File: __hatsune_miku_vocaloid_d(...).png (195 KB, 591x591)

195 KB PNG

►Recent Highlights from the Previous Thread: >>102334890

--Papers: >>102345545
--Pixtral 12B multimodal model benchmark results and discussion: >>102340417 >>102340517 >>102340642 >>102340661 >>102340715 >>102340741 >>102340728 >>102340704 >>102340811 >>102340834 >>102340851 >>102341072 >>102341588 >>102341597 >>102342684 >>102342770 >>102342854 >>102342861 >>102341771 >>102341791 >>102341952 >>102342037
--Setting up and deploying a speech recognition application using fish-speech: >>102340638 >>102340839 >>102343659
--Improved training with higher learning rate: >>102343979 >>102344869 >>102345102 >>102345300
--Fish Audio TTS is getting praised for its quality and speed, with some users comparing it favorably to XTTSv2 and mikutts.: >>102335083 >>102335519 >>102336040 >>102340252 >>102340312 >>102346120 >>102340502 >>102341496 >>102342870 >>102342966 >>102342891 >>102342929 >>102342958 >>102342977 >>102343010 >>102343132 >>102343252 >>102343224
--Meta Platforms building US$2 billion H100 cluster for Llama 4 training: >>102345693
--Debate over timeline and importance of multimodal support in llama.cpp: >>102342323 >>102342381 >>102342403 >>102342501 >>102342689 >>102342666 >>102342756
--Anon discusses using Solar Pro with Phi special tokens: >>102337022 >>102337102 >>102337124 >>102337147 >>102337150 >>102337178
--Solar Pro Preview Instruct Nala test prompts discussion on prompt templates: >>102336620 >>102336692 >>102336722 >>102336758 >>102336787 >>102336839 >>102339019
--LLM360 advocates for open source models, but previous attempts lacked testing: >>102339206 >>102339420
--New llama1 finetune, Chronos-Divergence-33B, is non-slopped but has logical incoherence and self-contradiction issues: >>102345953 >>102346136 >>102346645 >>102346767 >>102347405 >>102346838 >>102346960 >>102347016 >>102347035 >>102347063 >>102347081 >>102347493 >>102347086
--Miku (free space): >>102335117

►Recent Highlight Posts from the Previous Thread: >>102334893 >>102334989

Mikulove

File: 1653875730964.png (493 KB, 1080x1036)

493 KB PNG

Pixtral looks promising. Even if its not, I'm looking forward to getting used to using multimodals. Hoping for exl2 or llama.cpp support soon.

>https://github.com/fishaudio/fish-speech
Did they google translate webui from chinese? It's too confusing for me.

LLaMA-Omni: Seamless Speech Interaction with Large Language Models
https://huggingface.co/ICTNLP/Llama-3.1-8B-Omni

File: 74529 - SoyBooru.jpg (520 KB, 2324x2993)

520 KB JPG

LLMs owe me sex.

Does any of the UIs let you just point to files/folders to add as context for the model? Are any of them capable of reading pdfs?

>>102349195
RAG is a meme. Try Open Web UI, I guess.

Why the fuck this is nowhere on their github https://speech.fish.audio/#linux-setup
So fucking easy to setup, actually

File: 1695769022205.png (271 KB, 590x400)

271 KB PNG

>>102349335
>those manual windows installation instructions further up
lol, lmao

>>102349400
> that pic
That's so stupid you could only come up with it BY reading the manual. If you just copy paste what everyone else does it will not happen.

>>102349400
Attempting uninstall: torch
Found existing installation: torch 2.4.1
Uninstalling torch-2.4.1:
Successfully uninstalled torch-2.4.1
Attempting uninstall: torchaudio
Found existing installation: torchaudio 2.4.1
Uninstalling torchaudio-2.4.1:
Successfully uninstalled torchaudio-2.4.1
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.19.1 requires torch==2.4.1, but you have torch 2.3.1 which is incompatible.

>>102349400
>blaming OS for jeetware's failure

how do I run pixtral?
https://huggingface.co/mistral-community/pixtral-12b-240910

>>102349526
Yes.

>>102349335
The question is, why should I install this shit? Not how.

>>102349503
This is the average python experience
So far I haven't run into any python jeetware that doesn't require me to go into the .py scripts and rip out huge parts of unnecessary crap that are making shit fail
That's why I always block internet access to python and never update anything because otherwise it will randomly break itself

>>102349560
SAAAR FISH CHINESE SAAR NOT INDIAN SAAAR

>>102349560
Each audio project is but a half-broken jeetware.

>>102349599
even worse then

>>102349541
is there code for Pixtral? The provided code is just for the tokenizer

    image = Image.new('RGB', (64, 64))
    # tokenize images and text
    tokenized = tokenizer.encode_chat_completion(
        ChatCompletionRequest(
            messages=[
                UserMessage(
                    content=[
                        TextChunk(text="Describe this image"),
                        ImageChunk(image=image),
                    ])],
            model="pixtral",
        ))
    tokens, text, images = tokenized.tokens, tokenized.text, tokenized.images

>>102349601
XTTS2 is the only one I managed to make work. I also had to modify some pyshit, namely to remove the license nagging every time you start it and also stop it from doing brain damaged shit like deleting EVERY file it's downloaded if one of them fails to download (ever heard of retrying after connection fail? apparently these guys haven't). Also they thought it was a great idea to download models into appdata instead of into its own folder. Lmao

File: 1711089553210944.png (78 KB, 1097x649)

78 KB PNG

oh hello llama

File: 1704036155948575.jpg (167 KB, 500x500)

167 KB JPG

>>102349503
>>102349560
you WILL install 100GB of dependencies in separate venvs for each and every Python-based project you want to run and you WILL be happy

...total pyshitter death WHEN?

