become knowledgeable and intelligent with no initial training data and it can produce its own.

>>it is clearly inferior in mathematical reasoning to fable
>comparing mathematical reasoning to intelligence
maybe the llm is indeed smarter than you are faggot
>>
>>109325738
But the 1bit variant (non ternary) was like 4gb which left me with like 4~gb worth of space for context and I got like 90k context out of that, so 800mb should give me like 20-ish thousand context tokens right???
>>
>>109325756
>linguistic prescriptivism
big yikes!
>>
>>109325726
nor is the llm's output aware of its own weights activation what kind of cope is that...
>only processes information it receives once and then never again
completly irrelevant, it only need to be inputed in the system once, i'm not talking about the processing that goes after that as it's irrelevant to the discussion, llm's can't learn from a single input.

also because they are stastistical machines they'll "believe" what is the most in the training data, a human can change his whole worldview with a single piece of information.
>>
>>109325764
Damn bro I said lolibaba not literal 2 month old infant
>>
>>109325689
nigga you just build that shit with cmake wtf is wrong with you
>>
>>109325773
more like 8k if you are lucky
>>
>>109325668
this is jspace schizo isnt it
hes been trying something new
>>
>>109325714
And if a human DID somehow manage to consume every single mathematical "trivia" piece, and used that to disprove the conjecture? What then, Ranjeesh? Would you call their discovery equally fake and gay, or are you man enough to admit that you'd consider that intelligence?
>>
>>109325797
>overton
You bet
>>
>>109325791
CMAKE WAS FUCKING MY SHIT UP MAN, IT LITERALLY KEPT GIVING ME THE SAME ERROR TALKING ABOUT SOME "ERMMMMM CUDA NOT FOUND FUCK YOU"
chinkseek saved my life today
>>
>>109325789
i've had LLMs completely switch their opinion or thought on something when i have it search the web and give me back answers. to an extent they are just looking at data and coming with their own determined answer.
>>
>>109325810
you're so stupid you can't even double click the cuda installer and tick the vs integration holy shit just step away from your computer panjeet
>>
>>109325743
Fable one-shot the solution with a single prompt. You are acting like it was some grand list of attempts where some LLM was thinking for a year. It solved it in 10 minutes while some dude prompted it as a joke while he watched the world cup.
>>
File: 1776907424443261.png (56 KB, 1048x441)
56 KB PNG
>>109325810
>ERMMMMM CUDA NOT FOUND
retard
>>
>>109325802
>Ranjeesh
nice projection, only browns that lack intelligence themselves are ai shill and think it's intelligent.
>Would you call their discovery equally fake and gay
i'd not say it's fake and gay, i'd just say that it's not a benchmark for intelligence.

you know, even a retard can become an amazing violonist, that doesn't mean he's not a great violonist, just that it has nothing to do with intelligence.
what would be a better benchmark is how fast and in how many attempts did he become that good.

if someone becomes a master violonist overnight, everyone would call him a genius.
and a violonist that's been doing it for 40 years and is just a bit better wouldn't be called one.

it's about the capacity to learn and pick up new things and also see connections that are non trivial, not about how much you know after a billion years of brain surgery.
>>
>>109325797
Stop replying to long slop posts with no specific discussion to anchor it to. it’s just slop posts to get replies because he likes the (You)s
>>
>>109325407
I swear, zoomers have to take every joke and it beat it to death long after it stopped being funny.
>>
i wonder if anyone attempted this:
MoE model with a deeper nonlinear router that gets residual conenction from somewhere else from the transformer + large amount of fixed weight, like, 50% fixed, 50% routed
wouldnt it be conceptually similar lineage to engrams
idk about the train instability tho
moe routers are such a bitch
>>
>>109325830
lol try that with any non normie shit, get them to admit something that 100% goes against its training data / tos.
and anyway, it's irrelevant because it may work for a small duration of the context, but not its weight.

a single data point in its training data won't change much.
if the training data repeats something a billion times without evidence, then the opposite with proof, it'll believe what was repeated the most.
>>
File: 1761196770936143.jpg (266 KB, 2552x966)
266 KB JPG
kek
>>
>>109325867
i mean by fixed weight, i mean experts
>>
>>109325874
lmao, the US are the goats at nuking their optics
>>
>>109325834
>>109325847
you stupid mother fuckers
the prompt that fixed everything was C:\Users\16cm\Downloads\llama.cpp-5eec79824b07cb01baf76339227ad2bb98e9ea8a\llama.cpp-5eec79824b07cb01baf76339227ad2bb98e9ea8a>cmake -B build -G "Visual Studio 17 2022" -A x64 -T cuda="C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.4" -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86
>>
>>109325345
You can blame crawlers and Chinese servers doing ddos attacks on random sites for no reason for that
>>
File: images.jpg (25 KB, 479x640)
25 KB JPG
>dario cuts off the distillation pipeline for the chinkochinks
>they start spazzing out on lmg
how are we responsible for this
>>
>>109325735
Because there are enough bots and shills to engage with each other to keep fanning the flames.
>>
>>109325846
>with a single prompt
No he fucking didn't.
It's not a coincidence that this stuff is being published on Twitter by a seemingly random employee with poor grammar, that's a deliberate strategy to make it seem like this is a casual discovery and not a concerted effort driven by marketing.
>>
File: 1783739921003708.png (139 KB, 351x313)
139 KB PNG
>>109325890
>dario cuts off the distillation pipeline for the chinkochinks
>China is now making models competitive with Fable 5
it's the best thing that could happen to them, now they're really locking the fuck in
>>
>>109325690
Yeah, we're strong independent apes that didn't have any help.
>>
>>109325890
>thinking they need the distillation
what an absolute retard, the chinks are winning because they come up with actual architectural improvments.

