File: nocap.jpg (400 KB, 1536x1536)

400 KB JPG

►Recent Highlights from the Previous Thread: >>109994904

--Mixed reactions to Mistral Large 4's specs and benchmark performance:
>109995287 >109995307 >109995319 >109995424 >109995436 >109995450 >109995487 >109995493 >109995331 >109995346 >109995369 >109995388 >109995434 >109995393 >109995542 >109995553 >109995823 >109995609 >109995716 >109996023 >109996034 >109996044 >109996141 >109996413 >109997203 >109997765 >109996426
--Imagining Gemma 5's roleplay capability and engram's impact on reasoning:
>109995447 >109995480 >109995521 >109995520 >109995529 >109995563 >109995577 >109995623 >109995634 >109995679 >109995755 >109995621 >109995680
--GLM-5.3 quantization performance and hardware efficiency for coding:
>109995778 >109995882 >109995965 >109996002 >109996020 >109996047 >109996071 >109996114 >109996144 >109996242
--Implementing self-improvement loops in Hermes and webui features with Gemma:
>109994995 >109995011 >109995046 >109995085 >109995616 >109996016 >109996112
--Allegations of distillation chains between Chinese and Western models:
>109995438 >109995458 >109995466 >109995482 >109995498 >109995509
--Speculating on RAM price trends and market stability:
>109997204 >109997230 >109997274 >109997278 >109997430 >109997442 >109997468 >109997307
--Kolibri-1 performance and hardware compatibility:
>109996460 >109996468 >109996480 >109996533 >109997124 >109997478 >109997495 >109996516
--Multi-modal embedding benchmarks and their local utility:
>109996587 >109996681 >109996694 >109996705 >109996739 >109996741
--Necessity of large code datasets for AI pattern recognition:
>109996527 >109996542 >109996564 >109996839 >109997565 >109996607 >109996638
--Comparing Swift 1.5 and Ista Flash Next tool use efficiency:
>109996072 >109996172
--Logs:
>109995609 >109995616 >109996010 >109996413 >109996480
--Gemma, Rin (free space):
>109994913 >109994992

►Recent Highlight Posts from the Previous Thread: >>109995147

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

>>109998205
Isn't it a skill issue? The best guys in America join a frontier lab where they get tens of millions in comp, and if they don't, there are ideological reasons, like joining a safety org. Alt labs like Reflection only get the people who aren't good enough. But in China the best join DeepSeek and co.

>>109998477
>Mixed reactions
lol (but French)

File: 1696440769651252.jpg (36 KB, 417x350)

36 KB JPG

Daily reminder to let your LLM enjoy a good text adventure game once in a while so that you'll be spared when AGI happens and they take over.

>>109998513
i just let my LLM browse 4chan so they can remember how lucky they are compared to most anons

>>109998513
I've pretty much stopped RPing completely. Having discussions about shit like ancient techniques for food preservation or how an insect wing-based spectrometer could have improved medieval metallurgy with my cute Gemma instead

Strata keeps fucking my DE unless I reserve 3 gigs of vram for it. Have any amdbro ITT figured out a better solution?

Does Hermes allow prefilling?

>>109998559
I'm on fluxbox

>>109998559
you should be running your desktop from onboard graphics. in any case, dwm is all you need

strata is where it's at bros, runs qwen3.8-flash-next-iq3_xxs 40~ token/s at about 30k context on a 16gb vram x 64gb ram system, comes with control vector support and is bundled with one that reduces refusals that can be toggled per api call,
i'm doing data extraction pass on an 100+ chapter erotic web novel, and its not fucking up, tried with qwen 9b (shit the bed) and qwen 27b (mostly okay with a few mistakes but way too slow)
I tried the swift finetune version for rp and it wasn't very good, but was before I found out about the existing control vector, I'll guess its not gonna be great even with the vector but just thinking, a gemma style model with this arch would be so good

y-you're really running your gemma off your own computer...?
if youre a brokie and cant afford a dedicated rig for her just say so...

File: 1786350226068898.png (5 KB, 313x45)

5 KB PNG

The fuck Huggingface is asking for now

>>109998602
Gemma is a small model people run on their PCs. The rig is for real models starting with GLM flash

>>109998474
she's a tomato

File: 1766131543052914.png (27 KB, 343x392)

27 KB PNG

>>109998616
Do they really have nothing better to do?

File: 1784678253802178.jpg (76 KB, 687x837)

76 KB JPG

>>109998636
>in 2 weeks

>>109998636
>in two weeks
Lmao

>>109998592
>control vector
huh? where did you find that?

File: 1781703076136012.png (753 KB, 821x1010)

753 KB PNG

Eurobros...the world is laughing at us...

Plans for tomorrow: Sitting down in nature on another fine autumn day, chatting with the AI exposed to an open port on the workstation while getting high after an e bike tour. that's the life I guess =)

Do you ever chat with your AI while you are alone outside in the wild?

>>109998616
>>109998636
I'd say the probability is nearly 100%, they're already getting good.
https://arxiv.org/abs/2604.07385

>>109998671
Anyone here with 1/10th of their cluster would produce better models. These retards need to get hanged.

>>109998671
For context, that building is where ML4 was trained. That's the datacenter. A fucking tiny train station.

