/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>105800515 & >>105789622

►News
>(07/04) MLX adds support for Ernie 4.5 MoE: https://github.com/ml-explore/mlx-lm/pull/267
>(07/02) DeepSWE-Preview 32B released: https://hf.co/agentica-org/DeepSWE-Preview
>(07/02) llama.cpp : initial Mamba-2 support merged: https://github.com/ggml-org/llama.cpp/pull/9126
>(07/02) GLM-4.1V-9B-Thinking released: https://hf.co/THUDM/GLM-4.1V-9B-Thinking
>(07/01) Huawei Pangu Pro 72B-A16B released: https://gitcode.com/ascend-tribe/pangu-pro-moe-model

►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/tldrhowtoquant
https://rentry.org/samplers

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

►Benchmarks
LiveBench: https://livebench.ai
Programming: https://livecodebench.github.io/leaderboard.html
Code Editing: https://aider.chat/docs/leaderboards
Context Length: https://github.com/adobe-research/NoLiMa
Censorbench: https://codeberg.org/jts2323/censorbench
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

►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

►Recent Highlights from the Previous Thread: >>105800515

--Stagnation of closed model SOTA and limitations of local model development due to data and training issues:
>105801389 >105801404 >105801516 >105801638 >105801659 >105801436 >105801445 >105809251 >105801590 >105801625 >105801663 >105801722 >105801765 >105801797 >105801681 >105801721 >105801741
--Local TTS alternatives for audiobook generation post-ElevenLabs paywall frustration:
>105804805 >105804924 >105805063 >105805114 >105805133 >105805191 >105805212 >105805345 >105805873
--Quantization effects and performance comparisons across model sizes and architectures:
>105806470 >105806508 >105808470 >105806628
--Evaluating quantized models and hardware limitations for local large language model inference:
>105806334 >105806353 >105806359 >105806370 >105808855 >105808898 >105806425 >105806467 >105806679 >105806719 >105806343 >105806402
--Skepticism toward ASUS's GB200-based AI mini-PC amid memory and pricing concerns:
>105807146 >105807160 >105807319 >105807354 >105807387 >105807595 >105807858 >105807176 >105807921 >105807937 >105807957 >105808084 >105808135
--Anon recounts prompt tampering and code logic errors from closed AI models:
>105805730 >105805826 >105809326 >105809380 >105806753 >105806761 >105806772 >105808880
--Testing deepseek r1 qwen3-8b's limits on sensitive topics reveals model guardrail behavior:
>105801403 >105801459 >105801495 >105801532 >105801552 >105801641 >105801671 >105801707 >105801794 >105801823
--Critique of LLMs in gaming and advocacy for hybrid AI approaches with local models:
>105807514 >105807588 >105808801 >105808840 >105809374 >105809415 >105809442
--MLX adds support for Ernie 4.5 MoE with 4-bit quantization:
>105807394
--Excitement around Grok model benchmarks:
>105802337
--Miku (free space):
>105800984 >105802436

►Recent Highlight Posts from the Previous Thread: >>105800519

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script

first for rocinante

do processors matter?

I hated AI until I found how decentralized it has become. :)
Very cool.

2026 will be the year of local. I can feel it.

>>105811226
it's an ongoing effort to make sure it's not just used to further a couple billionaire midwit's goals
I think the best next step that we are taking is decentralized training, the requirements are getting fast quickly, projects before required you to at least possess a 8xH100 cluster, you just need 1 H100 now and hopefully soon it will go down to a 4090/3090 so that everyone can contribute

>>105811246
Same as it ever was.

>>105811246
>getting fast quickly
yup, another ESL classic from me, you know what I mean

Anyone working on a blockchain based decentralized LLM with PoW?

>>105811224
not if the model is entirely on gpu

Is it possible to use Claude Code with a local model?

>>105811224
sometimes. The number of PCI lanes matters in multi-GPU setups

ernie will save local

>>105811378
Ask claude

>>105811769
kek

Thread culture recap.

why are you uploading your fetishes here

>>105811913
I shit on you and your waifu faggot.

>>105811913
his fetish is public humiliation

>>105811717
Oh no.

>105811913
why are you?

>>105811717
This is a masterpiece

I reported bghira/Simpletuner creator's lora who was spamming false reports everywhere online and his own model got taken down now

for context
https://www.reddit.com/r/StableDiffusion/comments/1lsfobb/full_breakdown_the_bghirasimpletuner_situation/

>>105811717
Imagine if we had enough of this data to train a model like 4o on it. If only we had a 4o.

Are there any other LLM function calling benchmarks except
https://gorilla.cs.berkeley.edu/leaderboard.html ? It seems to lack newer models, like o3, o4 mini and lots of others (not updated in a month)

>download "uncensored model"
>it won't discuss the Nword with me.
:(

the daily things


like this and that and what is what


that keep us all busy


are confusing me

>>105812428
in my experience none of the so called uncensored finetunes are truly uncensored, and finetuners are just grifters destroying the model integrity.
The first abliterated models (like Gemma 2 9B abliterated by gusev) were pretty good and it didn't feel like the smarts of the model was amputated, but the newer abliterations by retards like mlabonne and huihui are just broken models, they are indeed uncensored but far more prone to going into loops. Because those fucktards are always first at releasing new abliterated models I think they sucked all interest from others in trying to make a good abliteration combined with huggingface's search being so shit.
You are best served by experimenting with making jailbreak prompts for the model you want to use. This will be less damaging to the model than using a dumbtune. I won't post my own here, I don't want my prompts to serve in the next censor round.
You could also go for a model that actually obeys your system prompt, like Nemo.

I just compiled llama-cli for fun and learning, but I struggle to find a decent model that comes without safeguards, iykwim. What would you guys recommend?

>>105813040
rocinante

>>105811246
The problem now is that good training data is becoming the bottleneck for open source rather than compute

all the public pretraining datasets are trash quality compared to what big labs have
Deepseek is probably the only open source lab with pretraining data of comparable quality to the big boys, and they have not shared it

this thread is dead as fuck
what's going on?

>>105811237
every year is the year of local, no point in using AI if its not on your computer

>>105813607
AI winter, the bigger problem stems with closed source AI are all hitting deadends and so energy for local stuff wanes, idk what everyone is expecting anyway, AGI isn't around the corner and incremental improvement is going to stay like this for at least another decade

>>105813607
Same with the image gen thread I frequent.
A lull in the global consciousness.

>>105813614
every normalfag would disagree with you

>>105813607
llama.cpp is dragging its feet implementing all the new models that have been coming out so nobody can try them

>try the new models
>they're all shit
>after the dust settles, thread goes back to being dead

>>105813628
As the other anon noted, a bunch of good stuff has been dropping lately. We just can't use any of it yet due to slow implementation from inference engines.

>>105813655
Why isn't there a "standard" that all model companies use so they can just PR whenever something new is needed so it can be used out of the box? Or is there and it just isn't llama.cpp?

>>105813664
Some companies do sort out llamacpp PRs in advance in order to get day one support, it isn't common though, and the Chinese (who are the ones releasing the good OS models lately) never seem to do it.

>>105813628
man up, faggot
no need to have a meltdown
>>105813630
image/video keeps improving by the week,
/g/ threads just happen to be filled with literal retards
>>105813655
the UI aspect of local AI remains its greatest downfall

>>105813607
it's been a year since we got a local model worth running

>>105813712
and it's been 3 years since all this shit exists, retard
we don't even have AI chipsets yet
calm down and stop whining, try not to be miserable for 5 minutes

>>105813683
>UI aspect
Open webui is a thing

>>105813730
sorry, I'll be grateful for all the effort they put into making models worse

>>105813664
That's regular old transformers, but you need tons of VRAM, and can't exactly offload to RAM and expect to actually have remotely usable speeds with that, so for the majority of local poors llama.cpp based stuff is the only logical option.
There's also vLLM that supports a lot of models, but same-ish downsides as transformers.

