I Taught an Open Source AI to Think Like OpenAI o3 (here’s how)

[@bunbox]

Going to give this a shot as soon as I get home. Thanks Austin, you make some of the best and most actually useful tutorials out there when it comes to opening up Open-Source AI space. Everyone else doesn't really give anything functionally useful, where as you give great stuff others can then build ontop of.

[@SCHaworth]

This is precisely how i thought it should be done, but i had no idea how to do it. Appreciate you.

[@xAgentVFX]

GOAT. I tried a very crude version months ago when you dropped the Memory Agent tut. Just made it reflect on the prompt and it's initial thoughts 3 times before the final response was sent, but hooking it up to 405B was basically triple the price

[@gnsdgabriel]

Nice video. Thank you for sharing.

[@flow.philosophy]

I'd be really interested to see how this performs compared to the vanilla model, compared to o1, etc. I realize 3.17b isn't sota, but I wonder how far just the virtue of the cot process will carry it

[@mixmax6027]

Deepseek and others. Hope it breaks openAI

[@JustinJohnson13]

With Deepseek R1 out now, how does this compare?

[@DantePowell]

probably have to build this out yourself and do some testing to see. i am going to do this myself.

[@Mono_Autophobic]

So basically : Website : censored, o1 quality but censored Locally run (671b) : everything happened upto July 2024, literally everything you can ask and it will answer with o1 quality (yes even illegal subjects) but only problem is you either need 1tb of RAM or 700gb of VRAM, which is hella expensive

[@Mono_Autophobic]

Alternatively if you want something that can run at 24gb vram (4090), you can use 16b or 32b sized deepseek r1 models, but those are at level of gpt 4o not o1

[@lio1234234]

Have you attenpted this with the non-finetuned llama model? I'd have thought when it comes to training a model specifically for generating the reasoning steps, training off the pretrained model would be better, no?

[@Ai_Austin]

i have not. it would require a much larger dataset and won't be able to work with Ollama. but could absolutely yield better performance if you wanted to put in the time to create that large diverse dataset.

[@Mono_Autophobic]

So basically : Website : censored, o1 quality but censored Locally run (671b) : everything happened upto July 2024, literally everything you can ask and it will answer with o1 quality (yes even illegal subjects) but only problem is you either need 1tb of RAM or 700gb of VRAM, which is hella expensive Alternatively if you want something that can run at 24gb vram (4090), you can use 16b or 32b sized deepseek r1 models, but those are at level of gpt 4o not o1

[@websurfer-x2o]

Deepseek r1 lovers like here😂