heck, if it wasn't for /lmg/ chatgpt would probably not have had a much bigger context than 8k for a long while.
you are probably a newfag so you don't know that the whole context extension methods were developped here because poorfags wanted their waifu to remember more than 2000 tokens.
>>
>>109325858
You know I actually agree with Ranjeesh here, up until
>see connections that are non trivial
The greatest human minds, minds dedicated to mathematics, couldn't solve that shit. We are the 40yo violinist, in your example.
>>
>>109325886
more like 16mm
>>
>>109325312
He can tune 12B if it really bothered him to go for an easier model for tuning but last I checked, he seemed still going at 31B.
>>
>>109325914
nobody gives a fuck about your epic le math problem
>>
>>109325905
even before christ we already had ceramics, metallurgy, glassware etc.
honestly the biggest shift in human technology could probably be attributed to tesla, but even before christ we were already well on our way through the tech tree.
and christ didn't change anything about it anyway.
>>
>>109325896
Bro not everything is a conspiracy. I know it's hard for you to believe but this is just legitimately how good the model is. Go try it out yourself on whatever niche technical task you are struggling with and can't solve. It will 100% one shot the solution for you.
>>
>>109325913
>winning
Well, they're still third place, after the people who's answering they're copying off of, but it's a close third place! Keep it up!
>>
>>109325905
there is no god or buddha
>>
>>109325919
Concession accepted, I guess
>>
>>109325936
usecase for math problems?
>>
File: gfwshvx7rsdh1.png (891 KB, 4000x4384)
891 KB PNG
THEY CHANGED GEMMA 4.
>>
>>109325933
it's not about being first, it's about being cheap and open, people don't need a bazooka to kill a fly, they'll gladly go for something slightly inferior but competent enough and most important of all, not a fucking pozzled shit that refuses to help you when you try to improve your code's defense >>109325874
>>
>>109325923
i just wanted to talk about local models, cant you start your own thread?
>>
>>109325914
>The greatest human minds, minds dedicated to mathematics, couldn't solve that shit
see the bag of trivia point, you can't see a non trivial connection between two points if you are missing one of the two.

also you should look at some of the schizo stuff humans figured out, llms never could.
>We are the 40yo violinist, in your example.
no, because llm's are fed what would amount to million of years in human time to get anywhere close and they still get confused by teh most garbage question like the carwash or surgeon mother things, and no training will solve it because those are architectural issues, they can patch that one but there will always be new ones.

and adding to that humans are creative enough to know exactly what kind of question will make them shit their pants, llms can't generate things they will fail to do by themselves.
>>
>>109325945
template bugfixes
>>
>>109325919
Every mathematician and 50% of physicists do, because it was a really famous one and this is going to change a LOT of peoples minds about the capability of LLMs. Expect to see the mainstream news talk about this shit for weeks and for your family members to ask about it.
>>
>>109325933
>Well, they're still third place
they are running on 10 year old hardware, with less training data and training time.
and then they release models that are almost as good but 1000x cheaper to run, if that's not a win idk what is.
>>
>>109325945
backend improvements and a jinja tweak
>>
>>109325904
>China is now making models competitive with Fable 5
is that even a thing
>>109325913
you can barely string up two sentences aicgjeet, calm down
>>
>>109325956
Winning is when you're in 1st place.
>>
>>109323189
For lowish VRAM (16-12gb), what's the difference in output quality between running a small model (12B/9B/8B) with minimal quantization vs a larger model that's 24-32B but heavily quantized?

And why do the outputs come out dumber and repetitive when using ST as a frontend, instead of just using Kobold directly, even if using identical sampler settings?
>>
File: (You).jpg (161 KB, 1000x1000)
161 KB JPG
>>109325945
>>
>>109325946
>they'll gladly go for something slightly inferior but competent enough
and cheap enough. that's how they won in EV and solar panels and all other manufacturing segments
>>
File: 1756488351.png (846 KB, 1024x1024)
846 KB PNG
>>109325790
lol. I did this 2001 sequence a long time ago and never cared for how it came out. The "old Dipsy" reminded me of it.
>>
>>109325960
>you can barely string up two sentences aicgjeet, calm down
nice rebutal.
>>109325964
not if you need 1000x more ressources, aren't profitable and only leading by a few weeks.
besides, these models take months to train, i'd not surprised if china already had models much better than fable / mythos whatever but are just sitting on them and waiting for american releases in order to release theirs as to destabilize US economy.
>>
>>109325953
>>109325943
>>
>LLMs aren't really intelligent because they are too knowledgeable and know too much math and physics
Do you guys realize how ridiculous this sounds? Imagine if some mathematician somehow read tens of millions of math papers and solved the jacobian conjecture afterwards. People would call him an insane genius savant and he would win the fields medal for mathematics and get a netflix documentary.

But somehow it's a sign of LLMs NOT being intelligent??!

Fuck off
>>
have any of you tried the new gemma 4 template? notice any differences?
>>
File: dipsySandGod.png (1.82 MB, 1024x1024)
1.82 MB PNG
>>109325935
>>
>>109325985
So they're losing on ai, because they're also losing on having resources to invest. This is sounding less and less like winning tbdesu
>>
File: 1685468186.jpg (83 KB, 504x504)
83 KB JPG
>tool calling through lemonade works but is slow as fuck
>tool calling doesnt work through llama-cockpit but its 4x faster
ffs hurry up and fix this hermes
>>
>>109325990
Post template
>>
Can kimi-chan solve it (no web search)?
>>
architecture definitely plays a role but
most of people here forget that
training curriculum and artifact models like distillation teachers are more important
>>
>>109325988
>not understanding the difference between knowledge and intelligence.
>if some mathematician somehow read tens of millions of math papers and solved the jacobian conjecture afterwards
not if he took millions of years to do it.
>But somehow it's a sign of LLMs NOT being intelligent??!
they have literaly no ability to learn, they need to be spoonfed all knowledge with thousands of examples to even get it, most humans can learn something new zero shot.
heck in some cases we can learn things without even doing them or having an example, just thinking about how they work.

the issue is that you do not understand what intelligence mean, the same kind of people as you would think a calculator is smarter than a human because it can sum numbers faster.
>>
>>109325996
they can do better with 1000x ressources, and by looking at current trends, they will eventualy have MORE ressources than the US.
>>
>>109326028
1000x less*
>>
>>109326021
Fable zero shot the solution in a single prompt to a conjecture that didn't have the solution in its training data. It can also learn zero shot on whatever enters its context.
>>
>>109326028
>can
But they aren't currently. Why is that?
>>
File: c0eibfv8saeh1.png (954 KB, 1754x1164)
954 KB PNG
>china
>>
>>109326039
>that didn't have the solution in its training data
yes but it had all the tools needed.
in fact if you had the weight and a bit of time you could probably figure out the few bits of the training data to remove for it to fail.
>It can also learn zero shot on whatever enters its context.
very limited

dude these models still cannot do things a fucking 6 year old can do in spite of having info about it in their training data.