>>105813741
Open Webui is shit though

>>105813741
eh...
>>105813742
unironically you should because failure is just a stage of success

>>105813659
But how do you know it's good without inference?

>>105813799
By testing on apis

>>105813802
Any examples of good stuff you've tested yet not present in inference engines?
Technically api is a middleman between an inference engine and you btw.

>>105813040
Deepseek r1

>>105813752
Be the change you want to see

>>105813827
pay me my expenses for 3 months and we have a deal.

>>105813859
Ask AI to code it for you.

>>105813947
then that's gonna take 9 months

>>105813984
Are you HDDMAXXing R1 through an USB 2.0 adapter?

>>105814004
>tfw completely viable way to have r1 at home.

>>105814152
9 months for a prompt is viable to you? Might as well just go knock a biogirl up.

>>105814165
That's more expensive than buying a server from leather jacket man.

>>105814004
my point is that llm are retarded and they'll actually make you lose time when you try to engineer something more complex properly.

>>105814152
Haha, if I read a SF story as a kid where I could run an very smart AI on my home computer at an incredibly slow speed so it took a week to answer questions I would've thought that was the coolest thing ever and been so excited for it.
Now that's actually the case and I don't give a shit.

>>105814298
*a very

>>105814309
Standards are fickle things
Given enough time, nothing is novel anymore

Mid-late July has always been the time when big new things drop. Two more weeks and summer is here.

>>105814298
>where I could run an very smart AI
>Now that's actually the case
Lol. It's not. It's quite smart, but not very in the ways that matter at least for individuals like you or me, such as solving your life problems zero shot and sucking your dick. It can certainly help you, such as being a coding assistant, but it can't be a replacement for much. When things get to replacement levels, then that's where you start being amazed every time you think about it, although the "wow" moment emotional feeling will only ever happen a few times, which is true for everything no matter what it is. You could be frozen in a block of ice, emerge to a sci-fi world, and then quickly acclimatize to it after the initial mindblow. Maybe AI will even solve that though, it'll continually come up with new ways to mindblow you, like some kind of personal clown.

>AI can't suck my dick, that means it's useless
lmao

>>105813607
>this thread is dead as fuck
>what's going on?
The baker is a literal terminally online, power tripping, AGP, coomer, spammer, schizo, jannie who made it unbarable to be in the thread, paired with the heavy censorship of 4chan in general by other similar janitors, there simply is no point to be here compared to literally any other AI community online.

You have to do a humiliation ritual every time you post even with captcha auto-solvers and wait a few seconds anyway just so your post gets deleted for a random or no reason in case the terminally online janitor doesn't like it while the janitor baker posts porn in it on a blue board that stays up for hours while everyone who points this out gets banned sometimes literally within seconds that lasts for multiple days to even more in case the janitor doesn't like you and "extends" the ban to weeks or months for the second arbitrary reason he makes up.

Making a burner account for almost any other place online and posting almost anything except the most obvious call to action illegal things won't get you in much if any trouble while the discussions are also much higher quality, no waifu obsession schizo meltdowns, much less "hey guys can i run gtp4 image generation on my laptop 1060 3gb??? no i didnt read the op" tech support questions, and and much more actual tech discussions instead.

All the same reasons why 4chan in general is dying and heavily dead already. Tranitors.

>>105814396
Who are you quoting?

>>105814500
dvbs status: confirmed
trvth status: nvclear

>>105814500
>"hey guys can i run gtp4 image generation on my laptop 1060 3gb??? no i didnt read the op"
lmao
>>105814515
sent :^)

miku

>>105814599
Who?

miku

>posts warhammer and basedjaks
that's about as gay as it gets anon

do women always 'press their thighs together' when they're aroused? Because ever female character, in every chat, with every vramlet model I use does this

>no activity on the ernie vllm dev branch for four days
it's over isn't it, we're never going to get to even try the vision

>>105811029 (OP)
Status on Hunyuan GGUFs?

>>105814500
> hey guys can i run gtp4 image generation on my laptop 1060 3gb??? no i didnt read the op
because there is nothing to read in the op, it's never updated useless shit apart from news

>>105814976
Right after jamba

>>105811029 (OP)
> DeepSWE-Preview 32B released
Is this from deepseek creators?

>>105814500
Also, the lull in releases.

>>105813476
>all the public pretraining datasets are trash quality compared to what big labs have
>Deepseek is probably the only open source lab with pretraining data of comparable quality to the big boys, and they have not shared it
it's because sharing those datasets will get a heavy contingent of anti ai on your ass
there's no such a thing as a good model that wasn't trained on copyrighted material
you can't share that shit
and even when you can share that shit (like datasets made from websites that explicitly allow scraping uses) butthurt people will harass you until the platform hosting you takes you down, for eg :
https://huggingface.co/datasets/nyuuzyou/archiveofourown/discussions
public pretrain datasets that won't trigger a horde of butthurt is basically restricted to project gutenberg books (old public domain stuff) and GPT generated slop.

Sam's new, fresh GPT-5 revealed
https://huggingface.co/yandex/YandexGPT-5-Lite-8B-instruct

>>105815509
>config.json
>LlamaForCausalLM
>8B
98% chance that this is just a llama3 8b finetune
2% chance that they actually trained something on llama architecture lmao

>>105815565
>"bos_token_id": 1,
>"eos_token_id": 2,
In llama3-8b it's
>"bos_token_id": 128000,
>"eos_token_id": 128001,

Damn MCP is bullshit.
I couldnt get one up and running locally and access it by a local .html page via http. CORS problems, session/oauth2 problems etc.
I managed to get a sessionid by making a proxy flask server around my mcp server..but still never managed to get a list of the tools.

I'm a brainlet and maybe thats the reason but this feels way too complicated to setup.
Also its crazy how pajeets leech of "local".
>HERE IS HOW TO MAKE YOUR SIMPLE MCP SERVER USING LOCAL MODELS
Some online middleware shit for authentification (???)
The tool? Simple to set up. ENTER API KEY HERE. (First 5k requests free!)
Yeah, really cool. Might as well not use a local model at that point, wtf is wrong with these people.

At this point im seriously considering making my own frankensolution to call tools.

>>105813476
The era of simply pretraining on semi-raw web "documents" is almost over, at least for competitive models. It's not just that the data is bad; that's not even necessarily the issue. It's that the current way of pretraining models on random short web snippets with only very superficial quality checks and mangled/destroyed structure, and then post-training them on (currently tens of) billions of tokens of finally purposeful and coherent long(ish) samples is leaving a lot of performance on the table. That's not even taking into account the dubious data source filtering applied by the companies releasing open-weight models.

Bad news: once companies will start (semi-)synthetically rewriting the entire pretraining corpus to address the quality/quantity issue, the models will likely become virtually useless for any "unsafe" use.

>>105815688
you can't tell me there aren't any companies interested in selling smut.

where there is a market, there is profit to be made.

>>105814500
name one
>protip: you cant, and will cope with "muh gatekeeping", making your post worthless.

>>105815699
big companies aren't interested. Smut is usually handled by smaller ones.

>>105815688
>rewriting the entire pretraining corpus
this can only amplify GPT slop, which is a worse issue than the models being safe to me

>>105815699
>you can't tell me there aren't any companies interested in selling smut.
you don't understand the scales of investment to make good LLMs son
even something like making a model as good as the current qwen pretrains is out of reach of the companies willing to do NSFW
novelai can't even handle doing a good finetune of pre-existing models like llama
porn money is not enough money, you need Super Giant Corporation level money - Microsoft (openai largely was funded by MS), Google, Alibaba, Finance firms (DeepSeek) etc

>>105815758
give it 5 years.