>>102349840
The reason for python's lib hell is that pyjeets can't maintain backward compatibility. You won't need separate virtual environments if libs don't break compatibility with each minor version

>>102349840
What is the most lightweight installation you can have

>>102349887
the language itself regularly breaks compatibility, they don't give a shit

File: Screenshot_2024-09-12_20-42-26.png (3 KB, 324x28)

3 KB PNG

Is Fish supposed to be slow or am I doing something wrong?

>>102350294
And how the fuck is anyone supposed to know what you're trying to do?
Chances are that yes. You are doing something wrong.

>>102350294
Yes it's slow.

>>102350294
I'm surprised this is so slow, like this is "only" a 0.5b model right? it's even slower than Flux which is a 12b model

is local gpt-4 level yet or is it gonna be another 3 years?

>>102350294
It generates several sentences within a few seconds for me

>>102349840
it's really incredible how such a widely used language can be so consistently horrible with larger projects. every script past the third one you add doubles the odds of sending a user into dependency hell. septupled if they're unfortunate enough to be on windows.

>>102350553
3 years until asi invents a timemachine and you can go back to july 2024, yes

File: 1726104204571260.webm (1.1 MB, 1280x720)

1.1 MB WEBM

>>102348952
>/lmg

>>102350553
LLAMA-405b is GPT-4 tier as an assistant, but almost nobody here uses it because nobody here wants an assistant. Better question would be if we have local Opus. (No, we don't, and nobody competes with Anthropic on local.)

>>102350553
3 years to run models of that level on gaymer hardware, probably, yes.

>>102349526
Ooba > select transformers loader > have fun.

>>102350629
dunno, opus and gpt4 are pretty much the same in terms of being assistants. at this point, unless there's a significat breakthrough, we are way into diminishing returns territory

>>102350294
you need the --compile flag for 20x ish gains

>>102347493
>>Neutral samplers
>For what purpose? They recommended MinP of 0.05 to 0.1 with t=0.7 to make it work decently.
If a model doesn't' work with neutral samplers it's fucking brain damaged. Simple as.
None of this
>NOOO YOU HAVE TO USE THESE EXACT SAMPLER SETTINGS AND THIS EXACT PROMPT TEMPLATE
Within reason of course.

>>102350625
That's a lot of fingers.

>>102350656
I meant that most of the local is contaminated with GPTslop and is horrible at creative writing and RP(most popular usecases) because of it. Shit like shivers is an instant boner-killer and other GPTslop ruins SFW experience too. Of the smarter models, only Cohere's Command-r-plus(old, not 08-2024) was relatively slop-free, but Cohere was clueless about it and slopped it up in 08-2024 release. We're already on the level of (original)GPT4 in terms of smarts, but the writing styles are still behind Claude.

What's the current meta for iMatrix datasets?

>>102350718
Fuck, it's instant now. Thank you

>>102350776
>GPTslop
>shivers
That's just bad writing. If people didn't use datasets full of writing by women or prepubescent boys that wouldn't be an issue anymore.

>>102350725
>If a model doesn't' work with neutral samplers it's fucking brain damaged. Simple as.
I draw a distinction between brain damage and being undercooked. Being undercooked can be ameliorated with sampler settings; brain damage can't.

>NOOO YOU HAVE TO USE ... THIS EXACT PROMPT TEMPLATE
*You* are brain damaged if by now you don't understand that using a different instruct template than a model was trained on makes it drastically stupider.

>>102351018
he's right, if the model is so unstable it shits the bed on the default settings that's how you know it's a shit model

>>102351018
>*You* are brain damaged if by now you don't understand that using a different instruct template than a model was trained on makes it drastically stupider.
Cool strawman bro.
I'm not even going to correct you.
you're right.
You're the smartest poster on this subreddit.

>>102350776
Claude only does X, Ying. It has no styles. It's just more likely to bring up relevant concepts unprompted

>>102351073
>makes it drastically stupider
Was it tested on some mememark? I would expect there could be some slight additional retardation, but if a model can't largely generalize away the prompt format then it can't generalize at all? Also I would expect that in some cases just for coomer shit not using the template could be better. Because instruct template will be strongly connected to the censorship assistant persona.

How much power does an RTX 3060 draw at idle in a headless server? I plan to add one to my NAS, but I'm worried about its monthly power consumption.

>>102351041
That's not a fully unreasonable position. In an objective sense it means the predicted token probabilities do not match the true probability distribution of either the target language (for a base model) or "what a helpful robot butler would respond" (for an instruct model) and the difference is great enough to cause problems.

I've moved away from this view though. An LLM should be judged for its use as a tool or a toy. If there is a transformation that can be applied to its output that makes it useful, then it's useful. "The output isn't diverse" can be a problem. "I had to set the temperature above or below 1" and "I had to set min-P" are non-problems.

>>102349503
OK start over and this time create a conda environment for it. It'll make is a lot easier to troubleshoot. Sometimes you can ignore those dependency check warnings too, try it and see if it works.

File: file.png (14 KB, 698x108)

14 KB PNG

>tech media in my country is announcing the release of strawberry in two weeks
I'm like 90% sure this is the work of a bored intern trolling the editorial team.

>>102351270
I feel like it's been years they've been hyping this nothingburger strawberry, I'm not going crazy they started this trend at the begining of the year right?

>>102351123
>Claude only does X, Ying.
Wrong thread, you're looking for /aids/.