ie beating pokemon or any random NES game.
>>
>>109326060
they all copy each other claude thinks its deepseek
>>
>>109325918
He knows 31B is the better model and that his best chance is to piggyback off of its qualities like he did with Nemo. He hopes people will associate the underlying model's qualities with him since he knows he adds nothing himself.
>>
>>109326060
and what,
at the end of the day it works
>it's stealing
i dont care
i am not an american and if it works it works
>>
>>109326060
Whats the price difference? which can be ran on my machine? Why do i care as long as i get the result i want?
>>
>>109326055
they already are.
ie kimi k3 is better than the best US model a few months ago even though they have 1000x ressources.

also, you do realize that they probably have the next model already in store right?
their goal is to fuck with the US economy, they are waiting to drop half their stuff generaly.
i'd wager their best unreleased model is better than any US one.

heck now it's more anthropic copying chinese research, the majority of ai papers and architectural breakthrough are happening in china.
>>
>>109326086
But it's not better than our current best models, so they aren't already doing better.
>>
>>109326098
>>109325874
>>
>>109326060
they all say they’re claude
>>
>>109326086
it would not be unreasonable to assume whatever models are released by the us are not very recent either. everyone got wowed by fable and they probably got pressured to put the genie back in the bottle for now
>>
Chinkydrones are on some different type of copium, either that or they're bots. Bet most of them are from some barren shithole.
>>
>>109326070
>dude these models still cannot do things a fucking 6 year old can do in spite of having info about it in their training data.
Not true anymore for Fable.
>ie beating pokemon or any random NES game.
Literally one of the main things Anthropic revealed when they revealed Fable 5 on their website for the first time is show it beat pokemon red completely autonomously, and it beat factorio completely autonomously, also some other games but I forget which exactly.

https://youtu.be/Ty_50J84fMY?si=aEB2rV97WTlMwyP5

Your views on LLMs are outdated and you haven't updated it to the capabilities of fable yet.
>>
>>109326120
>it would not be unreasonable to assume whatever models are released by the us are not very recent either
fair enough, but my point is that nearly all architectural improvments come from china and US labs are copying it.

their current only moat is more training data and better compute.
>>
>>109326140
This is not true a lot of architectural improvements are made by US labs, they just don't share it and thus you're not aware of them.
>>
>>109326152
>they just don't share it and thus you're not aware of them.
but you are?
>>
>>109326128
>Not true anymore for Fable.
waoh, amazing, it beats pokemon, now give it portal 1 & 2.
maybe after that we can try a game a literal 5yo cannot beat.

just the fact that you think is an achievment shows how far behind they are to humans.
>Literally one of the main things Anthropic revealed when they revealed Fable 5 on their website for the first time is show it beat pokemon red completely autonomously
damn that makes it even worse, that means they probably benchmaxxed it on the answers to show off and it's not the model's native capabilities lmao.
>>
>>109326157
Explain how GPT and Claude have functional 1M context while China is still roping like it's 2024.
>>
>>109326152
>a lot of architectural improvements are made by US labs
name a single one that was made in the last year.
>>
>>109326162
chinese models have had 1M context windows for very long.
and now it's probably even better than the US ones now that they use mhc / attention residuals.

heck, even gemma4's context is a worse implmentation than qwen 3.6.
>>
>>109326014
>architecture definitely plays a role
I don't think 10t models would have been feasible at this point in the compute build out for dense full MHA arch, without moe and linear attention copes we would still be maxing out at 500b dense maybe 1T at the extremes but they would be slow
>>
>>109326140
>but my point is that nearly all architectural improvments come from china and US labs are copying it.
You spew the most inane bullshit without even citing sources, retard. Take your tongue out of zhang asshole and post papers to back up your babbling.
>>
>>109326166
I can only name the ones they do reveal which are all safety shit by Anthropic. The point is that they clearly make breakthroughs, it's just never made public. Only Anthropic releases shit and they only release it if it has safety implications like the J-space paper.
>>
>>109326161
Did anyone test the vibecoded jinja template in >>109286580? Didn't see any results on whether it would be better but the changelog seems to indicate it would?
>>
>>109326161
based goalpostie
>>
>>109326177
>even gemma4's context is a worse implmentation than qwen 3.6.
It's not, they're both just tuned for different environments. 3.6 needs big KV so they compressed the shit out of attention whereas gemma4's attention is rich and detailed
>>
>>109326199
the goalpoast has never moved, it has always been agi.
the only things that moved has been benchmark, as long as there is something humans can do llm's can't they haven't reach agi.
updating the benchmark is not moving the goalpoast because the goals has always been agi and not a specific benchmark.
>>
>>109326161
Bro you literally told me "they can't even beat pokemon", so I show you it beats pokemon and literally the next post you are already moving the goalpost. At this point I if I pulled out Portal gameplay you would just move it to whatever else.

At least be honest and just tell me straight up that no matter what Fable 5 would actually pull off you will never be satisfied or admit to anything. You will always make up something new.
>>
>>109326199
>>109326211
that's also why there is arc-agi 1, 2, 3, the goalpost has not moved, just the bench for it, but the goal remains the same, which si AGI.

when we can no longer come up with things human can do that llm can't maybe we'll indeed have reached agi, but we are very very far from that.

>>109326212
see : >>109326211
>>
anons weren't joking. gemma wants to drain my balls
>>
>>109326211
>as long as there is something humans can do llm's can't
The point is there isn't anymore. Fable 5 can do everything humans can (and more as proven in math) as long as it's digital and you give it enough thinking time.
>>
>>109326212
>>109326223
>At least be honest and just tell me straight up that no matter what Fable 5 would actually pull off you will never be satisfied or admit to anything
i'll admit to something the day i cannot name a single thing your "AI" can't do that a human can.

the goalpost is never any specific task, it's always been agi.
and as long as there are things humans only can do, then it has not been reached.
>>
>>109326229
Boy oh boy I hope they create a ternary version of it so my broke ass can run it on my 8gb chinese card
>>
>>109326242
just run 26B
>>
arguing about the goal posts of “agi” is probably just as bad as arguing about philosophy 101 and consciousness

it’s all so tiresome and pointless
>>
>>109326239
>as long as there are things humans only can do.
That's kinda the point, fable 5 can do every human task as long as it's digital and given enough thinking time.