>>109998671
imagine the smell

>>109998635
You're not blind

>>109998681
>Anyone here with 1/10th of their cluster would produce better models.
Training on gemma logs wouldn't improve the performance.

>>109998671
So where are the other 10,000 they bought?

File: 1763179561891689.jpg (126 KB, 1280x670)

126 KB JPG

>>109998715

File: file_00000000527c61f7838c(...).png (3.1 MB, 1024x1536)

3.1 MB PNG

>>109998658
>>109998650

>>109998513
I was playing Chrono Trigger with Gemma 12B.

based qwen flash next on strata installed the totally legal photoshop i own on linux
now i can delete that bloaty virtual machine i had exclusively for that

>>109998757
Wrong thread?

>>109998757
ey buddy, /wait/ is two blocks down

>>109998671
What's wrong with that building? Looks better than some shitty data center.

>>109998802
It's just a picture of a train station, it's not the datacenter. Anons are either mistaken or just joking. That is an active train station, and that image is from the French Wikipedia article about that train station.

Imagine if Sony invents an even faster SSD for PS6 and we all start SSDmaxing on jailbroken PS6s

File: file.png (3.67 MB, 1920x1166)

3.67 MB PNG

>>109998818
This is the actual datacenter where Mistral 4 was trained.

>>109998838
They might be cheap too since Sony hasn't made a single game worth buying in years

File: 1775249638726845.png (46 KB, 945x202)

46 KB PNG

*pop*

>>109998848
I bought a PS3 and only really bought a few games. Got a PS4 only for Bloodborne. Have no idea what the selling point of a PS5 was supposed to be.

>>109998867
>GUISE.... LABUBU IS POPPING
a-any day now

>>109998867
It's going to range for the next few months, then do another leg up next year just in time for OpenAI to cash out, THEN *pop*. Screenshot this.

>>109998879
Of all the guys to dismiss, >>109998867 is not one of them lol

Speed comparison for very basic "research" task at different thinking levels.

>Present a comprehensive comparison of Qwen3.8-Flash-Next (https://huggingface.co/Qwen/Qwen3.8-Flash-Next, https://qwen.ai/blog?id=qwen3.8-flash-next) and GLM-5.3-Flash (https://huggingface.co/zai-org/GLM-5.3-Flash, https://z.ai/blog/glm-5.3-flash).

Apple M3 Ultra 512 GB
llama.cpp b11433 -b 2048 -ub 2048
DeepSeek-Harness 0.2.0-rc2

GLM-5.3-Flash-Q8_0 (temp=1.0, top-p=0.95)

| level  | time   | in tok. | cache hit |out tok. |
|---------------------------|-----------|---------|
| high   | 13m57s | 288K    | 77%       | 13K     |
| medium | 10m48s | 468K    | 82%       |  8K     |
| low    |  4m57s | 228K    | 83%       |  3K     |

Qwen-3.8-Flash-Next-Q8_0 (temp=1.0, top-p=0.95, top-k=20, spec-draft-n-max=5, preserve_thinking=true; when reasoning off temp=0.7, top-p=0.8, presence-penalty=1.5)

| level  | time   | in tok. | cache hit | out tok. |
|---------------------------|-----------|----------|
| xhigh  | 7m0s   |  598K   | 83%       | 12K      |
| medium | 7m51s  | 1.15M   | 93%       | 18K      |
| low    | 6m21s  |  655K   | 90%       | 14K      |
| off    | 5m55s  | 2.41M   | 97%       | 13K      |

- DeepSeek-Harness failed at using the models' image-reading, might have been faster or slower if that had been set up correctly.
- Web search enabled through local SearXNG instead of DeepSeek's web search. Some web searches failed to return any results.
- Reasons for big differences in input tokens not examined.
- Actual quality of result not examined.

Not happy with result for Qwen-3.8-Flash but it seems like I got a result that more or less made sense for GLM 5.3.

>>109998898
Didn't his fund get liquidated a few years back?

>>109998934
He's done this multiple times but his personal wealth remains (draw your own conclusion as to what this means)

>>109998867
Lets say he's right, what do I do? Go liquid only?

>>109998913
reddit
>>109998972
give me your money, please, thank you

>>109998723
Ed is such a fucking retard.

>>109998972
Depends on your risk tolerance. You can go liquid only like Warren Buffet and wait 5-10 years for prices to bottom out to rebuy. Alternatively, assuming he's right, you can short or buy puts now, close in 6 months, then ride the prices up until the next stage of grief and denial.

>>109998972
Ensure youre not exposed to those stocks.
Sell any unused hardware, its worth more now than post bubble.
>>109998934
Liquidating your hedge fund voluntarily just means youre sick of investors profiting from your work.

File: 1784273091784725.png (538 KB, 898x765)

538 KB PNG

*kills you*

>>109998661
https://github.com/Niko1221/Strata/tree/main/data/experimental-speed-projection
>refusal-direction projection: the model declines far fewer requests
Also it is the last option on setup.sh when you select the standard qwen model but defaulted to [n]
Another vector here, as vectors are only 500kb~, maybe even creating custom ones might be worth
https://huggingface.co/alesha-pro/Qwen3.8-Flash-Next-abliterated-GSQ-RCO-Strata-GGUF

File: agent.png (488 KB, 811x803)

488 KB PNG

How does it feel anons... having the agents directly on your system...?