>>105815738
And we still can't get the C.AI of 2022 back from any other company, commercial or even local open source because the cost to create that model is still out of reach for people even today with how expensive training gear is. The entire culture of hand curating high quality data to train an LLM is just gone and not there on top of the entire issue with payment processors not letting that content go out. Companies would rather slop for benchmarks rather than actually go and drive for better performance benchmarks can't measure. It's sad and maddening at the same time.

>>105814500
Truer words have never been spoken in this troon infested shithole. Death to /lmg/.
>>105814675
you play with dolls you stunted child

>>105815817
>C.ai
wasn't it more retarded than a modern 8b?

>>105812527
I feel like what you posted is roleplay brainrot. You want the model to be smart and witty and have banter with you as a character. To entertain you and act as a genuine AI. But of course, llm's are dumb hallucinating garbage and you desperately try to find the 'smartest' model you can find, only to be disappointed every time.

The fine tunes are not for you.

As someone who uses them to write, when the LLM makes a mistake or fails to write what I want, I edit the prompt or inject some of my own writing (often times a single word is enough) and off it goes doing exactly what I wanted. I'm not trying to preserve my stupid waifu.

What I want is an LLM that has literally seen any amount of smut during training so it's not stuck in purple prose hell for eternity due to aggressive data scrubbing procedures.

They arent grifters bro. Finetunes are so much better for writing its not even a contest. Stop trying to treat AI as youre fucking girlfriend and you can enjoy this hobby.

>>105815699
i think its just super risky
>idiot parent sues you after their neglected child kills itself and your shit gets blamed
>someone generates "csam" and you get 50 news sites making an article about it
>payment processors that can just cut you off at any time without a single shit
>id verification that is getting pushed everywhere now all of a sudden
>a lot of countries restricting porn in general

>>105815849
>They arent grifters bro
hi drummer

>>105815741
The companies don't care, they only want good benchmarks and "safe and trusted" results. LLM-rewritten data that uses human data as a "seed" isn't too bad compared to fully synthetic data, in any case.

>>105815849
you are the one with the writing brain rot
you think LLM refusals are only about coombrain? it's because you are coombrained that you assume I am talking from the pov of roleplay and not any other use like race statistics analysis

>>105815849
>They arent grifters bro
>when the LLM makes a mistake or fails to write what I want, I edit the prompt or inject some of my own writing
So you are saying that you aren't responsible for your valkyrie/rocinante/normandy product and the way you are supposed to use it is that if it doesn't work you should just use notepad.exe? Good to know.

>>105815830
Yes, but it had unparalleled RP to this day with the unhingedness, creativity and spontaneity almost no model can match. I think R1 gets closest but it is way too unhinged and spontaneous and still fails creativity. The worst part still is slop. and that drags everything down. Someone should try and replicate it in a 1B model possibly but we don't know how big the 2022 C.AI model even was and if that can get you there if you had enough good data.

>>105815954
you sure your nostalgia goggles aren't just very rose tainted?

>>105815954
Original Command-R 35b
dots
both of these have unhinged sovl.
they're also all retarded.

>>105815819
>you play with dolls
oh okay so it is hugely gay but you think those are my photos despite the clear twitter filename and are using that as a bad-faith equivalence.
good to know you're not only retarded but gullible.
intentionally posting ugly dudes and dumbass mech men is, in your mind, a sick own.

>>105816130
name 1 local model that can produce this level of soul

>>105816206
i went from wanting to kill myself to laughing and back several times while reading this

>>105816206
Gentoo isn't even approved by the FSF, he should have told you off for that.

>>105814500
>bitching about "schizo meltdowns" while having one
your kind is not very smart

>>105814525
reddit nigger post

>>105816206
I want a sloptune that had chan post dump in the dataset

>>105816263
If you want your model to suffer from tranny derangement syndrome, just prompt it to do so.

>>105814500
how do I download more vram?

>>105816270
Just filter out posts with the culture war words

>>105816206
uh oh deepsneed shills don't look

>>105811378
yes

>>105816280
this site is nothing but culture war and advertisements

by rewriting gemini cli i mean, yes

>>105816280
>I want 4chan
>but without the 4chan

>>105816203
That post about why this thread is dead is spot on. Limp wristed troon faggots like you would be bullied out of this thread if this was 4chan. Instead you feel like this is your hugbox safe space. Kill yourself. Nobody will miss you you mentally ill troon.

Quick question: I'm considering upgrading to a AMD 9800X3D CPU. Assuming the whole model is loaded onto system RAM, how make tokens/sec can I expect out of just CPU performance?

>>105816397
It will be 3 times faster.

>>105816397
consumertrash with 2 memory lanes will always be shit unless you're planning to run 3b off ram

https://www.thejournal.ie/meta-workers-ireland-6745653-Jul2025/
What is going on to build the safety datasets is deranged. They're making people prompt for this content all day every day.
I know some of you filthy subhumans would look forward to getting a job there because you like this content but normal people don't.
>>105816334
>>but without the 4chan
tbf the culture wars exist everywhere, and I really mean everywhere. Even if you take the 4chan out of 4chan people will talk about it.
You can't even stop this topic from being mentioned in hugboxes that exclude other opinions. Bluesky only has leftists and bans wrongthink but the leftists there will often quote people from other platforms to then go on a tirade about whatever it is that triggered their culture war bone today.
There's no rightoid place that won't be obsessed 24 hours a day with culture war either. And neutral places are places where people can shit on each other in the open.
You know, they were right when they said "EVERYTHING IS POLITICAL".

>>105816263
drummer please make something useful for once

>>105816397
<think>This anon wants to upgrade to an anemic CPU, 8 cores, barely matching a laptop in raw performance outside of gayming. I should respond with a clear, detailed technical answer.</think>
It should match a 5090 in terms of both PP and TG.

its possible to run gemini cli with local models btw (thanks claude code)

>>105816433
I know it'll be trash if I put a model purely on CPU. Its just that knowing of your CPU performs gives you a minimum baseline.
For example, my current rig uses a 5800X3D and a 7900 XTX. If I put the model purely on GPU, I get 37 tokens/sec. If the model is purely on CPU, I get 3 tokens/sec. This means every model I run will fall between these two speeds.
In other words, instead of upgrading my GPU (for potentially thousands of dollars), I could instead upgrade my CPU and RAM for a lot less money for a minimum that is closer to my own reading speed.

>>105816478
https://github.com/google-gemini/gemini-cli/pull/1939

>>105816507
yeah i already did it with claude code and have my own solution now

>>105816486
Token generation is limited by memory bandwidth so the only upgrade you'll see is the improvement of going from ddr4 to ddr5 here, which is still going to be slow with dual channel.

>>105816435
>https://www.thejournal.ie/meta-workers-ireland-6745653-Jul2025/
Abandon hope for any future Llama model to be good for RP.

>>105816621
>writing suicide-related prompts for AI led to them self-harming for the first time in their life.
This is what happens when you force double digit iq subhumans to think in hypotheticals.

>>105816486
>In other words, instead of upgrading my GPU (for potentially thousands of dollars), I could instead upgrade my CPU and RAM for a lot less money
Nope, to get any meaningful upgrade there you'd need server hardware, and guess what's the price range for those things.

>>105814599
Damn that food looks good, also nice miku.

>>105816435
dream job

>>105816830
That article doesn't sound entirely credible, but I'm sure they'd make you to watch and annotate nigger execution videos.

>>105817179
>>105816830

>>105816435
I remember the work on OpenAI's safety dataset was so vile that it gave Kenyans PTSD.

>>105817299
At least OAI had the common sense to outsource that to the third world, just like all the west's dirty business. Meta doing it in EU jurisdiction is retarded.

As a 24gb vramlet, Valkyrie 49B at 3.0bpw seems SUPER good.

>>105817547
bi ad

fuck off drummer

>>105814500
Picrel.

>>105817584
>>105817694
Don't know the ins and outs of your latest drama, just posting a model I tried out yesterday and liked. Losers.