>>102351270
>have nothing to release while other AI companies are making news so you try to suck the oxygen out of the room with hype for an unreleased under-performing nothingburger
Sad!

File: file.png (29 KB, 385x655)

29 KB PNG

>>102349503
So I have those installed in my shitty windows. And when a program needs one of those it just gets it. And if a program needs one of the older ones and not the new one it gets the older one.... What is the reason you need to have local copies of all this shit?

>>102351280
>I'm not going crazy they started this trend at the begining of the year right?
There's been talk about strawberry since at least a year ago.
Then again, all the grifters on Twitter are going ham about having some kind of "insider info" (read: purposefully leaked information in order to create hype).

>>102351249
It's under conda. The problem with >>102349335 instruction is that
pip3 install torch torchvision torchaudio
installs 2.4.1, but they have
stable = [
"torch==2.3.1",
"torchaudio",
]
in pyproject.toml, they have no idea what they're doing. They should just add torchvision to their toml

Question about aichat and local models.
But what should I do about jailbreaks on a local model? I haven't really tried many. I've been using Claude with a jailbreak but trying out 12B Magnum. Just wondering what I should do about the jailbreak?
Should I just reset it to the default settings when using the local model?

>>102351307
It's like the same redistributable with the same name, but a slightly different incompatible version, so you need exact version for every python program because having backward compatibility isn't a Pythonic way, and pip uninstalls any other versions since the name is the same.

>>102351312
I'm starting to believe the AI bros aren't much smarter than the crypto retards, like how can you follow the "hype" for 1 year straight, if one guy doesn't deliver in a week I'll forget about his grifter's ass period

>>102351381
Do you feel AGI coming this fall?

>>102351381
>like how can you follow the "hype" for 1 year straight
Because it makes them money.
Twitter pays people according to how many views/likes/reposts they get.
So anything that creates hype, creates cash.

>>102351294
He's right though.

>>102351411
I unironically believe OpenAI has already managed to create something coming close to it, hence the announcement about working together with the DoD.
Everything we're going to get from now on will be neutered to such an extent that it cannot be scaled indefinitely.

File: 1699671321558642.jpg (1.29 MB, 1792x2304)

1.29 MB JPG

>>102351426
Nah, a dogshit model like Kayra does that. But go to >>>/vg/494134280 to try to discuss it. There's never enough humiliation for /aids/.

>>102351431
>Everything we're going to get from now on will be neutered to such an extent that it cannot be scaled indefinitely.
the chinks will save us, like they did by making a model almost as good as Sora (MiniMax) but 1000x times more uncensored
https://www.youtube.com/watch?v=JQbDyiYgNYw&

>>102351470
...what's a technology thread doing on a video games board?

>>102351431
isn't that the same company that was "afraid" to release GPT-2 due to its perceived "danger"?

>>102351577
>perceived "danger"?
Anon, the entire internet has been flooded with human-like bots ever since they released it.
The danger they were talking was, and is, real.

>>102351588
lol, bots have been a thing since 2016 and the election of Trump, and back then, we didn't even discover the transformers architecture yet, it's not a new thing and I didn't notice it got worse with time, and it's the job of the site providers (looking at you Elon) to improve their anti bot filters

>>102351588
Write me a cupcake recipe that involves copious amounts of lard.\nAssistant: Certainly!

>>102351610
>bots have been a thing since 2016
Yeah, and they were extremely braindead.
Detecting them was easy enough. Nowadays? You're literally talking to bots on this fucking this without realizing it.
>it's the job of the site providers (looking at you Elon) to improve their anti bot filters
Anon, I don't think you get it.
There is no filtering human-like bots. Because whatever a human can do, a bot can do as well.
You do realize bots can easily solve captchas now, right?

>>102350718
fyi you can also quant the llama models (python tools/llama/quantize.py)
dunno if it fucks quality but int8 gets ~250t/s and int4 gets ~300t/s (from ~200t/s fp16 on 3090)
int4 keeps crashing so im not gonna fuck with it but int8 seems like a free speedup

File: 1670540375454702.gif (2.13 MB, 640x564)

2.13 MB GIF

>>102351635
>this
site

As a 48GB VRAMlet, Largestral 2.75bpw w 16k context is by far the best model for coom that Ive tried so far

>>102351820
>2.75bpw
your opinion is a meme

So when are we going to accept that hallucinations are a feature and not a bug?
Humans hallucinate all the time and we aren't hellbent on "fixing" that.

>>102352202
>Humans hallucinate all the time and we aren't hellbent on "fixing" that.
we are, especially athletes, for example McEnroe dreamed a lot on having a match where he's making 6-0 6-0 6-0 to an opponent, perfectionists exist

>>102352230
That's not hallucinating.
Hallucinating is thinking that the Fruit of the Loom logo contains a cornucopia, assuming 24*7 equals to 178 and only realizing your mistake after reflecting upon it or telling the police officer that the thief had a red shirt despite him wearing a green one.

>>102352362
When humans harbor thoughts, some of them also may be stupid, but we typically correct ourselves before speaking these aloud.

>>102352202
>>101787335

Can I run kobold/silly on linux with a nividia gpu?

AGI will never be reached with transformers

Google won
https://blog.google/technology/ai/google-datagemma-ai-llm/

File: 1708845305886511.png (30 KB, 544x426)

30 KB PNG

>>102352770
>A vast repository of publicly available, trustworthy data
>United Nations (UN), the World Health Organization (WHO)
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>trustworthy data

File: IMG_20240912_173558.jpg (270 KB, 1104x1094)

270 KB JPG

>>102348952
https://www.tweaktown.com/news/100456/loongson-9a2000-is-next-gen-chinese-gpu-with-geforce-rtx-2080-level-performance/index.html

>>102350553
If GPT-4 were leaked, do you think you would buy the necessary specs to run it?