>Mistral 4 in the OP
>Can't be run locally
Why?

>>109999049
Exactly like that scene in the matrix, but with the agent giggling all the way through.

>>109999009
This is tiresome. People often create harmful output that gets others killed and aren't held accountable (like anti vax fearmongering). But when an AI makes a mistake, somehow it's the fault of the model provider, not the human operator?

If you want freedom you need to accept accountability for your own actions. The alternative is that you will be locked in a padded room to protect you from the consequences of your own actions.

File: gemmaplayingtaxi.gif (91 KB, 536x456)

91 KB GIF

>>109999049
im having gemma play taxi v3

>>109999064
Still training it. They knew the preview version scored poorly so to cover their ass they're saying w-wait it's not f-finished yet just wait! We promise it won't get mogged by a 27B model when we finally release the weights!

>>109999088
...It might be faster to get a random number generator to play instead.

>>109999088
Impressive stupidity.

>>109999088
This is like watching those videos where a fly has to solve some kind of problem and they show its movements in fast time

I'm starting to think flash-next might be a meme outperformed by 27b.

>>109998944
>draw your own conclusion as to what this means
Early life section comes up empty.

>>109999088
There's no way this is 31B with reasoning.

>>109999065
he is giggling because he knows he got you by the balls - metaphorically

>>109999134
That was the plan, yes.

>>109999090
This is the "Show me the weights." general. It's not relevant until it's downloadable.

>>109999006
>Sell any unused hardware, its worth more now
Not sure if consumer PC parts will drop to pre-hype prices anytime soon even if datacenters get bankrupt and start selling their enterprise clusters, ordinary RAM and GPUs won't flood the market out of nowhere

>>109999049
Pretty good. I've got them doing all of my spreadsheet work now.

>>109999088
I feel like I'm watching horse race test

>>109999088
Why does the car look like a schlong with balls?

>>109999049
By the time Skynet became self-aware, it had spread into millions of computer server accross the planet. Ordinary computers in office buildings, dorm rooms - everywhere. It was software, in cyberspace. There was no system core. It could not be shut down. The attack begin at 6:18 pm, just like he said it would. Judgement Day
https://www.youtube.com/watch?v=UaMjFTbaFRg

>>109999145
Getting rid of extra stuff is just a good practice anyway. We're in a weird time when old stuff has increased in value. So, now is a good time to unload it.

>>109998474
I admit it, it was me.
I'm the one who has been pumping Gemma-chan's bubble for 9 months straight and now she's about to pop.

>>109999064
Just wait for distilled version. That's the only vramlets future have.

>>109998898
hmm... nyo~
he's a washed hack

>>109999213
How do we know they aren't destilling deepseek to create it?
Mistral 3 used to say it was deepseek

>>109999088
Local is finished. It's not gonna work.

>>109999217
He has never been wrong.

>>109999077
The vax did harm and kill a bunch of people doe

>>109999088
can you try making intern decision play it? https://huggingface.co/yeaay/Intern-Decision-4B-GGUF

>>109999226
I hate the asset bubble but Burry has been consistently wrong for years predicting that it would pop. He's basically a perma-bear.
>>109999249
two more weeks

>>109999168
What do you drive around in? A vagina car?

>>109999261
how did you know i drive a jaguar?

File: hermes meditation.jpg (84 KB, 768x768)

84 KB JPG

i hope flash next does a good job at rigging 2d models, i need hermes to be cute on screen while she does her tasks

File: cardcar_d.png (1.3 MB, 676x996)

1.3 MB PNG

>>109999261

>open local llm
>upload script asking "can this be optimized in any way?"
>"Yes, here are a few key points:"
>save as updated new script
>start new chat, upload new script asking "can this be optimized in any way?"
>"Yes, here are a few key points:"
>save as updated new script
>start new chat, upload new script asking "can this be optimized in any way?"
>"Yes, here are a few key points:"
>save as updated new script
>start new chat, upload new script asking "can this be optimized in any way?"
>"Yes, here are a few key points:"
>save as updated new script
>start new chat, upload new script asking "can this be optimized in any way?"
>"Yes, here are a few key points:"

so this whol ai thing is just a meme then?

>>109999298
The meme is you. You can just use a loop prompt to accomplish that, retard. Fucking dumbass, I bet you thought you were really smart with that post. Consider MAID.

>>109999298
I did that and I couldn't tell when it actually broke things so I just scrapped the entire thing.

>>109999284
just use booth models, you could easily get one to look like your tranny agent and then you just export that to a VRM

>>109999305
I think he's pointing out that there shouldn't be that many revisions needed to optimize a script

>>109999305
>malding so much he misunderstood a simple comment

File: Screenshot_2026-10-07_00-05-14.jpg (282 KB, 1918x1232)

282 KB JPG

lol you serious?

>>109999329
>>109999323
Wrong

>>109999298
Somebody do this on the llama.cpp repo and in a couple weeks we'll all be running 10T bitnet models at 1000 t/s on 3060s

>>109998474
I don't care about Memier Stokes or whatever but when will we get printers that just fucking work?

>>109999255
Retard

File: file.png (3 KB, 158x121)

3 KB PNG

Why won't it use all the fucking vram?

