this post was submitted on 10 Aug 2025

178 points (96.8% liked)

LocalLLaMA

localllama@sh.itjust.works

5,625 readers

3,221 active

122 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

Rules:

Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.

Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.

Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.

Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.

founded 3 years ago

moderators

MonsterBug@sh.itjust.works

Smokeydope@lemmy.world

pax@sh.itjust.works

noneabove1182@sh.itjust.works

When DeepSeek V4 and R2? (sh.itjust.works)

submitted 1 year ago by pepperfree@sh.itjust.works to c/localllama@sh.itjust.works

26 comments fedilink hide all child comments

all 26 comments

sorted by: hot top controversial new old

[–] Kyrgizion@lemmy.world 29 points 1 year ago (1 child)

I remember how we would get into trouble for copying each other's homework in high school.

hideshow 1 child comment

[–] riskable@programming.dev 10 points 1 year ago

Now we get in trouble for generating each other's homework 🤷

parent

[–] mtpender@sh.itjust.works 14 points 1 year ago (1 child)

OK Tech bros, repeat after me:

"Thou shalt not make a machine in the likeness of a human mind."

hideshow 1 child comment

[–] pepperfree@sh.itjust.works [S] 4 points 1 year ago

parent

[–] OmegaLemmy@discuss.online 13 points 1 year ago (3 children)

I wonder if a good fine tuned model beats every general purpose LLM if you need it for a really specific purpose

hideshow 3 child comments

[–] laz@lemmy.dbzer0.com 20 points 1 year ago

Yes it does

parent

[–] kata1yst@sh.itjust.works 8 points 1 year ago

Of course, and this is why the new hotness is a Mixture of Experts for one model that is effectively a bunch of experts arguing over the answer, or else on a different scale there's the Combination of Agents where different specialized agents perform specialized tasks.

parent

[–] pepperfree@sh.itjust.works [S] 1 point 1 year ago

There is new project which they share fine-tuned modernbert on some task. Here is the org https://huggingface.co/adaptive-classifier

parent

[–] Endmaker@ani.social 6 points 1 year ago (4 children)

I thought DeepSeek's selling point was efficiency?

hideshow 4 child comments

[–] laz@lemmy.dbzer0.com 13 points 1 year ago (2 children)

It was beating near frontier models at tenth of a cost could be hosted by you or any cloud service for you and the open weights were not censored for China's pr and could be jailbroken to say write code for shady stuff which any frontier model would refuse.

parent

hideshow 2 child comments

[–] mindbleach@sh.itjust.works 10 points 1 year ago* (1 child)

You dropped these: ,,,,,,

parent

hideshow 1 child comment

[–] Kissaki@programming.dev 7 points 1 year ago

No time for commas with how fast this tech is developing

parent

[–] Eyekaytee@aussie.zone 5 points 1 year ago*

AI as a tech product is advancing faster than any other tech I've ever seen, you mentioning DeepSeek's revelations from like 8 months ago already feels like an eternity

parent

[–] Jankatarch@lemmy.world 5 points 1 year ago

(None of them are THAT much better than previous, even worse in some areas)

[–] DragonTypeWyvern@midwest.social 4 points 1 year ago (1 child)

Wait a second.

Grok's symbol is a lightning bolt?

hideshow 1 child comment

[–] lastunusedusername2@sh.itjust.works 2 points 1 year ago

Grok 2 will have two lightning bolts

parent

[–] TriflingToad@sh.itjust.works 3 points 1 year ago*

how about you grok deez nuts (it is 1 am please forgive me)

[–] Eyekaytee@aussie.zone 2 points 1 year ago (8 children)

The most surprising thing about this image is that Grok is on it, they started way behind the 8ball and have caught up

hideshow 8 child comments

[–] magic_lobster_party@fedia.io 9 points 1 year ago (2 children)

It’s flip flopping between being a full on Nazi and arguing against MAGA with facts and logic.

parent

hideshow 2 child comments

[–] pepperfree@sh.itjust.works [S] 1 point 1 year ago (1 child)

The RL is so good grok changed it's personality by changing small part of it's system prompt

parent

hideshow 1 child comment

[–] mindbleach@sh.itjust.works 2 points 1 year ago

Anakin: It's aligned.

Padme: To be good?

parent

[–] pepperfree@sh.itjust.works [S] 1 point 1 year ago (4 children)

It changed after Grok 3

parent

hideshow 4 child comments

[–] Eyekaytee@aussie.zone 0 points 1 year ago (3 children)

yeah i hate musk but the fact grok was launched in nov 2023 years behind the competition and has caught up is shocking

parent

hideshow 3 child comments

[–] pepperfree@sh.itjust.works [S] 3 points 1 year ago (2 children)

They got the whole Twitter database. It's kinda the same with Gemini. But somehow Meta isn't catching up, maybe their llama 4 architecture isn't that stable to train.

parent

hideshow 2 child comments

[–] veroxii@aussie.zone 5 points 1 year ago (1 child)

Or maybe Facebook data is even worse than Twitter?

parent

hideshow 1 child comment

[–] pepperfree@sh.itjust.works [S] 0 points 1 year ago

Llama 3.3 was good, tho. For the multimodal, llama 4 also use llama3.2 approach where the image and text is made into single model instead using CLIP or siglip.

parent