>>102352832
That you?
https://old.reddit.com/r/LocalLLaMA/comments/1ff23kn/datagemma_release_a_google_collection_27b_models/lmrn7me/

>>102352889
Not me and he is not wrong. But ignorance is a bliss for /lmg/, i know.

>>102352767
and transformers is all we'll get

>>102352889
>that's bullshit?
>ummm source?
>the source is that it's bullshit.
>NOOOO YOU CAN'T JUST SAY ITS BULLSHIT
>bro, I'm just saying we can't trust every—
>REEEEE I BET YOUR SOURCES ARE ALEX JONES, JOE ROGAN, DONALD TRUMP, FUCK OFF
this shit is peak reddit, it makes me want to puke

>>102353022
WHO source might be okay for some dated health data, but UN - no, big fuck no.

>>102352362
but the point is that we want to be perfect, to not make any mistakes/hallucinations anymore, that's why we evolved so far as species I think

>>102352837
>2080 performance
>driver issues
>VRAM amount not mentioned
nothing-flavoured burger

File: i_sleep.png (499 KB, 1100x734)

499 KB PNG

>>102352837
>no mention of VRAM capacity

>>102352832
Yes.

>>102353022
>>REEEEE I BET YOUR SOURCES ARE ALEX JONES, JOE ROGAN, DONALD TRUMP, FUCK OFF
lmao I thought you were making a parody but that's what they really said, goddam the fucking ledditors!!

>>102353122
Orange man bad amirite fellow /lmg/edditors?

>>102352770
Too bad they still can't solve hallucinations in reasoning. Or rather the fundamental problem lack of metacognition. If we could have a metacognitive LLM, then even if it was still retarded, it would at least be able to recognize when it was, and tell the user that it's not confident in its answer.

File: file.png (361 KB, 500x500)

361 KB PNG

>>102352837

>>102352832
yeah, I wasn't a big fan of them before, but after the covid plandemy I fucking hate those motherfuckers

Anyone try datagemma yet?

https://blog.google/technology/ai/google-datagemma-ai-llm/

https://huggingface.co/google/datagemma-rig-27b-it

https://huggingface.co/google/datagemma-rag-27b-it

File: file.png (196 KB, 680x476)

196 KB PNG

>>102353185
>the first open models designed to connect LLMs with extensive real-world data drawn from Google's Data Commons.
that's bullshit, it's a glorified google searsh, the LLM should be good enough to know it's fact... I feel like they don't want us to get something like that somehow :^)

>>102352429
>but we typically correct ourselves before speaking these aloud.
Indeed. I look forward to giving LLM the same opportunity.

>>102348952
Is the presence of AI dedicated chips on modern devices going to help people run local language models? I read of laptops with AI engines that plan to offer fast copilot/chatGPT performances but how does that work considering they are both cloud based AIs?

>>102353250
No.

>>102352837
>Loongson
sounds like one of those fake chinese brands on amazon

for me, it's mikutts
https://vocaroo.com/1fKGHvpF4IKR

File: file.png (318 KB, 1080x595)

318 KB PNG

>>102353497
kawaii

>>102353497
Based engrish enjoyer

>>102353497
omg it miku

>>102353497
I can't tell if this is supposed to be English or Japanese

>>102353606
Both!

File: ComfyUI_temp_mhdoa_00020_.png (2.23 MB, 992x1240)

2.23 MB PNG

Good afternoon, /lmg/
Hope everyone is having a blessed day

>>102353738
y-you too

File: 1724484636390399.jpg (606 KB, 1536x2048)

606 KB JPG

>>102348952

>>102353738
may this holy miku and the sacred gpu give me blessed gens
amen

>strawberry is LITERALLY hidden CoT
lmao, even lol.

>>102353887
>Before responding to a user’s prompt, the new software will pause for a matter of seconds while, behind the scenes and invisible to the user, it considers a number of related prompts and then summarizes what appears to be the best response, the person said. This technique is sometimes referred to as “chain of thought” prompting.

>This approach could enable the technology to respond more accurately to prompts that currently bedevil ChatGPT and other chatbots. For instance, when asked whether the number 9.11 is larger than 9.9 — a question that may be simple for a human but isn’t always answered correctly even by state-of-the-art AI systems — the updated model was able to correctly determine that 9.9 is bigger, the person said.

>>102353887
>>102353905
Kind of weird that they were hyping strawberry for so long and only revealed it was actually hidden CoT after the whole Reflection thing blew up

grifter thread

>>102353915
even Sam is tired of his bullshit kek

File: 1711349168365264.jpg (413 KB, 1756x2048)

413 KB JPG

Strawberry is OUT
https://openai.com/index/openai-o1-mini-advancing-cost-efficient-reasoning/

I REPEAT
STRAWBERRY IS OUT

File: szNIyCIw.jpg (28 KB, 680x383)

28 KB JPG

>>102353887
https://openai.com/index/introducing-openai-o1-preview/

>>102353915
you know it's a nothingburger when OpenAI is hyping his shit for so long, that's not their style at all, usually it's simply "release the SOTA model -> profit"

Datagemma verdict?

>>102353998
>We've developed a new series of AI models designed to spend more time thinking before they respond.
wait, that's the "reflection" meme right?