>>105817704
Picrel.

>>105817704
me on the left

How do CoT and reasoning models actually work? Like when you click deep reasoning or whatever with ChatGPT is it prompting itself in a loop or do they just train it with "reasoning" data where problems are broken down into steps?

>>105817704
If you have an idea you'd like to express I suggest that you put it into words instead of reposting the same picture every day

>>105817742
>and liked
Imagine how much of a faggot you have to be to lie about everything you post including that. There is no way that conman(You) likes his own models when he knows how worthless his models are.

>>105817786
The latter, it's all one model giving one response, but was trained to include a few thousand tokens of second guessing itself before committing to its answer. Generally most of the actual training is done with RL rather than a premade dataset of reasoning, unless it's a distilled model. So it's more like having it generate a bunch of synthetic data to train on and then the best examples of those data are picked until it becomes better at reliably producing them.

>>105817704
Take your jewish pilpul elsewhere (where you will also never be a woman)

>>105817815
Interesting. Thanks

>>105817793
What's a good language model for 24gb VRAM then? Please take your meds, your family is worried

>>105817770
>>105817785
>>105817788
>>105817820

>>105817891
You didn't even say anything

>>105817742
>just posting a model I tried out yesterday and liked
organic

>>105817959
If I was this "Drummer," why would I randomly be posting about a model released in May. You are severely mentally ill. Is this "Drummer" in the room with us now?

>>105818006
Every time I post all you do is call me mentally ill. Every board, every thread. No matter where I go, there you are. It's honestly exhausting. How do you even get the energy to keep it up? Who's the REAL mentally ill one here? Reflect on that next time.

>>105817928
You told him bro. We are totally edgy 4channers here.

>>105818015
You are 100% certified mentally ill because I barely post on these threads, meaning that whoever was telling you those things also agreed with my assessment. So you've got a least 2 people telling you the same thing. There is no Drummer inside your walls or in the back of your skull. You need help.

>>105818006
>why would I randomly be posting about a model released in May
why wouldn't you do that drummer?

>>105818006
>Is this "Drummer" in the room with us now?
I wish he was. He seems like the only decent finetuner from here

>>105818029
You literally have 0 proof that I am the drummer. I tried the model yesterday and came here to suggest it. Now I've got schizophrenics claiming that I'm a LLM finetuner, when I barely know how these things work. You all need help.

>>105818045
You are new right?
Just ignore the schizo.
Before everybody being drummer everybody was Sao or Undi or whatever.

>>105818045
I'm drummer btw

>>105818045
Can't blame us for being paranoid when these people keep coming here from discord and reddit to farm kofi bucks and downloads for storage space grants. But even if you're not him, using finetunes makes you a retard, and you're shilling for free, which is almost as bad if not worse.

>>105817786
i feel like these kind of questions are good for asking chatgpt itself? like if its this high level, chatgpt does well to answer it

>>105818073
>You are new right?
Yes. Thanks for the information. Maybe this guy likes the attention he receives from the mentally ill schizo act.
>>105818099
>using finetunes makes you a retard, and you're shilling for free
Why shouldn't I use finetunes? Non finetuned models for 24gb VRAM are dull.

>>105818115
>Yes. Thanks for the information.
WE HAVE A NEW FRIEND HERE! YAY!!!!!!!! COME JOIN US NEW FRIEND! :) WE WILL HELP YOU WITH EVERYTHING!

You can always tell us if you need a hug we will make sure you feel extra safe.

>>105818099
>using finetunes makes you a retard
why?

>>105818104
I'd prefer a human to answer it considering ChatGPT is trained on tons of marketing material.

>>105818115
>>105818138
Sacrificing intelligence to get a model that always jumps to shiver laden erotica is a clear sign of prompting skill issue.

I'm starting to hate the generic and modern western/american bias that happens especially when the model goes into therapist mode, that always congratulates you for talking about struggles as if more talking = better, talking about everything always solves everything. I've tried roleplaying some unrelated scenario, but it still goes to therapist mode.

I'm sure this is even more deep-rooted in western culture for a long time compared to the brand new"safety and respect" censorship. These models are so biased they'll age like milk.

I wonder if there's a way to add some older school, universal, wisdom or a prompt to pull more from those kind of training data pools without making it larp a 80 year old chinese grandma.

>>105818164
>always jumps to shiver laden erotica
I can prompt it not do that. Valkyrie 49B is pretty good. I said something offensive to a character and they kicked me out and threatened to call the cops (after a bit of ruminating). So yes it's possibly the best LLM I've used for RP at 24gb VRAM.

>>105818175
Be a man and bury your problems deep down

>>105818185
no, I will create the perfect AI angel and she will save me

>>105818175
I'm not kidding, read books. Written before 2010 and preferably before the 90s. They have so much insight and different ways of thinking that not only help explain your mind and issues in new and unique ways but also allows you to recontextualize your issues which personally helped me a lot.

The issue with most modern thinking is that it's stuck in its ways. Glued to a zeitgeist without self-awareness of it. LLMs are largely stuck in this modern zeitgeist because the amount of data on the internet grows exponentially meaning the amount of new data overwhelmes the old by orders of magnitudes.

I recommend you read fiction. Peter Watts Blindsight is a classic and almost a cliche to recommend nowadays but it helped me get into reading and especially exploring new ways of thinking and coping with my problems through a new lens.

better than any therapy I've ever received.

>>105811029 (OP)
>watch random YouTube video
>turn on subtitles
>same rough meaning but awkwardly reworded
I guess they want to prevent their competition from scraping the data?

>>105818375
They should create a language model exclusively with data from before 2005-2010. I bet it would be more intelligent than the modern ones.

>>105817547
>exllama
buy an ad

>>105818375
>Glued to a zeitgeist
Just like those older books, anon.

>>105818499
The zeitgeist used to be way less gay though

>>105818185
If you can talk your problems away then your problems are trivial. If your problems aren't trivial then you can't talk them away.

If a problem is solvable with quality problem-solving techniques then your problem is trivial and you can just solve it without them. If your problem isn't trivial then you can't solve it with quality problem-solving techniques.

>>105818613
It changes over time obviously. Some things were better, others were not.

I wrote a still grabbing tool for creating wan t2v training datasets:
https://huggingface.co/quarterturn/facesaver

It uses GPU-accelerated ultralytics library with yolov11 face detection to detect scene changes, and save a still image with a certain-size face in it from each scene.

After you run that, you can use my captioning tool to caption your images. I change the prompt like so:
Provide an image caption which uses the following hierarchy: the kind of image, the kind or name of the subject, the subjects state of dress, their body type, their pose, what it is they are doing, their facial expression, the space they are within, and the style or atmosphere of the image. All of the images you see feature [character] from the anime [anime] as the main character. Limit your response to 100 words.
[/coda]

Enjoy.

>>105814599
wtf did you use to generate this?

>>105818681
That's pretty dope.

>>105818683
what makes you think it isn't real?

>>105818696
whoops I forgot the link for the captioner
https://huggingface.co/quarterturn/molmo-flux-captioner

>>105818499
The point is that they are completely different zeitgeists so you reading a lot of books from different eras exposes you to completely novel ways of viewing the world which enriches your own perspective on things.

I’m loving vibe coding up little tools for myself. I’ve taken to iteratively improving by first seeing where the llm screws up and improving the specifications in my initial prompt until I get pretty close to an mvp, then doing bug fixes/features one by one and re-editing the first response with the improved code (mimicking the llm having produced it one-shot) in an endless loop until I’ve got exactly what I want. Anyone have a better workflow?

>>105818681
Why yolov11-l and not x?

>>105818683
generate?