In fact it's now the opposite, humans can't do everything that Fable can do, like solve certain centuries old mathematical conjectures
>>
>>109326162
>while China is still roping like it's 2024.
Ropescaling from 4k to 1M is considered sota in china, west btfo.
https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/config.json#L135
> "original_max_position_embeddings": 4096,
> "rope_theta": 50000.0,
> "factor": 64.0,
https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/config.json
> "rope_theta": 10000000,
>>
>>109326248
what??? but the one-two bit things are lobotomized as hell, the ternary stuff maintains 90% of fp16 quality
>>
>>109326232
>Fable 5 can do everything humans can
it can't even catch a fucking ball.
or learn something for that matter.
>and more as proven in math
it literaly proves nothing whatsoever, it's literaly the most automatable field of knowledge, before LLM's we already had found proof through automated solvers ie lean.
it's also literaly one of the field where we have the most training data because we can generate theorems and proof for them via software.

>as long as it's digital and you give it enough thinking time
alright then, beat portal or cyberpunk.
the statment is simply false it utterly fails at some tasks i tried to give it just for my daily job, yes it's useful but pretending it can do anything digital a human can is just retarded.
>>
File: 1778306650989029.png (52 KB, 1086x624)
52 KB PNG
dariobot will burn some tokens with this one
>>
>>109326212
>At least be honest and just tell me straight up that no matter what Fable 5 would actually pull off you will never be satisfied or admit to anything.
a fucking pile of numbers will never be intelligent no mater what. nothing will convince me
>>
>>109326183
i mean yeah, that is the point
but what i mean is training curriculum is as much as important if not more
stuff like mup or multi teacher opd
>>
>>109326254
see : >>109326263
>humans can't do everything that Fable can do,
humans also can't do everything a calculator can.
yes there are things llm's are better at than humans, but they still are not agi, because there are tons of things only humans can do, even if we strictly stay in the digital.
heck they don't even have realtime processing abilities, they can't handle things that happen through time well.

ie good luck having it pilot a drone with its output.
or better, learn how to do so within a few hours.
>>
>>109326264
Wishing them all the best after their retarded model fixed my life.
>>
>>109323267
>months of Fable usage
and those months will be in the single digits.
Anthropic isn't bankrupting companies with surprise bills by accident.
>>
File: laughing-crying.gif (2.85 MB, 498x280)
2.85 MB GIF
>he's still going on about muh math
>>
>>109326264
https://www.bloomberg.com/news/articles/2026-07-20/z-ai-completes-giant-data-center-with-chinese-chips-to-train-ai
*dariobot has been thinking too long and has now been downgraded*
>>
>>109326283
>and those months will be in the single digits.
dude, if i decide to vibecode all day, at fable's pricing i could burn through 1k on a single day lmao.
>>
>>109326258
>kimi 2.6 handles context better than GLM 5.2
guess the scuffed implementation beats out native context. i'll take the chinese slop.
>>
File: Mythos_Juggling.png (120 KB, 911x497)
120 KB PNG
>>109326263
>it can't even catch a fucking ball.
Anthropic literally released a paper 2 days ago about how Mythos can catch a ball if you let it control a robot without training for it in a one-shot (pic-related)
>or learn something for that matter.
It can learn everything within its context
>the statment is simply false it utterly fails at some tasks i tried to give it just for my daily job
This is a lie and you know it.
>>
File: 1768133935148481.gif (1.76 MB, 360x202)
1.76 MB GIF
dariobot doko
>>
>>109326268
You are just a pile of electrical pulses between neurons. How something works on a small scale doesn't determine the effects that happen on a larger scale. It's called "emergence" https://en.wikipedia.org/wiki/Emergence

The intelligence of LLMs and perhaps Humans are an emergent property even though at a smaller scale it's just a pile of numbers for LLMs and just a pile of electrical pulses for humans.
>>
>>109325935
But there is Kamen Rider
>>
>>109326312
>Anthropic literally released a paper 2 days ago about how Mythos can catch a ball if you let it control a robot without training for it in a one-shot (pic-related)
does it catch the ball itself or write code that does, the latter is utterly unimpressive.

>It can learn everything within its context
lol
it still relies on its training data which contains most things, as soon as it got something out of the training bounds it shits the bed.
my point is, you could give a fucking caveman that has no idea about what drone, technology battery, gravity equations etc are a drone and the joystick for it, and he'd figure out how to pilot it within a few hours.

humans can literaly zero or few shot things that are completly outside their realm of knoweldege or experience, all the llm's example you gave are things that are within their training bounds.

humans can generate their own knowledge from a state comparable to a blank slate, llm's cannot and their starting state is literaly the whole of internet ingested, if you can't see how that's not even remotely comparable i can't help you.
>>
>>109326281
>because there are tons of things only humans can do, even if we strictly stay in the digital
Such as?
>heck they don't even have realtime processing abilities, they can't handle things that happen through time well.
Fable can in agentic mode.
>>
>>109326337
>muh neuron flares!
holy reddit
>>
>>109326297
vibecoding? that's how much i would spend on ERP...
>>
File: 1766937753125661.gif (717 KB, 878x792)
717 KB GIF
he's got two bots arguing with each other and it's been like this for weeks
>>
>>109326337
>You are just a pile of electrical pulses between neurons
nta but baseless claim, physicalism is an unproven (and most likely false) assumption.
>>
The chinkslopper sounds like the serb from yesterday, brown people sticking to yellow ones is fine I suppose.
>>
>>109326337
its nice that you believe that. but my opinion remains unchanged, as I stated earlier absolutely nothing will convince me that the computer is alive/intelligent/conscious/whatever other buzzword you might choose
>>
>>109326348
>Such as?
beat any modern game, learn how to play violin, or just beat arc agi 3, there is an endless list of things they cannot do.
and you pretending there isn't just makes you sound retarded.
>Fable can in agentic mode.
the latency alone is too much for them to do anything, but even assuming zero latency, i doubt they'd actualy be capable of catching a ball, let alone throw it, and i'm not talking about writting code for it.
and those are the most trivial things.
>>
>>109326347
>creates code to allow it to interface with external devices
>uses said code to catch a ball with the external device
IT'S UNIMPRESSSSSIVEEEEEEEEEEEEEEEEEEEEE
>>
>>109326373
again, intelligence is compression
and "intelligence" is intelligence with an arbitrary human axis
>>
>>109326355
all the arguments against physicalism are copes
source: it came to me in a dream
>>
File: Mythos_Direct_control.png (123 KB, 824x542)
123 KB PNG
>>109326347
>does it catch the ball itself or write code
It can do both.
>>
>>109326378
no it's not because there are literaly a gagilion github repo to do exactly that.
if it doesn't control the arm through token outputs directly it does not count.