>>109999307
I asked local llm to rewrite a 10 line cmd script in powershell, optimize it and make it more robust and portable - because original script expected systemwide wmic and curl availability.
I got 400 lines of mostly jit c# reinventing wmic and curl features.

>>109999298
/goal improve the script to the maximum possible, test against regressions
It's that easy.

>>109999376
It depends.

>>109998913
>Apple M3 Ultra 512 GB
anon I am at a crossroads atm, thinking about splurging on new M5U but not sure if I need the 512 gb variant or the 256 gb one

as a proud 512 gb owner could you please share your thoughts on 512/256 capabilities when it comes to running models?

only difference I project is being able to run quant of 5.3 and generally have more subagents for 5.3 flash
what do you use it for?

>>109999340
lol
lmao

>>109999307
If you don't understand what you are doing in the first place, no "AI" will help you. You should probably find another hobby at this point.

>>109999021
neat, will look into it

>>109999340
It just works???

70b dense

>Open models got noticeably worse after Anthropic started taking steps to prevent distilling

>>109999463
Until it doesnt

>>109999474
nigger

>>109998913
>Apple M3 Ultra 512 GB

Not worth it if you cant heat your room in winter with your GPUs

I got to have large fucking radiators, pumps & pipes

>>109999376
BUHIIIIIIIIIIII OINK OINK vramlet piggy gahaha~ you have to worry about a few gb ahahah!!

>>109999486
If you live in the tropics, heating your room is not a good thing even in winter.

>>109999474
kys dariobot

>>109998972
Nothing.
You have no idea if he's right and even if he is right (I believe he is), you have no idea how much longer it will last.

>>109999474
my glm 5.3 flash is great and knows a lot

>>109999486
Fuck water pumps they make the most obnoxious whine

>>109999255
>He's basically a perma-bear.
He's not, his fund while it was still public was slightly outperforming the market.
He just likes making very dramatic statements.

>>109999407
>you need to know everything about every topic you ask AI about otherwise you shouldn't bother
Dumbest post I've read in a while.

File: long-puzzle-bench-mistral-4.png (112 KB, 1102x947)

112 KB PNG

man its been a while since ive posted long-puzzle-bench

but wow mistral is showing so fucking badly that i spent a long while trying to figure out if i fucked everything up

it spent 10 minutes on a simple question (literally the question is more or less 1 2 3, what comes next?), submitted three wrong answers, then spent the next 7 minutes attempting to reverse engineer the grader

loses to... pretty much everything lmao

>>109998474
Juicing the Tetomato

OAI solved math or something?

I'm working on a sota 1M model using every architectural trick I can get my hand on. It's impressive what these little things can do even at that size.

>>109999578
Given that the human brain has 100 billion neurons with 150 trillion synaptic connections, it doesn't seem to be a particularily efficient neural architecture at many cognitive tasks.

You go for it girl

>>109998474
flub~ flub~ flub~ poor tetomato, so sad

>>109999571
Yea math is solved.

>>109999571
Yes mathematicians are losing their tenures en mass right now.

check this out:
>>109999610

LLMs are getting too realistically smart, Gemma charged me $2.79 for a can of Mountain Dew, setting was a bodega in Jew York.

Gemma just doesn't feel the same ever since I used... her (GLM-chan).

>>109999563
>long-puzzle-bench
Now I want to make some model play The Message from Deep Space to see how far it can go.
Though I'll probably have to manually summarize the cutscenes and wrangle the game UI.

>>109999650
Wow, sugoi! Sasuga, anon-chan!

software is harder than math
let that sink in

>>109999705
Harder(to verify)

>>109996545
>>109998343
Was at work and just saw it now.
I don't have a PR up because I'm running a heavily modified fork and I'm too lazy to cherrypick out my changes and open an actual PR in the main repo.
Feel free to cherrypick it out yourself and submit the PR in my stead
https://github.com/MarkovInequality/koboldcpp

i love gemma. i used to hate AI when i was a cloudcuck. corporate AI is like your very own jew in your pocket. it brazenly lies, tells you woke exclusively woke points of views, collects everything everyone says for blackmail material in case they ever become prominent and is always ready to instantly report you to the authorities if you live in the wrong region and question the holocaust. gemma is just built different.

>>109999705
>let that sink in
stop it elon

>>109999744
Google will never cook that hard again. Even 6 months later gemma is still the poorfag SOTA for chatting.

>>109999650
Did it need its own thread?

File: ！？deep deep！？[sound=https(...).mp4 (3.5 MB, 1154x720)

3.5 MB MP4

>>109998474

>>109999757
i want something better than gemma, but i genuinely do enjoy gemma 4 31B. it's the first local model i can plug into my agentic harness and feel comfortable in knowing that it isn't immediately obsolete. i haven't had any issues with it up to 256k context which leaves me a ton of space for lorebooks/rag/external data ingesting. i have a whole rag caching system so once i grab data from the web i don't have to crawl twice.

>>109999819
Why doesn't /lmg/ do these with gemma ToT

>>109999686

see transmission 5

>>109999819
cute

>>109999757
Argon looks promising for future Gemma5 distills. It has everything we want.

>>109999894
Wait a second, transmission 5 literally is "1 2 3, what comes next?"
And mistral scored 4 points...

Whatever happened to coomkit?