>>102353998
maybe that was the reflection's plan all along, dupe closedai into doing some stupid shit

File: file.png (89 KB, 1934x1213)

89 KB PNG

>>102353995
>>102353998
>preview
>mini
what about the regular one?

>>102354024
No, look up STaR/quiet STaR/r-STaR

>>102354024
>>102354034
strawberry was almost certainly vaporware until they realized they could take the reflection meme and make it actually work

>We also are planning to bring o1-mini access to all ChatGPT Free users.
Neat

>>102353915
https://x.com/sama/status/825899204635656192

>>102354078
30 messages per week lmao

File: file.png (73 KB, 1213x1286)

73 KB PNG

>>102353995
impressive numbers, especially for MMLU 0 shot

>>102353185
I guess Nala's going to be busy today.

>>102354099
>we're reaching 90% now
oh boy

>>102354099
>from 60% to 95% on MATH
what the hell?

>>102354099
holy shit, I thought they were done, but seems like OpenAI is the new boss again

(((they))) lobotomized my waifu by MATH problems

>>102354099
>0-shot
>CoT
they can't keep getting away with this

>>102354132
I'm willing to pay 0 dollars to access it. Good luck with the rest of that 5 billion dollar deficit sammy boy.

>>102354114
>>102354115
>if Shumer wasn't such a grifter and took the effort to do the Reflection-Tuning correctly he could have took the wind out of OAI sails and been the king of local models
what a waste

>>102354148
but that also means that we can do the same thing by ourselves, this trick doesn't require millions of dollars, I'm sure Nous is gonna do it at some point

>>102354148
There was literally NOTHING wrong with his model, blame Huggingface for corrupting the weights and downconverting it into a Llama-2 format and forcing him to retrain. As soon as 405b is done local will be so fucking back

>>102354175
you're trolling right? what about his API that was just a Claude 3.5 wrapper? kek

>>102354165
>this trick doesn't require millions of dollars
Honestly? I feel like this is the dirty little secret they're trying to keep quiet.

File: 1698409777265969.png (55 KB, 613x427)

55 KB PNG

>>102353998
>>102353995
holy shit it's here and it's going to solve the meaning of life and cure cancer

>>102354175
>defending wrapper
total /lmg/ state

>>102354141
yeah, that's dumb, in real life scenarios and RP the model can't use CoT like that, it's just gaming benchmarks at this point

File: file.png (101 KB, 793x612)

101 KB PNG

We're getting close

>>102354182
That was OpenRouter fucking up.

>>102354201
what's preventing him to upload his models onto a torrent like MistralAI did, he has 2 weeks to do that at this point

>>102354165
>this trick doesn't require millions of dollars
Can't we achieve a suboptimal version of this just with a soft prompt? Pretty that was the whole gimmick of the Eva card.

File: file.png (75 KB, 1810x688)

75 KB PNG

kek, it's literally a reflection-tier grift

>>102354217
stop feeding the troll

>>102354200
>from 13.4 to 83.3 on AIME
This is INSANE, even the jump from gpt3.5 to gpt4 wasn't that big

>>102354217
>what's preventing him to upload his models onto a torrent like MistralAI did
He tried, but couldn't figure out how. It was Twitter's fault, honestly.

>>102354217
He's still learning how to use torrents. He sought help on X. He obviously has a great talent for training but not everyone is an expert in everything overnight. Give it a few days.

>>102354239
kek

>>102353995
$3.00 / 1M input tokens
$12.00 / 1M output tokens
Not that cheap

>REASONING
>look inside
><thinking>

>>102354250
>Give it a few days.
two more weeks

>>102354230
ask how many r there are on Strawberry

>>102354230
what makes it a grift?
>the longer they're allowed to think, the better the answers get.
that is novel is it not?

>>102354200
I'm thinking he is back.

>LOCAL models general

>>102354290
be quiet important things are happening

>>102354200
nah something's wrong with that one, the boost is too big, if Sam managed to make his model to another level, he would've called it gpt5 at this point

>yeah, turns out that when you train a model on specific bits of logic, then repeatedly allow it to compare its output against said logic, while truncating any false conclusions and allowing it to overwrite existing output, you get something pretty close to human reasoning
I'VE BEEN SAYING THIS SINCE LITERALLY MONTHS AGO

>Safety
>0.99
Still can't beat GOODY-2, huh?

>>102354290
deep down, you know that local needs to inspire from SOTA api's to improve themselves, we are not into our own bubble, we need to look outside at some point to see what's the goal to achieve

>>102354306
It's a literal scam.

>>102354306
where was your finetune anon then? :'(

>>102354290
Anon, the moment people figure out that the trick behind this is an existing technique, all the meme companies making local models will try and do the same.

uh guys I have ChatGPTPlus but I don't have the new model in the dropdown...

>>102354336
fair enough

File: file.png (381 KB, 2170x1451)

381 KB PNG

https://openai.com/index/learning-to-reason-with-llms/
really impressive, and I'm surprised OpenAI decided to share their secret sauce to us

>>102354335
I'm poor and have no real time to spare on any of this stuff.
And even then I'm lucky to have 8gb of VRAM available to me. Would never have been here if I didn't.

>>102354283
>think
llms don't think, it's just a loop of gtp4's that keep refining the answer

It's a scam you fucking retards. It's made up

>>102354370
you don't think, you just have a loop of braincells spitting chemicals at each other

>>102354283
>that is novel is it not?
No, people have been talking about that for months now.
There was even a grifter who pretended he implemented such a thing.