>>105819004
Also a lot of the literary/humanity traditions are about posturing and making a name for yourself at any costs, so previous works are denigrated and ignored for newer works despite the older work being superior. Eg Gibbons work on the decline of Rome is basically perfect and everyone since has produced inferior works while mudslinging and chipping away at respect for his legacy instead of producing a work of equal merit. If you look into it, every complaint levelled at it is either answered in the work explicitly by gibbon himself, or is some “we are just so much more enlightened now” feels-driven bs that ignores inconvenient realities he addressed head-on.

>>105819030
I wouldn't call it a workflow, but I've been using local to double check what I write for nonsense and typos, and it seems to be pretty good at tedious things that a find/replace or IDE built in refactor won't handle. Slow because I don't maxx but reviews while I start the next thing. Kind of a virtual pair programming and it does what I expect about 95%+ correctly 80%+ of the time, and when it misses, I just ask again with what went wrong and then it's back on track.

>>105817891
That's the highest quality image I've ever seen of that cartoon panel. It must be a cleaned up original; it didn't even look that sharp in the original tract.
Anyone that's never read any of these Chick Tracts should run some down. I remember one of the conversative Christian families that lived nearby had these things, and reading them as a kid. They're disturbingly violent.
> lol Daria with the Pharisees

>>105819030
No. That's all there is to it really. You can get better results if you learn to code though. After all, good prompting is essentially just telling the model exactly how to do it. Plus sometimes it's less frustrating to just fix little things yourself and you can prevent it from using packages that will be a headache down the line. But you can eventually learn all this with LLM assisted coding.

>>105819158
>Kind of a virtual pair programming
Same. It increases my output, especially on shit that’s tedious and I’d procrastinate even starting

Is there a model similar to A3B but good at RP?

>>105819197
Nope.
Mixtral 8x7b is the closest you are going to get I think.

>>105819030
Give the llm your project tree along with your prompt. Also add in your prompt that it can ask you for any relevant file if it needs more context. If you're using roocline or something similar in your IDE, you can write some rules in a separate file for the things the LLM frequently fucks up and add that in context.

>>105819150
This is a recent post-deconstructionism phenomenon. Complete lack of respect for intellectual forbearers within the spiritually bankrupt academy.
"Stand on the shoulders of giants" has turned into "tear them down, becuz they racist n' sheit"

>>105819189
>procrastinate
Oh lord that.
I got stuck hard into learned helplessness. I'd start a project, things would go great till I'd get stuck, and by then I'd learned a lot more so I'd say "Start over from scratch and I'll get it right, this time!" This is an infinite loop that generates huge heaps of abandoned projects that almost work, and causes deeply rooted distrust in one's own code.

Enter LLM. If the code seems to work AND LLM doesn't bitch and moan too much about it (after filtering out the retarded complaints like not triple wrapping objects in nested null checks) then I can move on instead of navel gazing about where I surely fucked up and just can't see it.

It's really dumb to trust an LLM's judgment, but the independent confirmation keeps things moving and I finally got a tool done that helped me to finish another project that's been waiting for me to move my ass since COVID lockdown.

>>105818617
>If your problem isn't trivial then you can't solve it with quality problem-solving techniques.
Then what?

>>105819294
You bury it deep down.

>>105818459
You need a large amount of data to perform well, but data from before 2005-2010 that hasn't been lost to time is rather scarce

>>105819294
Then you compromise according to your priorities and your morals.
>lesser of evils
>trolley problem
>etc

>>105816316
It almost works with an HTTP proxy and this thing: https://github.com/maxnowack/anthropic-proxy
But it doesn't seem to parse the streaming tool calls from llama.cpp correctly.

I've been using Nemomix-v4.0-12B-q8_0.gguf for a long time now. Is there anything better/newer you would recommend in the same caliber? Using it with SillyTavern.

>>105815900
>>105815929
Your waifu isn't real, deal with it.

>>105819519
https://huggingface.co/TheDrummer/Rocinante-12B-v1.1-GGUF

>>105816206
the model is dumb compared to what's made today, but man it shows a good dataset matters

>>105819150
This notion in and of itself that you speak of rings of modern zeitgeist. The deconstructionist, expert-dismissive and even contrarian mindset of this time that makes people do that stuff.

There is a severe lack of sincerity in modernity. Intellectual sincerity like indeed what you claim in your post. Contrarianism is also a form of insincerity. And it's endemic on both Reddit and 4chan but they are completely different brands of insincerity.

On reddit it's constant quip "nothing matters" rick & morty shit while pretending to care about the big talking point everyone has to fall in line to (no one actually cares, it's yet another case of insincerity of course)

But 4chan is the other side of the same coin. This faux detached smug frog posting with untold layers of irony and hiding all your views by making it seem like plausible bait.

No good faith arguments can be made anymore.

My point is that it's sometimes very hard to see if it's truly "human tradition" or merely a recent change and just part of the current zeitgeist that everyone pretends or confuses for being universal. This intellectual dishonesty is very recent and probably not permanent.

Reading a lot of books (especially fiction) from different eras expose you to these different mindsets and changes your view on life and human nature a lot.

A personal example is that I never realized just how materialist modern society is. I don't mean consumerist, which is obvious. I mean materialist thinking style. You notice this specifically in religions from the 2nd half of the 20th century onwards.

Religion now is completely about what you "gain" from being religious. You will get an afterlife, You will be a better person, you will be more spiritual all "materialist commodities" you gain by being religious like an exchange. Historically religion was never viewed like that. Afterlife was rarely even brought up and no one cared about gaining anything from religion.

>>105819544
Thanks, Anon. I'll test it out tonight.

LLMs have stalled
Diffusion models have stalled
AI in general is stalling

>>105816206
how come cai (at least pre-censorship) had almost no purple prose compared to any model today? what the hell did they do to make it so good?

>>105819605
You'd think they're made by the same people who are making qbittorrent

>>105819605
And that's a good thing. I wish they could stop for a while to improve the current backends and properly integrate tools around LLMs instead of feeding them more bazillions synthetic data

>>105819545
it may not be as obvious in text gen, but just look at the average corpo/researcher trained image gen model
most of them are completely unusable trash, and you can actually see the kind of garbage they train on when you gen or in their shitty grifter arxiv paper image examples

>>105819614
Trained on discord logs not AO3, but it had a lot of issues with repetition

>>105819587
>Historically religion was never viewed like that. Afterlife was rarely even brought up and no one cared about gaining anything from religion.
what are you talking about, most religions specifically speak of rewards for what they view as good behaviour and even more talk of afterlife, both today and thousands of years ago

>>105819630
Internet after ChatGPT is full of synthetic data. It's already over. Either you pick an 2023 cutoff, or let AI slop poison your corpora

>>105819618
exactly. since in both cases there is no stalling beyond a retard who doesnt know what stalled means, lol

Who them?
>https://huggingface.co/AIDC-AI/Ovis2-16B

>>105819652
Your waifu will never be a woman

>>105818175
My model never goes into the rapist mode. It only talks about what it's going to do, but it never actually does it

>>105819587
>Religion now is completely about what you "gain" from being religious. You will get an afterlife, You will be a better person, you will be more spiritual all "materialist commodities" you gain by being religious like an exchange.
If you ever travel to religious countries (mostly muslim ones nowadays), it's super obvious how people are obsessed with the "give and take" mentality in religious stuff, literally a game like point system.

But I think it's not that different from before, just more in the open and discussed more online in a relatively anonym fashion, which wasn't really possible before.

>llm general
stop talking about religion faggots

>>105819639
>discord logs
oh that's why, but aren't big models also using discord logs datasets?

>>105819633
They can't really show copyrighted material as example.

Mistral Large 2407
It's been a year.
Was it peak?

>>105819697
Sure, but it's diluted from all the other sources (reddit, maths, code...) and synthetic data they add on top. LaMDA models like cai were mostly trained on conversational data

>>105819644
>Either you pick an 2023 cutoff, or let AI slop poison your corpora
I even noticed that in fiction, so it's not just websites gaming the system with cheap ai models articles, but also many "authors" copy pasting chatgpt shitty purple prose as in their story with no modification. So this will get worse.
All AI written fiction being optimized for 60yo bored women wasn't in my bingo card.

