-Automaticity

joined 10 months ago
sorted by: new top controversial old
This model is extremely good in c/localllama@poweruser.forum
How to upgrade to the next VRAM breakpoints, and is it worth it? in c/localllama@poweruser.forum
[–] -Automaticity@alien.top 1 points 9 months ago

If Nvidia isn't upgrading GPU's past 24GB for the RTX 50 series then that will probably factor into the open source community keeping models below 40b parameters. I don't know the exact cutoff point. A lot of people with 12gb VRAM can run 13b models but you could also run 7b 8-bit with 16k context size. It will get increasingly difficult to run larger contexts with larger models.

Some larger open models are being released but there won't be much community there to train on a bunch of datasets to the huge models to nail the ideal finetune.

X.AI Grok could potentially be open sourced on a 6 month delay from launch in c/localllama@poweruser.forum
[–] -Automaticity@alien.top 1 points 9 months ago (1 children)

I've seen enough examples of it ragging on him actually, Lex even had Grok criticize Elon on that podcast. Elon even got owned by his own community notes feature a few times on Twitter too.

X.AI Grok could potentially be open sourced on a 6 month delay from launch in c/localllama@poweruser.forum
[–] -Automaticity@alien.top 1 points 9 months ago (5 children)

He open sourced the Twitter algorithm.

Grok-1 is a completed model which isn't equal to compare to something that's been in extended development.

And I will say again. He open sourced the Twitter algo.

1
X.AI Grok could potentially be open sourced on a 6 month delay from launch (alien.top)
 

Nobody mentions this from Lex Fridman. https://youtu.be/JN3KPFbWCy8?t=5090

Maybe after he gets the infrastructure needed to fully support all interested subscribers he might consider open sourcing Grok-1, or maybe he might release Grok-1 when he is ready to deploy Grok-2 on X. I doubt 6 month delay for open source following commercial release is an exact mark but it remains possible for Grok to become open at some point.

As of right now access to Grok on X is too limited for us to get any independent benchmarks and attempts to search for Grok on Twitter just flooded with unaffiliated memecoins. Post comparisons between Grok and GPT3.5 or GPT4 are limited and seems subjective to opinion. My observation is Grok can give you serious responses without the "humor" so we should be able to prompt it to not act "humorous". It's responses aren't large as GPT4 but it can be very concise with half or 1/3rd the written content while not lecturing on disclosures.

When it comes to being uncensored. I noticed benchmarks for original and uncensored Llama based models with the uncensored models losing points in logic which suggest that being uncensored or minimally censored from the foundation is better off. The technical details of uncensoring an existing model is beyond my expertise.

So what are thoughts about Grok open sourcing, without turning this into drama. Be civil please.