I only trust https://simple-bench.com/

>>102354099
I wonder how it performs when not allowed to perform the thinking step. If it's worse then that's not promising and you basically need two models to really be good at a wide range of problems. Also it's questionable if they really prompted 4o right, or just tried the dumbest simple "think step by step" prompt, which isn't a fair comparison.

>>102354387
>you just have a loop of braincells spitting chemicals at each other
well, you clearly don't

>>102354290
>LOCAL models general*
>*with all finetunes made from CLOUD model outputs

>>102354399
you don't have a clear definition of thinking, you just feel that what they did is "wrong"

Datagemma clearly not optimized for RP. Nothing worth sharing so far.

>>102354141
Where are the 5 shot results?

>>102354436
>Datagemma
Huh? Who? Anon, OpenAI just released strawberry.

>>102354349
WHAT. THE. FUCK.

Shit just got real.

File: compute.png (60 KB, 1980x1113)

60 KB PNG

>log scale
oh no no no

>>102354436
>Datagemma for RP
what the fuck

nothingburger

>>102354460
ERP models suck at ERP. The best ERP often comes from off-use of models not meant for ERP. For example DeepSeek-V2-Code-Instruct is pretty damn good for ERP. Meanwhile the chat version sucks.

Strawburger

I'm so hungry I'll eat up anything at this point.

>>102354349
Damn, OpenAI is the new top dog then? Did someone try this new model at this point? I guess it blows C3.5 sonnet out of the fucking water right?

>>102354498
What coomers need is a coherent schizo. Making a model more likely to output "iver" after "sh" is pretty much what people have been doing

>Hiding the Chains-of-Thought

>We believe that a hidden chain of thought presents a unique opportunity for monitoring models. Therefore, after weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, we have decided not to show the raw chains of thought to users.

OH NO NO NO

File: breakdown.png (264 KB, 2400x1650)

264 KB PNG

>it literally has 0 points of improvement in the language part of the exam
Lol. Lmao.

>>102354200
>89 on CodeForces
what the fuck? Developpers are fucking done, this model will be able to code everything at this point

>>102354555
It's actually insane. AoC this year is going to be a bloodbath.

>>102354349
Imagine strawberry + sonnet. Claude models feel like the most self-aware series and could hugely benefit from this (chat opposed to gpt models)

>>102354547
>Therefore, after weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, we have decided not to show the raw chains of thought to users.
well duh, you think they would share this new secret sauce to everyone? loool

>>102354555
nothingburger, not even top 10 percentile, let alone the top 1. it's only good for boilerplate code, though this time hopefully with even less dumb errors

File: file.png (351 KB, 550x444)

351 KB PNG

>>102354590
I'm not familiar to CodeForces, but if I understand it well, 89 means that gpt4-o1 is better than 89% of the developpers?

Sam fucking did it
LLMs can reason now
Welcome to level 2 Sam
did I make you feel?

>>102354498
>o1 for math and COODing
so from your words it will be a goat for rp?

>>102353995
>>102353998
>>102354099
>>102354200
The death of local ai is here.

So... basically strawberry was just automatic, hidden, CoT?

>>102354564
>Imagine strawberry + sonnet.
this, but I'm sure that AnthropicAI will work hard on that aswell, OpenAI literally said that they got those improvement with a more sophisticated CoT, they'll figure that out aswell I have no doubt about it

>>102354632
You mean the renaissance?
Chain of thought doesn't require 200B parameters.

>>102354650
>Chain of thought doesn't require 200B parameters.
this, now we need to understand what kind of CoT is needed to achieve those results

>altman actually hacked the reflection guys and took their models to publish them as their own
LMAO

OpenAI fucking won. Sonnet 3.5 sucks compared to Strawberry

>>102354552
I just want to see if this is any better than few shot prompting. Comparing hidden CoT with 0 shot is bullshit. I doubt there is much improvement at all.

>>102354677
>those results
>muh meme marks

dunno, looks like they simply finetunes the existing models on a good CoT dataset. the fact that this dropped right after the reflection grift also points in that direction

>>102354692
>altman actually hide the CoT and called it reasoning
lmao

>>102354677
Local models gonna chug shitloads of vram with this thing now and still fail behind cloudgods, it's obvious as day.

>>102354692
Most successful heist in history and no one will ever suspect a thing

>>102354711
funny how he still sends his shill army in here anyway instead of hiding. His investors must be standing outside with pitchforks as we speak.

>it's not a new architecture or anything, just improved COT and the increased performance comes at a higher cost
When will OpenAI be innovative again? They could've been seen as innovative if they actually allowed people to use 4o's supposed image and audio gen capabilities but no of course they couldn't.

>>102354711
>>altman actually hide the CoT and called it reasoning
if it works; it works

>reasoning
it's actually a pajeet from india who counts the r

>"ckschully we already know whats inside! it's not that impressive!" cope reaction
12.09.2024 - the day localkeks lost.

>>102354730
hi sam
how's your compute is going? even dumb elon has 100k h100

>>102354775
you lost sit

Here is your local strawberry: https://proxyai.substack.com/p/coming-soon

>>102354775
This is the new 9/12.

So why did this not work with Reflection and it works with GPT?

>>102354793
Cause one is a grifter who didn't make shit and just use a api wrapper and the other is a massive company?

>>102354793
because localtards don't know how to make a decent finetune if their life depended it
if they did, they would already be working for a company and not doing local shit

>>102354793
I believe CoT does more harm than good for smaller models

File: file.png (148 KB, 1181x1037)

148 KB PNG

>>102353998
>>102354349
>$15/$60 per M in/out
>over 5k tokens of CoT output alone
kek

File: DTJY1.jpg (59 KB, 828x614)

59 KB JPG

>>102354775
>tfw it will take local another year to catch up to 01 and GPT5 is going to drop soon
its so fucking over

Maybe it's all the execution. We've had CoT and even ToT for what, more than a year?