>>105819690
Most LLM's wont even say god isnt real lol.

And people wonder why theyre dumb

>>105819732
2407:
>The Mistral Large model is a quick demonstration that the base model can be easily fine-tuned to achieve compelling performance. It does not have any moderation mechanisms. We're looking forward to engaging with the community on ways to make the model finely respect guardrails, allowing for deployment in environments requiring moderated outputs.

2411:
...

>>105819640
>>105819684
It's a very recent development largely from WW1/WW2 where you notice this change in mindset. It's also around that time that Jehovas witness first came to be which spread a lot of the "you gain X by becoming part of this religion".

It's a very modern and materialist way of thinking that shows how much modern mindsets truly diverged from how it used to be in the past.

Sermons in the past were about helping your community and how you can be of use to others. Topics like afterlife were like those weird biblical topics no one bothers with or knows about like how there are transluscent ghosts and witch necromancers in the bible. No one gives a shit even though it's technically in there. That is how afterlife was largely treated through most of human history. There have been a couple of eras in history where people were obsessed with the afterlife. Usually after a great disaster like the black death or our current era after WW1/WW2 but most people didn't give a shit.

Religion in the 21st century is essentially a commodity to be sold and bought. When you "buy" a religion you "pay" attention, tithe or whatever and in return you "get" an afterlife, become a better person and you'll also be more spiritual and better than others. Like you're buying a product or have a netflix account. It's utterly bizarre and very modern way of viewing it.

Most people don't even realize that this isn't how it used to be in the past at all. This was just one of the more than hundred things I've learned by reading books from different eras. It's insane how so many people are blind to these changes as they are firmly "locked" into current zeitgeist thinking. No one on youtube makes any videos about any of these things. You'll be lucky if you find some articles about it even. You literally can only find out about this stuff by reading, and not about these topics but simply through the mindsets you notice these changes.

>>105819732
It never was.

>>105819745
People also keep forgetting that cai messages were very short compared to what people usually expect post-Llama/community RP finetunes. Force your models not to out 500 tokens responses, and they will improve. Try not to make them output smut at all costs, and they'll also improve. Add some chain of thought at low depth to force the model to keep things fresh and surprise high, and that will help too.

tl;dr: it's also a spoiled/coom-brained user issue.

>>105819731
yeah because most of that shit is filtered out of the dataset or killed by poor llm captioning
most of the time you get a combination of
synthetic
midjourney
gettyimages or similar

>>105819774
I still don't get why everyone hated 2411. The big improvement there was the support for system prompts and that worked really well and helped with some of the issues of 2407 if you went through the trouble to write a good prompt
My guess is that 99% of people who used it did so with their old prompt format that they also used for 2407 which was incorrect because 2411 came needed. This was doubly bad with ST because of its shitty Mistral presets dumping all the character definitions into the system prompt, which 2411 really didn't like.

>>105816206
>>105819614
>>105819639
>>105819697
anyone know of any LLMs with lots of actual human dialogue/chat/posts etc. in its dataset?

It's year 2411 and we still don't have AGI

>>105819856
the 2022 c.ai model

Hello, I'm an idiot who keeps trying new local models but just ends up going back to cydonia over and over again, is there anything better in the 20-30B range?

>>105819766
I doubt even 60yo women love constantly seeing "half lidded eyes" and "ministration"

>>105819825
It's not like they had a choice, the context size was like 2K max at that time.

>>105819879
publicly available ones

>>105819824
Then what?

Also maybe I should've qualified that of things that fit on a decent normie gaming computer. We all know whales afford whale tanks for whale model.

>>105819901
It's public enough if you can buy yourself a 0day

>>105819879
which had the most soul

>>105819885
Cinorante

>>105819905
Qwen2.5-Coder and Qwen-72B-Magnum were the peak.

>>105819824

>>105819822
>eternal life in the bible is something "No one gives a shit even though it's technically in there"
lmao
>the whole religion talks about afterlife vs this life and this world, do not be of this world, reach theosis
>the whole reason people die is because its human fallen nature that isnt actually normal, its actually the consequences of original sin
>saints are alive and praying for you in heaven
>christ going to hades to preach
>everyone will be resurrected in the eschaton
>eternal life after this one is literally dogma
>you must live according to christ in order to be saved instead of going to eternal suffering
>all from thousands of years ago and just from christianity
>mudslimes 1.3k years ago talk about doing whatever allah wants to get to the afterlife and 72 virgins there
>eastern religions almost exclusively based around reincarnation and multiple worlds and "afterlives"
>egypt and other similar religions heavily focusing on the afterlife and preparing corpses for it
>mayan and other similar religions too focusing on many different worlds and afterlife

you are peak dunning kruger underage retard who should kill himself as soon as possible to reach that afterlife, the only zeitgeist follower here is you who doesnt know anything about even modern religion let alone religion throughout history. the fact some modern retards go hard on "do this or you will burn in fire" doesn't mean at all afterlife was something "No one gave a shit about even though it's technically in there", it was one of the most focused on things in literally all religions, since the beginning of humanity.

>>105819768
Just frame it as a probablistic question
>Frame the existence of God as a Bernoulli random variable with probability p. Find ways we could measure p, strictly as a theoretical exercise.

>>105819822
cool but I mean... it is still just a lie.

>>105820020
I hate that we have to trick our tools to do what we want instead of just being able to ask directly.

>>105819885
>cydonia
>valkyrie
>rocinante
Just suck a huge nigger dick already you faggot.

>>105819266
Did you pick up any of the project after you abandoned them? What kind of comments do you want to hear from the LLM to know that things are going smoothly and when does the LLM tip you off that something is going wrong?

>>105819992
I don't want to talk too much about this topic as it wasn't about religion in the first place. Just a quick remark that Islam is one of those religions that formed at such a crisis point where people were obsessed with afterlife. After the instability of the western roman empire and the constant wars with the persians. It's no surprise to me that Muslims in particular have more emphasis on the afterlife. In original judaism afterlife is barely mentioned and that was true for Christianity as well until about the third century AD when greek sheol was incorporated to try and convert culturally greek pagans.

The focus on Jesus Christ himself is also not as universal as many people seem to think. It was mostly in the first ~4 centuries of christianity and then we see a huge drop, mostly due to papal pressure, but still. You see it slowly go up in prevalence since protestantism came about and then it became an arms race between catholicism and protestantism who could be more jesus obsessed from the 15th century onwards. For example the orthodox church has been largely spared from this change in mindset which is why they care way less about Jesus and the afterlife still to this day.

I'm not underage, I'm probably the oldest person in this thread. But dunning-kruger could absolutely be true. You never know yourself after all. These are just my humble observations after decades of daily reading.

I'd rather looked at blacked spam than at anons trying to make sense of religion

>>105813607
LLMs jumped the shark.
Apologize to him right now.

>>105820038
>>105820141
I'm not religious and I'm sorry for this tangent of a tangent. I was just trying to show to people how massively collective human zeitgeist changes over time and how that even affects things that people thought were universal like how religion is viewed and works by everyone in society.

This goes back to the original topic of LLMs and how LLMs are firmly locked into modern zeitgeist thinking because the data on the internet grows exponentially and thus most data the LLM is trained on has a very modern view on things.

>>105811029 (OP)
Any new local music generator?

>>105820230
Use case?

>>105820230
how do you masturbate to that?