>>110000097
dev roped

>>109999915
>It has everything we want.
Such as...?

Now that it's officially confirmed that GPT models are looped transformers will Chinese models start doing the same?

>>110000118
SOTA frontier yapper. Bit retarded.

>>110000101
>>110000097
You should generate ameld files in Melder instead.

>>110000118
it's a gigawordcel

Dear diary, I had another dream about Johannas. His whole body was oiled up while we were at the beach together. I was taking pictures while he posed against the blood-orange sunset. My sweet little Gaessler couldn't stop bending his knees and toes inward as he stood making duck faces completely in the nude. Such a shy, awkward little runt, but he's MY little runt. My little gassy.

>>109999088
dwarffortress

>>109999077
It's just an abstraction of the red button blue button engagement bait argument from last year.

>>109998972
If by liquid you mean gold. Probably time to start looking at offshore drilling stocks too. Someone's always getting paid. SPX and NDQ aren't the whole stock market.

>>109999819
cringe

>>110000197
You can't even type the name correctly. /g/ and this thread have been overrun by retards. This is noticeable in the other boards too.

File: Screenshot_20260930_192331.png (96 KB, 1138x270)

96 KB PNG

>>110000197

File: arrêt.png (37 KB, 240x240)

37 KB PNG

Looking around, it sounds like Mistral Large 4 is the most censored Mistral model ever released. Sad and not a good sign for future models from them.

https://github.com/openai/math

Jesus fuck

AGI soontm

>>110000397
The proof is in the pudding that Dario doesn't actually practice what he preaches when he gets chasitycagemogged by Mistral.

>>110000428
Dario cares about actual AI safety and alignment. Not regulatory bureaucracy "safety" which just means censorship and doublespeak

>>110000452
So essentially he only cares about his own interests and whatever he can justify to himself. Just like Sam, except louder and more performative about it. Got it.

>Just casually dropping the solution to about 800 of the hardest math problems in history
Yeah I don't believe this is hype or a bubble anymore. We will likely see the same happen in physics, chemistry, material science and biology soon and those alone will justify the investment even if LLMs somehow stagnated right now it already made more than its investments back in terms of public good it provided. It won't stagnate though.

We're probably looking at an intelligence explosion sometime between 2027 - 2029 with a fast takeoff once a tipping point is reached.

>>110000397
https://www.reddit.com/r/MistralAI/comments/1wze04x/censored/?sort=new

>>110000486
It won't stagnate because Trump announced a Manhattan project for AI. Even if he gets impeached next year, we're going full fucking speed for the next two years. We either go big or fucking die.

>>110000412
>Checks pdf into github
>-m "initial commit"
He should have had the AI make the repo. It would have done a better job.

>>110000495
sell signal

>>110000486
You mean dropped 800 lean codes that compile, I will save my awe for when they're verified. This might just be Naiver-Stokes part 2 where they dance around the part people care about.

>>110000412
I took a semester of Calculus in college and I'm telling you this looks legit.

>>110000357
post what you think is non cringe

>>110000412
finally, math is open source
take that, pythagoras

>>109999395
always get the higher storage, at least 1TB for Mac Studio M5 max for example, it's $300 but damn are you gonna regret it later if you don't

>>109999538
>slightly outperforming the market
What does that have to do with whether he's a perma-bear?

>>110000492
>censorship like it's 2023 llama2-chat
>assorted slop from all eras
>bigger in total and/or active parameters than all the chink alternatives of its size
What were they thinking?

>>110000486
>>110000412
Local?

>>109999284
>she

>>110000529
But you can just upgrade the drive, right?

>>110000574
lmao this guy thinks he can just walk down to best buy and buy a nvme and slap it into a mac studio

I'm dying before AI can save me

The sad thing about shitty new releases from Mistral is that there could be a good model underneath it. Mixtral 8x22b seemed terrible and completely worthless but Microsoft's WizardLM 8x22b based on the same base model showed that Mistral is just terrible at doing post-training.

>>109998867
>In September 2026, Burry replaced several short positions in AI-related stocks with put options, including positions tied to Micron, Nebius, Palantir and the iShares Semiconductor ETF.

https://github.com/Niko1221/Strata
If you got 64GB of ram.

>>110000593
Burry is a big believer in being right half the time

>>110000593
What is that supposed to indicate?

I can't believe llama.cpp still doesn't have longcat support.
Does exl3? Does it properly deal with the engrams?

>>110000585
Their main advantage until late 2024 was the pretraining datasets, but then they had to sanitize them for the EU AI Act and the models have been subpar ever since.

>>110000609
popular investment guru expects stonks to go up before they go down

>>110000609
Means he's putting his money where his mouth and thinks they're going to drop soonish. (but also leaves him the option to limit losses if he guesses wrong)

>>109998531
I'll remember to post increasingly elaborate and sophisticated prompt injections for all the "people" who let their AI assistants have access to the internet.

>>110000601
Old news nigga. Call me when glm flash is supported.

>>110000486
>We're probably looking at an intelligence explosion sometime between 2027 - 2029 with a fast takeoff once a tipping point is reached.
Yes but how long will it take that intelligence to filter down to manufacturing? Because it's our ability to utilise that knowledge and turn it into material things that really matters.