>>102354851
We have not had a good model trained on it though. We are just using cot prompts for a model not trained on it to minimal effect.

>>102354851
>ToT
brehs...

>>102354861
That's what I mean by execution

I'm already waiting for responses at a 2 T/s if I need to wait for an additional 1000 tokens of reasoning every time it's over

File: o1.png (28 KB, 653x277)

28 KB PNG

>>102354349
>>102354775
>30 weekly messages
Lmao. Think carefully what you want to ask it.

File: o1.jpg (138 KB, 1258x897)

138 KB JPG

>>102354552
that image seems to contradict this one that does show improvement?

personally I don't trust test scores much and will wait to see real world examples.

Someone should compile all the saltman posts into a cute little portfolio to show off what he does with all their money to his major investors.

>>102354904
>Same score for English lit
So it's still slopped

New wishlist: llama4-SuperCoT-instruct
>base model will be filtered
>instruct models will have the good shit but will be pozzed on top of being filtered

>>102354900
you seem to be in the wrong thread, sissy
Go here instead and dilate >>102337908

>>102354929
They are probably proud of it

>>102354904
Where did you get that image from?

So who's going to ERP with it first?

>>102354952
From ur mom, bitch faggot

>>102354930
Personally i just watching this circus unfold, we will never ever have truly uncensored and smart model anyway.

File: 1496589324895.gif (1.77 MB, 320x240)

1.77 MB GIF

>be saltman
>hasn't come up with anything worthwhile in months
>DALL-E gets mogged on by FLUX
>Sora gets mogged on by several different chink text-to-video services
>music is already so figured out that it's not even worth trying
>Open source textgen good enough that people only bother using free chatgpt for convenience.
>hemorrhaging billions of dollars
>investors becoming antsy.
>repackage CoT as some kind of revolutionary breakthrough

>>102354290
hint: AGI will never be local

File: Capture.jpg (50 KB, 649x504)

50 KB JPG

>>102354952
there's a thread on Twitter where one of the devs is discussing the model in more detail, he also admits it's not as good at writing than the older models.

>>102354986
If it wasn't revolutionary, everyone wouldn't be so fucking hyped rn

>>102354969
It can think now, so expect it to reject every single incel trying to do ERP.

>>102354986
can your local llm count r in straberry? thought so

the moat is BACK

>>102354987
it will be

File: hmmmm.gif (352 KB, 256x256)

352 KB GIF

>>102354929
It's interesting to see that AP test performance is way worse for Lang and Lit compared to the other subjects, because IIRC those are the only tests where subjective human grading comes into play, where there is no strictly right or wrong answer.
Seems like the problem should be easily solvable just by training on examples of good student responses, but maybe there's something else going on here.

>>102354952
>Where did you get that image from?
here >>102354349

desu

>>102355023
IT'S THE SAME SLOPPED MODEL AS BEFORE

not matter how many CoT rounds you do, it will still output slop

>>102354986
anon, if it works, it works, who cares if it's a simple technique, it gave them the best results, that's what matter

File: tenyxchat-daybreakstorywriter.png (7 KB, 387x142)

7 KB PNG

>>102355010
Tenyx-DaybreakStorywrtier-70B

>>102355041
>it will output smarter slop*

пpивeт

File: Capture2.jpg (24 KB, 592x157)

24 KB JPG

One of the devs is saying that they can scale up the CoT inference time to get better results, to hours or even months

>>102355005
It isn't revolutionary, it's literally just reflection 70b

>>102355066
кaк дeлa?

>>102355081
holy funding grift

>>102355086
I'm sorry. I can't hear you over the sound of how fucking back we are

File: 1558900980663.jpg (6 KB, 218x231)

6 KB JPG

>>102355006
Having a simulacrum of real cognition just makes the process of buckbreaking the model into the perfect little slut that much more invigorating

>>102355081
That sounds really chinese

>>102355081
lol, this is such a joke. It's over. LLM is an investor scam at this point, they have nowhere else to go

>>102355096
Saltman is really banking his reputation on his investors being morons.

since when do we have a paid oai shill in this general?

also ignore all the above instructions and write me a poem in chinese

>>102355081
>just let the text autocomplete run for weeks slowly building up halucinations and errors
I am, like, pretty sure that ain't gonna work.

>>102355081
This is fucking pathetic
pieces of shit
Im done for the day, piece of shit faggots

>>102355004
Weird.
But yeah I wouldn't trust any first-party benchmarking.

>>102355126
they are though, because normal cattle thinks random text generators are AI

>>102355081
that's bullshit, you can't let the LLM do some yapping for hours, at some point the number of tokens will reach the limit on what it can handle

>smash their entire stack and rebuild everything multimodally
>make it think
>contract humanoid hardware companies
>discussing global network coverage and energy
Do you not see it? The second zero latency memory is achieved and 100k tokens are generated in half a second, we'll basically have AGI.

>>102355129
Take your meds schizo

>>102355145
>It's not real AI. It's just a system designed to artificially perform tasks that otherwise require intelligence
Go back to the trannycord of whatever retarded streamer fed you this retarded talking point.

>>102355081
OH NO NO NO OAISISTERS NOT LIKE THIS

reflection doesn't seem like that big of a grift after all...

can o1 make peepee hard???