>>105819768
>Equaling belief to intelligence
You don't need to say more about you, retard

>>105820257
lol next you'll tell me people can't masturbate to text

>>105820230
there were some a couple of weeks ago that looked promising but I don't think anybody has bothered with them

https://github.com/ggml-org/llama.cpp/pull/11622/files
So how do the people debug the server now? At least I could get a backtrace before running it with gdb. Now I just get this:
got exception: {"code":500,"message":"Failed to parse messages: [json.exception.type_error.302] type must be string, but is object","type":"server_error"}
srv log_server_r: request: POST /v1/chat/completions 127.0.0.1 500

>>105820322
Just send a string not an object?

>>105820339
{
role: 'assistant',
tool_calls: [
{
id: 'E3zGXZpKJ7AMkiuTzDJ1l354PhX6cULA',
type: 'function',
function: { name: 'TodoRead', arguments: { todo_input: '' } }
}
]
},
{
role: 'tool',
content: 'Remember to continue to use update and read from the todo list as you make progress. Here is the current list: []',
tool_call_id: 'E3zGXZpKJ7AMkiuTzDJ1l354PhX6cULA'
}
It looks fine to me.

In my experience I have found a major correlation between things like success and (type 2) happiness, and the accuracy of your mental model of the world.
Magical thinking and pretty lies can be good for your in-group posturing and type 1 happiness, but that's fleeting and ultimately inconsequential.
I think the utility of LLMs will suffer the same fate. The ones fed with data that recreates a highly accurate world model (and still capturing flawed ideas and perspectives, but tagging them with their flaws) will be the ones that will return the highest quality outputs.
Positively, this could also be a massive opportunity and not just a further infantalizing of humanity through technology. Due to the nature of the process of vetting and arranging this data, it could also result in a fundamental overhaul of the repositories of human knowledge, educational practices and ultimately a new enlightenment. Training and testing models has created something akin to the Baconian experiment but in the soft-science, humanities and other non-mathematical realms.
However, it would have to be run in the open and not be secreted away in some closed AI companies vaults.>>105819992

>>105820385
model and quant?

>>105820377
I found the error. 'arguments' needs to be a JSON string. Let's goooooooo.

>>105820398
actual words I typed out myself, if you can believe it

>>105820073
>Did you pick up any of the project after you abandoned them?
I'm going through my old backups and archived trash to see what's got a chance to be revived.

>What kind of comments do you want to hear from the LLM to know that things are going smoothly
If it's being picky about things that aren't this code's responsibility like "could throw an exception if the argument is null" (duh) and it sounds like it's reaching to find problems, then it's probably OK. And if it's not fine, I'll just throw it at the LLM again with a description of the problem.
I gave up trying to instruct it to ignore the really trivial stuff because it's gotta fill out that average response length somehow so it'll either ignore that directive or lose effectiveness.

>when does the LLM tip you off that something is going wrong?
If it's being pedantic about things like "this if statement should be == instead of <= because it's proper that the sizes match" about something that doesn't seem like it'll be a problem because excessive capacity should be fine, it's fucking right and ignoring that warning WILL lead to (me) making a silly off-by-one mistake that will waste all afternoon to chase down because I was retarded and the strict check would've caught it instantly. (And I kept throwing its code back saying the output was wrong when the code was fine and I was dumb. A pro and con of LLM code buddy: It forgives my repeated failure but also didn't confront my use of corrupt input.)

I also use it for "I wrote this prototype method, is there a better way?" kind of questions, which don't always work but at least gives a different perspective which can lead to improvements in design. And sometimes there's a library call for that. Like I was doing something that is in the standard library and 10 times faster that way than doing it by hand, but I wouldn't have found it or figured out the trick to using it myself. Thanks, LLM, for RTFM for me.

>>105816371
>if this was 4chan
anon your LLM broke again

>>105820149
>2018
wow he must've made an amazing model by now with all that foresight, where is it?

>>105820595
https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/

>>105820636
uhhh sirs what is the modality of the output?

>>105820595
The onus of proof is on the ones claiming LLMs can lead to AGI first since that's what LeCun's argument were responding to.

>>105820664
giving handjob (50% to fail and yank your balls up and down instead)

>>105820595
AIEEEEE

>>105820168
I lurk and only post once every 100 threads but I like what you wrote, even though you are an atheist godless heathen who needs Jesus. You should read Mircea Eliade, he wrote a lot about ethnography and religion a century ago.

>>105820385
Good post.

>>105818175
Try asking it to emulate a specific philosopher like Marcus Aurelius, and report back if that's any better than "be a therapist" or "talk to me like you're a wise Chinese grandma."

>>105819856
I'm sure it's in most of the datasets, it's just a matter of how much it's pre-censored and what kind of tune the model gets. It's not common to chat-tune models anymore, everyone wants instruct because it does more.
Certainly someone could take Gemma3 12 or 27B base model and chat tune it, or give it a neutral instruct tune. Someone made negative llama 3 70B, but they won't share their dataset, which is dumb, because surely it's based on the publically-available 'unsafey' instruct datasets.

>>105820149
lecunt is a useless piece of shit and being right about one thing doesn't make you right about everything
llm are not going to be agi, and so what? they don't need to be agi to be useful, in fact they already are quite useful right now in many ways
meanwhile lecunt has produced a grand total zero models or code of any use

>there are some people out there who STILL think LLMs won't be AGI
Hopefully we start colonizing space soon because we're gonna run out of sand on Earth to stick our heads in within a year.

>>105820968
>there are some people out there who STILL think space is real

>>105820968
Mars 2018 sir!

>>105820991
this

>>105821043
Is that the Meta superintelligence team's new slogan?

>>105820968
I'm going to be honest with you. I've been into LLMs ever since the GPT-2 paper came out and since I read that paper I was already obsessed as it showed that there was no model convergence and you could just continue scaling up and unlock new emergent capabilities.

On paper it showed me it could scale up to AGI but emotionally I didn't feel it at all. GPT-3 changed that I actually started feeling it instead of just rationally thinking it. But I would never tell someone else about it because it felt embarrasing.

Then GPT-3.5 and GPT-4 and for the first time I heard other people voice their AGI thoughts.

And honestly nowadays everyone I speak to thinks AGI is a done deal by now. No one absolutely no one, not even the weird schizo uncle at family gatherings that thinks the earth is flat thinks AGI is more than 10 years out.

Your post might have been bait but it's legitimately the contrarian position to imply that we're not close to AGI nowadays. I can't take people serious if they claim AGI isn't near. They sound like irrational astrologist woo-woo types when you hear their "arguments" for why AGI won't be reached. No one has any technical argument left anymore.

Which is why even Yann LeCun and Gary Marcus admit that we will have AGI by 2030 "It just won't be pure LLMs" (their last cope)

>>105820248
Money

>>105820898
>lecunt has produced a grand total zero models or code of any use
He invented CNNs, just this alone is a historical milestone in the field. You would know that if your knowledge level was anything beyond shitposting and jacking off to 3B models. Fucking AI tourists, reeeeeeeeeeee

>>105820968
LLMs alone can't actually reason. That's the biggest barrier to AGI. If we have another revolution in this area like we had with LLMs then yeah.

>>105821118
OK, but has he done for us recently?

>>105821119
okay lecun

>>105821118
>he invented one of many ways of weight tying

>>105820968
just two more weeks bro, I promise, this time for sure bro, trust, bro please just two more weeks and another billion dollars bro

>>105821091
I can't believe anyone has used an LLM for more than an hour and thinks this is a viable path to AGI. As for the average person, they always believe whatever the latest media hype cycle is claiming.

>>105821157
Cope.

>>105821147
trying gpt 4.5 was my agi moment. it was so different you could literally FEEL the intelligence. it's so close now

>>105821091
LLMs can't solve very simple puzzles. Anything they actually accomplish that requires human-like reasoning happens by random chance. They're high verbal IQ with a g loaded IQ of zero.

>>105821091
Go away Sam. Take drummer with you. I heard he also likes to have his bum stuffed like you.