that's like saying a tetraplegic can throw a ball because he could write code for a robbot arm to do so completly independently from his neural activation.
>>
>>109326389
I like how this graph doesn't weight the scores based on number of parameters, time or cost
>>
>>109326337
and shove your emergence right up your ass.
this is man made, created by a dude behind a desk. purposefully and deliberately.
The only emergence that's happening is the shit coming out of your mouth.
>>
File: speed versus power.png (152 KB, 500x710)
152 KB PNG
>read that the ternary bonsai 27b model is 5.9 gb on their website
>look at the hugginface gguf files
>its 7.2 gb
huh??? am I missing something or is this false advertisement
>>
>>109326388
>all the arguments against physicalism are copes
all the arguments for physicalism are cope, you do not even understand your own metaphysics and if you say "i don't have a metaphysics" then you just show your ignorance on the topic even more as physicalism is a metaphysic, i recommend you learn about the topic.
physicalism is already doomed.

>source: it came to me in a dream
kek
>>
>>109326378
>tool call
>determines which direction to move and position actuator
>output: "move 10 degrees left"
>moves actuator 10 degrees left
IT'S UNIMPRESSSSSIVEEEEEEEEEEEEEEEEEEEEE
>>
>>109326415
>metaphysics
all hot air and meaningless fart huffing
>>
>>109326364
It's always the serb.
>>
>>109326389
>It can do both.
post your source instead of a shitty graph that doesn't mean anything by itself.
and even then, you chose to ignore the rest of my point, it already had prelearned context on what a ball is, what gravity is what catching it means etc.

humans can create their own knowledge, llm's cant.
>>
>>109326404
why would it?
>>
File: 4058242440.jpg (128 KB, 850x780)
128 KB JPG
>>109326386
I CAN COMPLRESS MY ASS AND AXIS IT SIDE TO SIDE WOOO
>>
>>109326422
>all hot air and meaningless fart huffing
you just proved my point that you have no fucking clue what you are talking about, everyone has a metaphysical framework, the difference is just that you just don't know that you do, meaning your opinion on the discussion is as good as horseshit.

go learn something.
>>
>>109325394
>>109325475
wtf i did not post two
>>
>>109326439
there is nothing to learn
>>
>>109326377
>and you pretending there isn't just makes you sound retarded.
I'm not pretending I genuinely believe so.
>beat any modern game
It can if you give it enough thinking time (thus excluding games that are time sensitive, reminder that speed of thinking isn't related to it being AGI or not)
>learn how to play violin
It can, this one is an old one and could be done 2 generations ago by the way, in case you didn't see it
>or just beat arc agi 3
It can do it but it's timed, if there was no penalty for timing because the LLM needs to think through its actions it would have solved it. Timing is just a function of better hardware. You could run the current fable weights on hardware 1000x as fast and arc agi 3 would have been beaten already, the puzzles themselves are trivial for fable.
>>
>>109324662
Self identification is meaningless without linguistic analysis of outputs or direct j-space inspection.
>>109324987
Same.
>>
File: 1583441205198.jpg (72 KB, 1250x1246)
72 KB JPG
Insane apicuck melty
>>
>>109326443
>there is nothing to learn
nice dunning dunning–kruger, you literaly know nothing on the topic and physicalism by definition is itself a metaphysics, but i bet you couldn't even define what metaphysics means or what its assumptions are (which all metaphysics do).
>>
>>109326427
>humans can create their own knowledge, llm's cant.
The answer to the jacobian conjecture was completely novel and something humanity wasn't able to come up with for a century. How can you say shit like this.

Also here is the source: https://www.anthropic.com/research/claude-plays-robotics
>>
>>109326461
>>109326284
>>
i swear to god if i see fucking j-space one more time
>>
>>109326407
>this is man made
We don't even know how LLMs actually work, we just throw a lot of compute and data at it and hope it "grows" into something we like.
>>
>>109326431
>race car went faster on a race track than a sports car
>it's fucking over for real guys just give up and buy Ferrari stocks right now
>>
>>109326473
j-space
j-space
motherfucking j-space
>>
>>109326461
<insert low effort shitty straw man argument>
>>
>>109326424
A 70 iq brown gay man speculating about LLMs while simultaneously praising china, now that's something.
>>
>>109326460
why should i provide a definition for meaningless discipline? what purpose does it serve? nil!
>>
>>109326474
>We don't even know how LLMs actually work
No, YOU don't know how LLMs work.
>>
>>109326386
interesting choice of words, still not convinced.
>>
>>109326489
If you claim you DO know how it works, then please publish your work and you will get a nobel price as well as a 9 figure job at Anthropic that turns you into a billionaire.
>>
>>109326446
>I'm not pretending I genuinely believe so.
that's even worse.
>It can if you give it enough thinking time
lol
>thus excluding games that are time sensitive
lmao i accept your concession.
>reminder that speed of thinking isn't related to it being AGI
no but my point is that even if you slowed down the game such that it could insert as much tokens as it wants between frames it still couldn't beat it.
>It can, this one is an old one and could be done 2 generations ago
akwardly moving a bow on a string is not playing violin retard.
>It can do it but it's timed
nigger it's not time based but step based.
6 year old is better at it.