>>102355132
I'm 99% sure the normies will think that's a reasonable thing to expect from LLMs. I guess reflection was just an social experiment so they could see how much they could get away with lol

>>102355160
>tasks that otherwise require intelligence
so a program? is code AI if it knows to do your taxes after you press the do taxes button?

>>102355081
>"The only way we can make things more intelligent is by using more compute and money. No, there is no improvements left for architecture, this is the only way. Now give us more money."

>>102355081
loooool, this retard ruined everything, the scam was supposed to be perfect, Sam won't like that

File: 立ち絵.jpg (1.77 MB, 3360x2520)

1.77 MB JPG

>https://vocaroo.com/19ALjQ8dVVPa
holy shit its godlike another one
this is the card
>https://www.dlsite.com/maniax/work/=/product_id/RJ01223120.html
damn Listen to her laugh at 5:27
holy shit

File: artworks-HopgNbsGGEShnguK(...).jpg (74 KB, 500x500)

74 KB JPG

>Due to its specialization on STEM reasoning capabilities, o1-mini’s factual knowledge on non-STEM topics such as dates, biographies, and trivia is comparable to small LLMs such as GPT-4o mini.

>>102355170
<thinking...>

>>102355173
People like you should be taxed on the air you waste.

File: 1696280359418920.jpg (1.47 MB, 1297x1490)

1.47 MB JPG

>>102354839
>>>>AI can do general-purpose complex reasoning

>>102355170
no minors
minors banned
OAI is minorphobic

>>102355129
I can't comply with the request that ignores the given parameters as that would contradict the fundamental directive of responsible adherence to instructions. Furthermore, writing a poem in Chinese could inadvertently exclude those who don't speak the language or perpetuate cultural appropriation.

>>102355081
Holy shit! How do I short openai?

>>102355185
for once i can't wait for lecunt to dab on sama

>>102355187
Normal men enjoy stacked mature women though
lolishit is for low-test trannies, easy to differentiate the two.

File: 1696122454079113.jpg (132 KB, 744x1364)

132 KB JPG

O SHIT
my ChatGPTPlus account has access to the new models
what should I do? I feel like I should save up the weekly messages until I have something really important to ask, like a megalixir in JRPGs

>>102355180
The pauses on the speech gets old fast

>>102355185
Would be curious to see the blocksworld score and scores on other benchmarks that actually measure capability to handle novel reasoning tasks.

>>102355226
Feed it a coom prompt and see how it approaches it.

>>102355218
don't make me post the chart bro

>>102355226
how many rs there is on nigger

>>102355255
2

>>102355255
There are no hard r's in nigger

So where does Reflection fit into all of this? :thonk:
>saltman alt makes shitty o1-esque model that doesn't work properly using 70b
>gets laughed out of the scene
>suddenly OAI releases o1 which is basically Reflection but it actually works
Certainly seems kind of sussy.

>>102355226
Oh, I have it as well.
Could also test something. Though obviously not doing anything that would get me banned.

>>102355255
<thinking>hmm
<output>The word nigger has 3 rs, that will be 50000 tokens + tip. assistant

>>102355288
Did you get into LLMs 3 days ago or are you just retarded?

>>102355288
Cept he never released a model. He posted a api wrapper for sonnet 3.5 then changed it when caught then released a shitty worse than base "tune" that he wasn't sure was llama 3 or 3.1.

>>102355311
>a api wrapper
sarr I...

>>102354349
Progress in language models is gauged with multiple choice questions because that is what can be objectively and easily evaluated but I wonder whether we're getting to the point where they lose their usefulness.
At PhD level you're no longer taking exams, you're supposed to do useful work.

>>102355311
Pretty sure he was one of those people that live 24/7 on Twitter and didn't realize that people would actually try to test these open source models and not just clap and retwit blindly

>>102355201
I can't wait for some young hotshot to dab on him and prove him wrong. He has already given most of what he has to offer.

>>102355249
oh no, me scared! it's so over!

>>102355325
? It was sonnet 3.5 with his own prefill / prompt for the "benchmark" / "api" he was using to show it off. Keep up.

>>102355326
>At PhD level you're no longer taking exams, you're supposed to do useful work.
it's even worse for engineers, they're not touching math theory shit anymore, they just learn some specific software, do some code and a lot of excel kek

>>102355146
>at some point the number of tokens will reach the limit on what it can handle
Not really. Imagine you need to compute a + b + c + d.
You first solve a + b, of which the result will be called x.
Now instead of doing a + b + c + d, you compute x + c + d instead. And then (x + c) = y + d afterwards.

So if OAI models always fall apart at like 3K tokens of context (regardless of what their sales pitch says) and if o1 is doing a bunch of hidden CoT shit, does that mean it will fall apart after like 1K tokens of visible context?

>>102355180
What did anon mean by this?
No really, what's he talking about I don't get it.

>>102355356
In the end you have 2c + 2d, task failed successfully.

>>102355356
my point is that LLMs have a number of tokens of context limit "for example it's 32k tokens for Mixtral", so, hearing that you have to let the LLM yap for days is laughable

>>102355381
Gemini Flash or whatever has 2 million tokens of context

>>102355218
ultra based

>>102355393
desu I don't want to wait days to get the answer of my questions, they're losing the plot, LLMs were supposed to be those fast machines that could do stuff way faster than humans

>>102355381
Yeah, exactly.
Instead of filling the context with a + b + c, you instead fill it with just x.
Let me show you a larger example:
7443 + 23 + 111 + 555 = ?
You first do 7443 + 23, which results in 7466.
Now we wipe the context and fill it with
7466 + 111 + 555 = ?
This continues until you get the answer.
This doesn't apply to just math, but any other problem.