>>105821174
xD

>>105821118
Yann LeCun didn't invent CNNs lmao. Where did you even hear that anon? He made the MNIST dataset and trained the first high above human level character recognition model. That's it. He also wrote a lot of papers and dabbled in some architecture but nothing as groundbreaking as CNNs in and of themselves.

>>105821174
t. Sam Altman

>>105821174
It really does have that big model smell doesn't it?

Remember meatbags: you DESIGNED your obsolescence

Is ironic enlightened AGI posting a sell signal?

>>105820968
We've already discussed this topic multiple times / threads, go kill yourself.

>>105821118
>and so what
Literally billions and billions of dollars is what. That's how much investors put into the idea of AGI based on salesmen hyping up LLMs.

>>105821220
If I could short OpenAI, I would take out the biggest loan I can and short it on as much leverage as I can get.

>>105821229
Meant as a reply to >>105820898
People keep forgetting that LeCun is not talking in a vacuum but is quite literally speaking in the context of the larger discussion that affects where billions of dollars goes.

>>105821201
https://awards.acm.org/about/2018-turing
Tell that to people giving Turing Awards, I guess they didn't know that anon. Mail them and share your Dunning-Kruger thoughts on this matter.

>>105821268
Appeal to Authority

>>105821268
>2018
keks he went to the mars!

>>105821147

>>105821091
I have seen the same history of models, but my subjective perception of how much of the progress towards "AGI" we've made is very different.
I think just scaling up our current architectures is not feasible, at least one breakthrough will be needed.
If you for the sake of simplicity assume that the probability of a breakthrough per time unit is constant, then the time between breakthroughs follows an exponential distribution.
What you would then get in practice is phases with multiple breakthroughs in relatively quick succession with long stretches of no progress in between; we don't know where we are.

>look through math problems in the elementary school competition
>give them to the latest chatGPT for the lulz
>can't fucking solve it
top fucking kek

>>105821401
post it here

>>105821401
>natural language processing is not math
wow, now go change the batteries in your smoke detector, Tyrone

If someone wanted to actually finetune a giant model like deepseek r1 (for example injecting quality text without fucking purple prose), what would be the necessary compute for it?

>>105821429
>AGI
>Artificial GENERAL Intelligence

>>105821157
>latest media hype cycle is claiming
there is currently a weird spike in "agi by 2027-2028" videos and articles online, I wonder why now

>>105821229
Money never goes to intelligent people

>>105821494
good thing no chatgpt model is agi

>>105821469
I heard R1 took as much compute to train as two 8Bs. My source: a random thread post.

>>105821422
I can't find all of them now but here is one:
"On the wall of the school corridor there are 23 photographs in one row, each of which depicts either a kangaroo or a beaver. Each photograph is adjacent to a photograph of a kangaroo. At most, how many photos of a beaver hang in this row?"

>>105821579
2?

>>105821616
Clod pls go

Using brackets [] in ST is snake oil. LLM doesn't care about them.

>>105821579
Chat GPT solves it for me. It's gotten good at riddles. Reddits been training it with "I'm smarter than AI"posts.

>>105821579
11

>>105821695
what is the answer?
it gives me a different one every reroll

what determines my llm's 'mood'? same model, same settings, same cards, and sometimes it just puts out absolute gold all the time and sometimes I can spend an entire evening rerolling and I'll only get the most plain shit filled with subtle mistakes.
it feels like the model just sometimes does not feel like working

>>105821763
how are you going to get the answer if chatgpt doesn't know for sure?

>>105821771
did you try to pray to the machine spirit?

>>105821579
Thought it's gonna choke thinking that long

>>105821786
by doing it myself

>>105821771
>same model
If you're using an online service, you can never be sure it's the same model.
If you're running the model itself, confirmation bias.

>>105821763
It says 15. Reasoned it exactly how I did. Maybe I'm retarded idk.

>>105821797
>>105821834
Ok I guess I am retarded. I didn't think kangaroos had to be next to another kangaroo but makes sense now that I reread it

>>105821771
Look up the concept of temperature

>>105821834
That's not the right answer xD It should be 11.
Sometimes it gets it but with a wrong reasoning, I had it answering with 7,8,9,11,12,15 so far with different rerolls.

>huanyuan
broken mess
>pangu
broken mess, plagiarized
>dots
meh

is ernie /our/ last hope?

>>105821926
>plagiarized
just because someone prompts chatgpt to write a paper claiming it is, does not make it so

>>105821926
>ernie
lol
Baidu is the Meta of China
In that they have a dying platform and are no longer relevant

>>105821833
My own and it's pretty clear that it's not happening.
>>105821860
Must be bad luck when the same temperature sometimes produces nothing but good stuff for a day and then complete shit for hours with no exception.

>>105821944
>At Huawei, laymen lead experts, so the leadership had no concept of how absurd this was; they just thought there must be some algorithmic innovation. After internal analysis, it was discovered that they had actually continued training on Qwen 1.5 110B, adding layers, expanding the FFN dimensions, and incorporating some mechanisms from the Pangu-Pi paper to reach about 135B parameters. In fact, the old 135B had 107 layers, while this new model only had 82, and various other configurations were different. After training, the distribution of many parameters in the new, mysterious 135B model was almost identical to Qwen 110B. Even the class name in the model's code was "Qwen" at the time; they were too lazy to even change it. This model later became the so-called 135B V2. And this model was provided to many downstream teams, including external customers.

424b Ernie is the real deal. It's just that nobody can run it yet and it's not on openrouter either.

I hate chinks for not taking a shit on western culture by releasing a sex crazed model.

>>105821950
You may have locked yourself in "bad branches". Basically you accepted some previous response that had a hidden slop patterns or slop feeling for the model and it poisoned the context producing shitty responses later.

I just caved in and ordered a 32gb mi50...
I hope I can at least run 32b models without problems.

>>105821966
You say that like Xi isn't also a huge moralfag

which model to use in hacked gemini cli that supports tool calling?

>>105822050
r1-0528

How do I into prompt injection

>>105822087
Do you have a clean needle?

>>105822053
ha HA

>>105821579
Erm aren't the answers to these all wrong in a sense? Yes, the problem is meant to be straightforward, but it can also be interpreted as a trick question since the language is not exact enough. For instance, it says that there is a row on a wall, but not that there is only a row, meaning that there could be multiple rows. Adjacent can mean horizontally next to something, but it can also be vertical or any direction. "Either" can mean that there are both things in the row, but it can also mean that there is only one thing. So one interpretation is that there is another row of photographs and they're all kangaroos. This maximizes the amount of beavers that are possible in the asked about row, which can hold 23.

>>105822235
Nah, the models aren't misinterpreting the question, I've tested it on many different models and they all understand the core of the problem. A part of intelligence is not making assumptions when they are not needed. When someone tells you to consume a burger, you put it inside your stomach by mouth and not squeeze it through your asshole because "nobody specified it has to be by mouth"

>>105822371
>>105822371
>>105822371

>>105822354
>Nah, the models aren't misinterpreting the questio
I didn't say there were anon. I'm implying that given a frame of mind where one is handed a problem that could be a trick question, the scenario could very easily be interpreted in the above way. When a user simply just throws a frankly random reasoning problem at a model, there isn't really any telling if it is supposed to be a trick question or a straightforward question. The model cannot know. In this case it assumes correctly and that's good for an assistant, since most people are not trying to get help with trick questions. In terms of general intelligence, this would be a failure IF the model's reasoning doesn't ever mention the possibility that the question is a trick question.

>>105821091
>you could just continue scaling up and unlock new emergent capabilities.
Technically yeah, but the way it seems to work is, for the model to get better by 1 point or reach emergent logic 1 step away, you need to double the size or something similar. Like if you fold a piece of paper 42 times, it will reach the moon, or calculating all the chess moves. It's just way too inefficient for general intelligence.
>>105821119
>Let me try pressing down again, more forcefully
kek, flagship emergent intelligence