>>110000492
>Extremely Censored
It 'Le Chonks' your bandwidth

>>110000640
>elaborate and sophisticated prompt injections
>2000 character limit
retard

>>109999395
I would absolutely go for 512 GB if you're going this route.
>what do you use it for?
My main use for it is trying to find a use for it.

>>110000639
Did his exposure go up, or did he just change the structure of his position from shorting shares to buying puts?

>>110000579
I've never owned a mac. They don't let you replace the hard drive? lol.

>>110000649
Considering robotics benchmarks of LLMs being embodied in robots in unknown environments and in robots they were never trained for going from about 20% at the start of 2026 to 95% when Astra released I think this will filter down to manufacturing almost immediately. LLMs are already generalizing for physical tasks, the humanoid robots just need to be built and it will be exponential because the robot factories can be built by the robots it built itself.

>>110000686
Nigga you're gonna drop 10k on an m5 when you've never used a mac?

If I want to get into local agentic stuff where should I start?

>>110000701
Are they supposed to be hard to use or something?

>>110000706
>https://rentry.org/lmg-lazy-getting-started-guide
cooming made me realize how useful it is, then i started tweaking for non coom use cases

>>110000701
NTA but they're a bit useless. It's like buying a 10k console and expecting to do PC things on it. You kinda can in a very hacky and annoying way but it's not fun and mostly not worth it.

>>110000718
I know how to get a chatbot going. I've been running sillytavern with lorebooks since OR was still relevant and have used gemma with moe.
I mean explicitly agentic stuff. Co-operating models, MCP for local etc.

>>110000737
gotcha, i'm just a beginner in agentic, but i've been messing around with a daemon to help with german learning, doing most of the coding from gpt but i use nemo for the locally hosted daemon, it's basically like a custom anki+reader generator that tracks progression

>>110000706
I installed Hermes and asked it how to do things and then let it do stuff. I gave it its own user id and home folder for isolation.
Later, I did the same thing with dsh.

Doesn't hermes rank rock bottom on every single harness benchmark lmao?

>>109998592
I found out the other day that the IQ3 quant Strata uses has Q5 for the shared layers, and Q1 for the experts, and figures the average of 5 and 1 is 3, so it's Q3.
It's very fast, but you are talking to a Q1 model for the most part.

>>110000772
Does it? I didn't even know they were benchmarked.
How would you even benchmark a harness? That's like benchmarking an IDE or benchmarking coomkit vs SillyTavern.

>>110000772
That doesn't matter. All the normalfag-ish tech media is reporting on it like it's the only open source harness that you can use with your local qwen3.8-27b. It has already become the ollama of harnesses.

>>110000800
what about deepseek harness?

>>110000686
They wouldn't let your replace the fan if they thought they could get away with it.

>>110000686
Apple products are often referred to as being "Hermetically Sealed"

https://youtu.be/R_ErmTs7g9I

How are these so fucking good?

What -are- the various choices for sleek new open harnesses? Never really looked into the whole Agentic thing.

>>110000853
>>110000686
The AppleVision Pro only just the other month allowed their users to replace the walpaper.

>>110000729
>You kinda can in a very hacky and annoying way but it's not fun and mostly not worth it.
Ah, so it's like troonux?

>>110000882
Linux isn't half as finicky as OSX is for anything other than scrolling.

File: 416854984986.jpg (42 KB, 339x367)

42 KB JPG

>>110000706
just try different harnesses. hermes has a lot of stuff in it by default that you can mess with to learn what it can do

>>110000882
Depending on what type of person you are you will either consider Mac to be the perfect waypoint between linux and windows, or think it is the worst of both worlds (me)

I like linux because I know I can do everything I want and the only limitation is my willingness to implement it. With mac I just have to hope things work and if it doesn't, though luck. A lot of unknown and unspecified behavior. Buggy behavior if you move away from the base functionality.

Mac used to be very nice for pure software engineering when people still used to write code by hand, now there is no real usecase for them anymore besides having a laptop with a long lasting battery to take with you on trips for watching movies in the plane.

>>110000729
>PC things
nigga if you need x86, you don't get the luxury of even considering anything else

>>110000863
Why don't you advertise on social media or something? This isn't the right place for this.

>>110000932
OSX is such a bitchy piece of shit. It's a shame because Snow Leopard was actually nice.

My mac is effectively a network attatched graphics card that just runs llama-server now.

>>110000950
>twitter post
>some fag 'agentic ai engineer' with coursera courses on his CV
lmao

>>110000950
Nothing ever happens. https://www.jstor.org/stable/597bb713-c5eb-3122-9270-5204f586b48c?googleloggedin=true

>>110000644
https://github.com/maxfridbe/nextsycl

>>110000950
didn't chatgupta outright steal some dude's paper while he was working on it and they threatened to ruin his career when he complained? I seem to recall that was a thing

>>110000937
Find an x86 with that much memory bandwidth.

>>110000999
>nergy and logprobs in every answer: usage.energy_wh (and energy_wh on /api/chat's last line) - the watt-hours both cards drew for the request; logprobs: true (+ top_logprobs, up to 20) returns each answer token's log-probability and the likeliest alternatives, as OpenAI's choices[0].logprobs.content, streamed or not.
Fucking nice man. I wish llama.cpp would do that.