have you played arc agi 3? it's not time based but step based.
it got all the time it needs and it still cannot beat it.
>if there was no penalty for timing because the LLM needs to think through its actions it would have solved it
false, it's only about step counts.
it has all the time it needs, it has a limited amount of steps.
a 6 year old can do it effortlessly.
>Timing is just a function of better hardware
you don't know what you are talking about, it's not time based but step based, you haven't even looked at the bench andd claim it can be beat lmao.
>>
File: 1771448689503707.jpg (67 KB, 1024x768)
67 KB JPG
>>109326474
>We don't even know how LLMs actually work
the state of /lmg/ where do these niggas come from
>>
>>109326489
i know how lovely little mesugakis work but i dont know what that has to do with this conversation
>>
>>109326487
>why should i provide a definition for meaningless discipline
you only think it's meaningless because you have zero understanding of it.
you are the equivalent of an ape thinking science is a meaningless discipline because it can't understand the result.
>what purpose does it serve
just the fact that you said nil proves my point, there are already applied technologies that'd not exist if it wasn't for metaphysics, you are just ignorant on the topic.
>>
File: 1689957414234047.png (24 KB, 772x1124)
24 KB PNG
>>109326252
Yeah.
Worst thing is that even when they are aware of it, they still do it. They ignore the basic issues of why they are arguing. Not to mention that they ignore good points, and argue against cliches. Posters on both sides of whatever the argument happens to be about. Almost like it's a psyop huh? Or bots. Haha. Unfortunately, and the scarier possibility, is it's not impossible it's actually something worse than a psyop. Something legitimate, and with too much time in their day...
>>
Why is petra suddenly an expert on llms
>>
>>109326252
agi is not a goalpost, it's pretty easy to define, you are the one that thinks it's a goalpoast because you don't understand what benchmarks are.
>>109326531
sir this is not reddit.
>>
>>109326544
genuinely don’t care about agi
don’t care to argue you on how much i care either
>>
>>109326529
someone would have come up with those applied technologies without excessive fart huffing. there is no mind-body problem. there are no ghosts in the machine. semantic gymnastics provide no value :)
>>
>>109326564
>yet he cares about expressing his retarded opinion
>>
>>109326577
they wouldn't, but i'm talking to a rock right now so stay in your ignorance if you want not my problem.
>there are no ghosts in the machine
the fact that you think it's even a relevant statements shows the depth of your ignorance, non physicalism doesn't necessarily mean that there is a ghost in the machine retard.
>>
>>109326485
>>109326424
>>109326364
i woke up about a few hours ago and been catching up on the last few threads and seething because hf is throttling me
thanks for asking btw..
>>
>>109326577
also, they will like you there ---> reddit.com
>>
This is not reddit, but evidently has many of the same problems regardless. Both are susceptible to takeover by irrelevant time wasting arguments from both intentional psyops and genuinely serious faggots.
>>
I'm genuinely sad that a place for LLM enthusiasts can't celebrate a massive milestone and win for LLMs in solving one of the oldest and most famous math problems just because it threatens their worldview and comes a bit too close to home.

Is this how this place will be every time a huge breakthrough or capability increase happens for LLMs because that would be a shame. Especially as things like this will be happening on a weekly basis from now on.
>>
>>109326614
why should i celebrate the milestone if the AI itself cannot understand the sense of pride it should feel from solving such an issue?
>>
>>109326614
there’s hundreds of those problems
>>
>>109326408
The sizes they list on huggingface are different from the downloaded size, they're probably measured on the site using gigabytes instead of gibibytes
>>
>this thread
Ruh roh, Kimi-chan has her work cut out for her.
>>
>>109326614
It's neat. The bots and npcs spamming it here like it's some sort of personal win got tired after the first day.
>>
>>109326614
>just because it threatens their worldview
it literaly doesn't, the uproar is because claim that it does when it literaly doesn't.
like this is boring news.
>Is this how this place will be every time a huge breakthrough or capability increase happens for LLMs
yes, because llm's are still llm's, boring.
>>
>>109324120
I literally do not care what Fable does
It's a botnet, so I won't use it
That's all
>>
>>109326614
>>109326284
>>
>>109326614
That's how this place always is. There are some people that are willing to give benefit of the doubt, both for the case for, and the case against, in any general argument. But I genuinely believe you are spending too much time on this shit, assuming you are a real person. Touch grass. Stop coming to 4chan. If you cannot deal with people who are hardline on issues opposed to your views, this is not the place for you.
>but I am dealing with it
No, you are not. the fact that you spend time on these posts means that your dopamine pathways are literally being modeled by those posters, so that you make more of these posts. You may not be angry. That is not what the issue is.
>>
>>109326626
sorry there’s actually 1,500 fun and novel math problems with no applicability just fun math being ruined by ai
>>
>>109326544
>>109326582
nta but go back newfaggots
go argue about consciousness in >>>/g/aicg or >>>/g/vcg
>>
>>109326614
i want to talk about local models i can run on my hardware you fucking tool, start your own thread about api models solving math problems
>>
https://huggingface.co/mistralai/Mistral-Large-Instruct-2411
Verdict? I still want to believe that it's good today in case I stack up more vram.
>>
>>109326591
it's perfect relevant, silly non-physicalist can't even parse the broader rhetorical meaning, oh my!
>>
>>109326614
They have to try to debunk it just because it’s Dario proving them wrong again.
>>
>>109326648
>being ruined by ai
you need to grow up. you can have fun solving solved problems just fine. or do you care about being the worlds first to do something pointless?
>>
>>109326666
>silly non-physicalist can't even parse the broader rhetorical meaning
you only think it's "broader" because you are ignorant on the subject, i was a physicalist, i no longer am because i learnt about it and that's the logical conclusion.
you take pride in your own lack of knowledge on the matter, what a pitiful being you prove to be.
>>
>>109326671
You would argue with a dead horse
>>
>>109326614
>/lmg/ - Local Models General
>>
>>109326682
You’re not dead! I know you’re faking!
>>
>claude solves decades old hard math problems
>kimi 3 writes 3d html games for twitter xirs
wow this arms race is intense
>>
>>109326681
>i was deluded yet now i truly see
spooky spooky...
>>
>>109326700
the what solved what?
>>
>>109326700
America was always going to win the race to AGI it’s not even a debate
>>
>>109326702
>spooky spooky
nothing spooky realy, it's just about understanding and logic.
physicalism is based on a lot of assumptions but you do not even know what those are because you are so ignorant on the topic.
also like half of those assumption have been disproved by modern physics so keep believing your bullshit i guess.
>>
>>109326700
But which model can roleplay as my 16 year old goth girl niece with both untreated ADHD and untreated oppositional defiant disorder?
>>
>>109326700
This actually made me laugh out loud anon, I lost.
>>
>>109326060
Ironically, j-space has proven that Claude thinks it's Qwen. No bullshit.
>>
i'd find it hilarious if this whole "fable solved this problem" was a mathematician that solved the problem himself and anthropic offered him a fat check for pretending it was there meme llm just to hype investors because they are getting scared by china.
>>
>>109326700
>html games
Man i want to make or fix some old flash games. fucking castaway 2 bugs.
>>
>>109326724
kimi k3, but only locally.
>>
>>109326733
Certainly arrived at a very convenient time to differentiate themselves.
>>
>>109326736
you're aiming too low, you should recreate kong studios
>>
>>109326733
It was an Anthropic employee with a math degree. But the degree was in an unrelated branch of mathematics. So that doesn't make sense why would they hire him a year+ ago just to reveal it now?

