正在加载视频...
视频加载失败
Me explaining the task to a 1B model:
130,321 次观看 • 13 天前 •via X (Twitter)
42 条评论

this is exactly what happens with grok 4.7

hahahah

use

what’s that? 👀

A way to write instructions and dynamically generate them for the model. When you speak to @grok, you need less text than when you speak with gemma4. So the skill and the goal are not plain MD files, they are being generated depending on the env.

@dedene Adaptive instructions by model and env make sense. Static files lose relevance fast across capability tiers. Leaner text works better here than denser hand-holding for smaller models. Dynamic skills keep the context sharp without the entropy.

😂1B has a long way to go

someone should try to implement 1 billion parameter model inside a TI 64

Use JEPA.

Can we use it already?

[2610.39302] Proving and Improving the Platonic Representation Hypothesis with JEPAs

@dedene "hey hermes, implement this paper to instruct claude code to update my website"

@wavedecompiler 😂

Tbh this is what it feels like testing the affordable mid range model the labs tell me is like 2% worse than the SOTA

What if we have 1B model with internet (search-first) and evidence based (memory)! That may work better or no!?

Quite literally.

Me using a Q1 on my phone

you just need to get your bigger models to live surgery on the weights.

True story, even with Gemma 4😂

Or a frontier model explaining the tool it just made to you...

Do not! I repeat: DO NOT explain your solution!, explain your problem. You'll be fine unless you have a problem nobody had before.

u should use 1.7b

😂😂😂

sounds about right 😂 but hey thats its purpose! We don't need to bring out big Groot just for a few button presses

Hey guys, no matter what size the model is, just dont give him the option of that kill button.

Felt the same while testing Bielik in agentic loop setting 😅 Also reminds me the GPT 3.5 Turbo era, the models were so dumb not so long ago. What a time to be alive!

That's like extreme sport 😆

sounds like grok

accurate

try reverse psychology with the model. lol

some times in real life, you have to talk to your new guys just like that.

You have to finetune the model, poor thing cant understand the "explaining" 😭

🤣

And it still messes it up 😭

Try qwen slm they are good

Just get a several T model adjust its weights 😜

well this is also a nice presentation of that it's our responsibility to help the model :D. I saw the movie and love it, but in this context why put two red buttons that do the same thing next to each other? We need to make sure they can't make a crucial mistake this simply.

just remove the other button? problem solved, no confusion for groot

That is me explaining the task to a nerfed ds 4.1 flash 😅

I truly hope in the near future even 1B models will be as capable as today's frontiers. They just need a great learning/cognitive core and can look up and learn everything else.

🤣🤣🤣

I think you need 10T params before your model has a button that can kill everyone.