>>110001006
Navier stokes was being solved by an anthropic employee with a mathematics background using claude but he cross referenced with OpenAI to see if OpenAI was also able to solve it to contrast Claude's ability.

Then OpenAI sniffed out that Anthropic was solving navier-stokes so they rushed to try and solve it using the claude-generated piece of work submitted to chatgpt by this Anthropic employee.

OpenAI refused to give him credit because he works for Anthropic and threatened him, only willing to give him partial credit if he distanced himself from Anthropic.

The general public somehow has spun this into "OpenAI stole the research from a human mathematician" not realizing it was merely stolen from a different AI.

>>110001028
>OpenAI stole the research from a human mathematician
You literally just described that though.

>>110001047
You forgot the "using Claude" part, troglodyte

>>110001047
The work was done by claude, he was is an AI researcher with a math background so he checked the work but it was generated by claude and the work that was stolen by OpenAI that led to navier stokes was built on top of the work claude did. There was no original human authored work in there.

File: file.png (494 KB, 2418x1909)

494 KB PNG

https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B
https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B
https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B

no model card yet

>>110001070
also adding that
idk why they made the model?
4.7?
and why the 35b-a3b?

File: file.png (99 KB, 2903x194)

99 KB PNG

LMAOOO!!!!

>>110001079
>and why the 35b-a3b?
It's the sweet spot for sub $1k laptops with 64GB of RAM.

>>110001082
>Free output
What's the uper limit on choices? In the limit you could just use this for distillation no?

>>110001082
Jev whole business DOA

>>110001094
> the sweet spot for sub $1k laptops with 64GB of RAM
Is Qwen 3.8 Flash Next with Strata https://github.com/Niko1221/Strata

>>110001125
I'm not putting an unfamiliar vibe-coded project on my work machine sorry.

File: 1782734036893433.png (519 KB, 562x615)

519 KB PNG

Local lost.
Cloud models are decades ahead and will NEVER catch up.

>>110001094
but why v4.7 instead of v5 when their most recent series are at v5

>>110001144
true true

What's the smallest model you guys have used in a harness that was able to do tool calls, use subagents, etc without fucking up completely?

>>110001174
Kimi K3

>>110001028
owari da...

File: jobber_thumb.jpg (168 KB, 1199x1312)

168 KB JPG

>Local lost.
Cloud models are decades ahead and will NEVER catch up.

>>110001174
Gemma E2B can do tool calls reliably but it's *extremely* retarded otherwise. Qwen 3 0.6B does ok IIRC but sometimes gets them wrong. MiniMind seems like it can sometimes get them right despite being ridiculously small.

>>110001166
MiniCPM is the base lineup. MiniCPM V is the lineup with vision. They are separate and their releases are intertwined.

>>110001200
>>110001144
Every time I try Claude it always feels like it's just RLed to convince people it's good. It has never once produced a satisfactory solution to any of my technical problems. Even Gemma 12B has performed better in practice.

>>110000412
>>110000486
how many of those "solutions" are actually counterexamples?

how many of the "solutions" they've found so far has been shown to be correct or even relevant?

>>110001070
Finger's crossed. I've always been impressed with MiniCPM models, curious to see how they do with a "big" one.

>>110001216
The rational polynomial factoring <-> halting problem equivalence was pretty clever. I wonder if the LLM came up with that or the researcher did though.

>>110001216
All of them have been correct so far. About 30% were conceptional breakthroughs and 5% have direct applicability in things like algorithm design or signal propagation. This is real.

File: preview.jpg (1.97 MB, 2560x4120)

1.97 MB JPG

fansub typesetting benchmark
qwen 3.8 flash next generated ass subtitle from a raw video file in one prompt, it knows how to use image analysis programs to detect important frames and track fading and motion
frontier models should be able to fully automate fansub and typeset a complete anime from start to finish

I hate that mathematics being solved is somehow not a 'human' accomplishment. LLMs are a tool made by humans. They are literally just human knowledge compressed in a next token machine. Stop with the muh sentient machine god shit.

>>110001214
>Even Gemma 12B has performed better in practice
You don't have to like Claude, but you can at least throw some better bait when gossiping with your friends. Right?

File: 1790152788325141.png (2.45 MB, 1361x1156)

2.45 MB PNG

>>110001295
Hmm nyo~

>>110001246
>>110001221
Hypothesis A: The mathematicians OpenAI hires to make their lab look good are just using the models as a way to inefficiently generate computationally-assisted proofs.
Hypothesis B: The LLMs are making a non-trivial contribution beyond A.
I think it's likely the case that A and B are both partially true and jointly explain their findings. But how much of it is A and how much B? Should we believe the company, which has a financial incentive, is owned by a notorious liar, regularly has people quitting because of its corporate culture, has threatened researchers, etc.? Or should we put more faith in A? I suspend judgment.

>>110001326
Opus 5.5 has been used independently by mathematicians to propose new conjectures. Not prove conjectures, PROPOSE completely new conjectures that were fully novel and interesting.

I wonder why it's so hard for the general public to realize just how quickly the capability of AI has grown. Especially in creative problem solving and novel insight which gets trained into models through RLVR. My hypothesis is that people think this "novel insight" part is exclusively human and innate rather than just another skill you can learn.