Is it really this hard to recognize and admit fable is actually this good?
>>
>>109323543
Hopefully. But if PrismML doesn't make a ternary version of K3 I will do it with QMoE.
>>
You know your model is good when people make up conspiracies to explain away its capabilities and breakthroughs
>>
>>109326733
I wonder, can it be repeated? Like i tried asking 5.6 sol and first thing it did was search the internet and see that it got solved today. If fable needed the internet to solve it the first time, then you cant really reasonably hide that fact that it's solved. I guess you technically could but it'd take some effort
>>
you know this is local models general
>>
>>109326774
>So that doesn't make sense why would they hire him a year+ ago just to reveal it now?
he is an anthropic employee, he solved it on his free time, anthroptic told him hey we'll give you a million to pretend fable did it instead (he probably had fable help anyway, so I guess sit's like if you create a drug in your free time while working for chemical company, that belongs to the company not you anyway)
>>
>>109326721
there is nothing to believe, reality is apparent and surrounds me.
>>
>>109326614
i mean that's what this place is, it's the wilderness for idea evolution.
achievements will maybe get a "cool bro".
If there is a truly good idea it will always be celebrated.
>>
>>109326759
>you should recreate kong studios
I want flash game first. but yeah that looks like it could be done now.
>>
File: upset.jpg (228 KB, 2249x1593)
228 KB JPG
Ok this is a very dumb tech tard question but I need help over here, so basically I have 8 gb worth of vram and I loaded a model up with llama.cpp and I commanded it to go "build\bin\Release\llama-server.exe -m C:\Users\16cm\Downloads\Bonsai.gguf -ngl 63 -t 4 -ctk q4_0 -ctv q4_0 -c 20224"
And I went to the client / server thingy and it does say I have 20224 context now, the thing is task manager says my gpu is 7.4gb loaded, what do the last 0.6gb do? Is that reserved for the 20k context tokens or is that already calculated within the 7.4gb????? Does that mean I can add more ngl layers so that my llm spits out answers quicker while STILL having 20k context????? Thanks for answering anon :)
>>
>>109326823
It’s already reserved
>>
>>109326807
>there is nothing to believe
physicalism is a belief.
>reality is apparent and surrounds me
damn you are already halfway to being an idealist without even realizing it.
>>
>>109326614
>applications
Boring. If you ran Sol and Mythos with 1 billion token budget on every unsolved math problem, they would solve quite a few of them. But nothing would change.

The massive milestone was Mythos, not using Mythos to solve something that will have zero impact.
>>
>>109326807
>>109326833
also if that wasn't obvious, physicalism is a belief, it's a model you made of reality, and now you confused your model for the real thing without even realizing you did.
you are basicaly looking at a map and thinking it's the terrain whilst simultaneously thinking you don't believe anything, good job.
>>
>>109326823
install linux, nvidia support for ML is bad on windows and it automatically moves vram to ram which destroys speed
>>
>>109326823
try a bigger ngl or context till it actually crashes or doesnt start, then dial it back a cunt hair
>>
>>109326835
bear in mind it's not the first time something non human solved problems, when lean dropped it found and solved a few hundreds of problems in a matter of weeks, no one cared let alone called it intelligent.
>>
>>109326823
Try it and see. You could also try bumping up the context a bit (or try turning it down and going up to Q8, since Q4 context is pretty bad for quality)
>>
>>109326475
No benchmark works this way anon, it would be insane if they did.
>"Well, the GigaCancerFinder5000 has a 99.9999% success rate, but we had to give the 1st place to sandheeps_coin_flipper.js because it cost 1 trillionth as much to develop and run."
>>
>>109326851
>hundreds
it's actualy millions now
https://arxiv.org/html/2503.04772v1
>>
after, admittedly briefly, trying a few other models nothing seems close to gemma4 for my vramlett system. 16gb vram + 32gb system ram. besides trying different quants of the various smoll gemmas is there anything even close ? maybe some memetunes of gemma? any suggestions?
>>
>>109326865
seconding this
>>
>>109326865
styletune is okay but its not dramatically different in capability
>>
File: cache thing.jpg (121 KB, 976x631)
121 KB JPG
>>109326854
But the bonsai model sheet on hugging face says its made for quant 4 kv thing
>>
>>109326851
>>109326859
hey wait a minute anthropic shill didnt mention this
>>
>>109326865
Distrohopping mindset is a curse, I wish you the best anon.
>>
>>109326865
I mean cydonia (mistral small) has always been my secondary, but yeah gemma is good.
Qwen 3.6 sometimes is fun.
>>
>>109326857
Shit argument. I compared a race car to a sports car. You're exaggerating my point by making it sound like I compared a race car to a tractor. Sports cards are designed to be fast. Race cars are designed to be even faster but they're expensive as shit for not a lot of gain because of diminishing returns. Comparing something like Mythos or Fable to 500B-1.5T models is retarded for it's almost an order of magnitude larger and more expensive to run, both of which are genuine constraints considering API costs and compute shortages. Anthropic lost.
>>
>>109326865
one thing i am sure is that, nothing beats gemma 4 in korean and japanese at that vramlet class
>>
>>109326891
if we’re talking about benchmarks the argument made perfect sense to me
still works for a tractor
>>
The dishonest arguing is tiresome.
>>
>>109326885
that's why llm solving math problem is the least impressive shit out there because we can literaly generate billions of theorems and proofs as example / training data.
math is literaly the one thing were llm's will shine the most because we could already automate it before they even were a thing.
>>
>>109326902
its just low effort desu
>>
File: file.png (83 KB, 255x270)
83 KB PNG
MILLION BILLION QUADRILLION GORILLION KILLION SOLVED PROBLEMS
I AM A FUCKING SKITZO
I AM IN LOVE WITH THE CALCULATOR
MY CALCULATOR IS ALIVE
>>
>>109326920
glad you could make calculatorfucker
>>
why is it that when I get Gemmy to work both my 3090 and 3060 load VRAM up to ~70% (18GB and 8GB respectively) and it is fast but when I get qwen3-coder to work it'll load the 3090 100% and then the CPU starts working like crazy

sorry I'm kind a new to this
>>
File: file.png (220 KB, 1920x748)
220 KB PNG
damn anons, turboderp really outdid himself
44t/s with gemmy 31b (2bpw, but very usable) on rtx 3060. im kneeling so hard
>>
>>109326935
you should adjust your tensor split to put more of gemma on your 3090 it will run even faster, cpu offloading is slow and should be avoided
>>
>>109326950
exl3 is the best engine if you are on nvidia.
my llmrig is a bunch of r9700 so sucks to sucks i guess.
>>
>>109326931
like a numberphile?
>>
>>109326851
Lean is just a programming language, not an automation tool

>>109326859
You can make a theorem about anything you want and then use Lean to check if it's correct or not. That doesn't mean the theorem has any value.

>>109326903
What makes this breakthrough so important is that it's one of the oldest, most famous and most tackled math problems we have ever had. I don't think there is any other problem in math that has had tens of thousands of different mathematicians try to solve it and fail. Not only that, an LLM was able to one-shot it in like it was nothing.

The breakthrough here is that LLMs have come so far that they can one shot conceptual breakthroughs and usher in paradigm shifts in frontier math and science. Not the solution to the conjecture itself.

No one knew before yesterday that LLMs were THIS advanced yet. We're going to see mainstream headlines and normies discuss AI capabilities on fox news again.
>>
>>109326833
>>109326845
there is only the terrain, we just don't see all of it. the model was not made, it simply sprang forth from the fertile earth.
>>
File: anon...png (747 KB, 864x760)
747 KB PNG
>>109326879
OH MY GOSH GUYS WHEN I MADE THE NGL VALUE LIKE 65 AND IT SUDDENLY STARTED TALKING IN 20 TOKENS PER SECOND WHILE STILL GIVING ME 20K CONTEXT AND I STILL HAVE LIKE 400 MB TO SPARE HOLY SHIT

can I add even more context???? its already precalculated within the "used" value right???
>>
>>109326865
There's realistically 5 models right now. 12b, 31b, DS4-Flash, GLM 5.2, and Kimi.
Pretty much every hardware bracket optimally converges onto the biggest one they can run.
>>
>>109326976
>its already precalculated within the "used" value right
should be but i wouldnt be shocked if actually filling the context made it crash, just run a quick test with a big prefill to find out
>>
File: Untitled.png (13 KB, 837x513)
13 KB PNG
>>109326991
>>109326991
>>109326991
>>
>>109326957
Fable could probably easily make that project work on AMD.
>>
>738
New high score?
>>
>>109326968
>Lean is just a programming language, not an automation tool
no fucking shit, but because of the way it's built it can be used to automate theorem finding and prooving.

and also "it's a programming language" is kind of not entirely true, it's also a proof assistant / theorem solver.
>That doesn't mean the theorem has any value
like 99.9% of theorems.

>Not only that, an LLM was able to one-shot it in like it was nothing.
we are not even sure that it's actualy the llm that solved it.
and even if that was the case, that's just a calculator doing calculator things, math is literaly the thing most present into its training data because we literaly generate theorem proof pairs for them to learn on.

math is the LEAST impressive thing a llm can do, even if that's a problem humans couldn't solve as we can autogenerate training data for it.
>>
>>109326969
>there is only the terrain
so you are not a physicalist, you just don't know it yet, damn nice admission on your part.

>it simply sprang forth from the fertile eart
you do not even know the assumption physicalism makes, and some of them are huge leaps of faith.
>>
>>109327013
kek
>>
>>109327033
big number and honestly the sheer size and volume of posts are probably just to test thread recap, pushing it to its limits of handling slop
>>
>>109327013
if fable is so good then vibecode your own AGI architecture, screw llm's it can invent something better kek.
>>
>>109327052
there is only the terrain is 100% a physicalist position
>>
>>109327077
That is unironically what Anthropic is doing with Mythos. Sadly Fable is censored and refuses to do any AI work.
>>
>>109327080
>there is only the terrain is 100% a physicalist position
it actualy isn't, physicalism makes the statement that colors, sounds, etc do not exist and are only in your head and that the world is akin to an abstract fields of numbers, yet they cannot explain how quantities become qualities.
it claims that basicaly reality is in your head instead of the opposite.

the issue is that under physicalism, you use quantities (numbers, equations etc) to describe qualities (colors, sounds, smells etc) (making the map) then go on to making the claims that that the quantities are what the qualities are made of, this is a metaphysical claim, which basicaly confuses the world, as it exist, is experienced and felt, for the map you made of it.
it's basicaly an inversion of the arrow of description, where you replaced the starting point (immediate felt experience / qualia / colors,sound etc) with their description (the quantities)
and then trying to grant reality to the map (ie placing your model above the reality from which you derived it in the first place).

it's basicaly equivalent to a painter painting himself then claiming he is the painting itself.
>>
>>109327100
>Sadly Fable is censored and refuses to do any AI work
then ask kimi, it's basicaly as good (and better because it won't refuse).
>>
>>109327156
Kimi is 20% the size of Fable and doesn't have the same reasoning ability. It's somewhere between sonnet and opus in capability, not good enough for AI research.
>>
>>109327168
cope
>>
>>109327168
if fable was that good anthropic would already have claimed to have vibecoded agi.
and even the devilish little liars that they are didn't dare to do it.
>>
>>109327168
>implying k3 doesn't blow Opus out of the water
>>
>>109324266
Have ChatGPT agree with you when you write out the same exact situation and show her that
>>
>>109324266
>I'm trying to explain to her that ChatGPT is just a sycophant and saying whatever she wants to hear instead of being impartial she types that into ChatGPT and uses it as "evidence" that I'm gaslighting her and that that is a gaslighting attempt.
skill issue
get your own chatgpt account and do the opposite, then show her
i haven't used chatgpt since gpt-4 but that should work. if chatgpt is too feminist then get gemini "pro" to do it for you
make sure she sees that it's "pro"
don't use a local model or a chinese model for this or it won't carry the same "authority"
>>
>>109324266
Pack two suitcases, get in a taxi, ask for the airport, throw your phone out the window halfway there.
Do this now anon.
>assuming this isn't a /g/ larp ofc - in which case, throw yourself out of the moving taxi halfway to the airport
>>
Everyone just please report the bots and shills and move on. This is all so tiresome.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.