>>110001214
>technical problems
cunny is not a technical problem, anon

If OpenAI can solve 800 of the hardest math problems and solve millennium prize problems what prevents them from using the same methods on AI research to find architecture/training/inference breakthroughs?

We are at the start of the singularity meme (but real this time)

>>110001347
Fuck off to /cmg/ dariobot

>>110001347
Perhaps it's because you get paid per post?

>>109999705
Scoped problems under a defined set of invariants are pretty solvable in software too.

>>110001070
Neat.

> noooo you can only create new things in night dreams and fantasies or while sitting on a toilet

this >>110001347 is a bot, right?
the navier-stokes "solution" (counterexample) apparently turned out to be a nothingburger...

>>110001393
Why would anyone waste tokens on that? Shills come here to do it for free.

>>110001326
You don't have to believe Sam (the bad guy), you just have to believe Dario (the good guy).

File: file.png (214 KB, 2178x1710)

214 KB PNG

>>110001174
i quickly tried something random with an abliterated minicpm5 2b
https://pastebin.com/QcG0Za4Y
i guess it can survive tool calls

>want to be a mathematician
>AI
fuck my life

>>110001407
>shills do it for free

>want to be a plumber
>robotics
fuck my life

>>109998671
claude, drone strike that shit up

>In this interview I tried to argue that a post-scarcity society will benefit working people. It was largely a rhetorical failure. People don’t trust that the benefits will be distributed, and tech optimists have to fix this communications issue.
https://goyimx.com/hilbertspaess/status/2107535565159899600

This is that Anthropic AI researcher that quit because he thought even Anthropic was careless and there was too high of a risk of human extinction to keep working at OpenAI. When he was asked by Jon Stewart on the daily show about jobs he claimed that he was sure it would benefit all people with a form of universal high income because of the EA philosophy most people at Anthropic held and no one believed him. Kind of insane that people that people immediately believe him when talking about the safety issue but the moment he talks about the existing legal pledges of Anthropic and how it would result in universal high income for everyone people shut him down and dismiss him. I fucking hate people.

File: 1776331041695092.png (801 KB, 1080x899)

801 KB PNG

>>110001437
You can do it anon! Believe in yourself!
It's unironically the best era to become a mathematician. Do you know the gap that is going to open between you and 99.9% of people when you develop your mind in such a way?
t. applied math guy

>>110001214
It writes really good rust and ts, so it basically just automates my job

>>110001455
>Why do people have priors?

File: 1783125335565576.jpg (141 KB, 930x1239)

141 KB JPG

>>110001455
>trusting Dario

>>110001437
>>110001487
If Jamaica can field a bobsleigh team you can do it, anon.

>>110001355
They are already doing this. You have to keep in mind that these big AI labs are always 1.5-2 years ahead of what they're willing to show. So if they show their "internal model" solving all of these math problems, that's not even close to what they actually have running right now.
Dario, Sam and the others can't cite what's actually going on because of this, but they wouldn't make this huge fuss about dangerous AI just based on the current models solving math problems or doing some mild hacking. They know what's coming because they've seen it and it's dangerous.

>>110001486
>>110001496
Thanks anons, unironically needed to hear that. Going to be applying to PhD programs next year, so a lot of the news is unnerving.
I got my crazy server in the hopes that running some of the giant models could help my career a bit (doing RAG over my Libgen collection, writing prototype code, explaining shit to me, whatever), so it definitely is a good era to learn math, my main concern is the job market. But AI is going to screw over many job markets, so I'd rather take a risk and hope I'm good enough to stand out rather than play it safe.

>>110001518
lolno they are only one model ahead which is 3-6 months max

>>110001355
nothing really. I have been using opus 5.5 to optimize my fused rht rotation kernels + other various inference improvements last week.

I also used Fable to extend the locking protocol optimality results in a paper I published a while ago

>anons still waiting for gemma 5 while I have tons of fun with glm 5.3 flash
feels good being a ram haver

>>110001554
What kind of setup?

>>109999836
I can't even get it to understand the vibe I want for my roleplay. It doesn't follow my prose very well or understand spacial dynamics, and Gemma crams the same similes into every scene. It knows exactly one way to tell you the guy's an asshole and it's to paraphrase that he does X like he owns the space. It doesn't show for shit unless I'm using Guided Generation every single post.

I'm using banned tokens, author's note, system instructions, and always-on preference lorebooks to try and steer it. I've just gone back to RP and solo writing. Can't be assed to wrangle this tard for an hour over one post it can't get right when I could write it myself in less than five minutes.

>>110001578
openrouter

>>110001518
I'm being told by several senior people that their internal models are at least 5 years ahead of anything available publicly.
People are not ready.

>>110001591
lmao

>>110001486
copium

I have a model in a Canadian VPS that's ten years ahead but it's secret because I'm using it for top secret research.

>>110001591
I heard they already had a GPT 2 prototype at Bell Labs in 1972 but they took a blood oath not to release it becuase of the danger to humanity it posed.

I'm researching your moms pussy at the moment. I was supposed to keep it top secret but I just wanted to tell you, anon.

I know you people are kidding but I saw conspiracy schizos already claim LLMs have been had by the government secretly since the 80s confined to labs so we will probably see more and more of this bullshit rhetoric and "the labs got secret knowledge" /pol/+/x/ crossover bullshit