@dj4aces
Safe, locally-sourced AI! From server farm to table!
@andre-le-bone-aparte
@01:57 - Do we want: Offline (Local) AI Video Series.... YES
@SuperMari026
Wendell: Safe local AI!!! My Ada 2000 card: I'm tired boss
@DrevisNEWS
Yes... Please a series would be great!
@TekmanRO
What I really couldn't find anywhere and I think you guys are well positioned to build is a live benchmark dataset across various hardware configurations (ranging from an M2/M4 mini, through entry level GPUs like 3060s, 3090, multiple 4090/5090, etc. to more professional ADA and all the way to the H100, as well as various CPU-only configurations and whatnot) against a large set of models in various quantized and distilled versions. It would be super useful for everyone delving into this space to have access to a quick reference on what hardware can run what model.
@tehsimo
When Wendell is this excited. My ears perk up
@PreRendered
More Wendell AI videos would be extremely welcome!
@Nobody-vr5nl
"It's never going to end up on the cloud" *Microsoft taking 4 pictures a minute of everything ur doing.*
@phantom8700
Speaking on the Industrial revolution, it also introduced new machines of warfare that devistated two world wars.
@lhl
One interesting thing about those "distill" version is they're not even really model/knowledge distillations in the classic sense, which typically do training to match the soft outputs (logprobs). The way these distill models were trained were with a straight output-based fine tuning/output imitation training. While the latter is fine (as we're seeing with LIMO, S1, etc can be incredibly effective), I do think that it's still worth pointing out the distinction just because different things are actually happening with these "distill" models are not doing the same thing as "true" distillations (which other groups like Arcee and Mobius and others have started doing and getting better results on!)
@Strykenine
Running locally is the key. Cloud-based AI may be more powerful overall, but if AI is being trained on all of humanities collective knowledge then it should be collectively accessible to humanity.
@claybford
Love it. Please make more of these videos!
@prosetheus
Superb video. Absolutely would watch a series. Really exciting times. Ive been checking your channel daily for this video. You did a great job cramming in so much dense info into such a small video
@jwiz1911
Would love to see performance comparisons for different hardware budgets: $500, $1000, $2000? Do I buy used servers with lots of cores? A single modern GPU? Multiple old GPUs?
@humblekhadjit
"In an AI world, it may be possible for that sole proprietor to establish themselves, grown an audience, and build a company; and capture a billion dollars of valuation... I think this is a good thing" I think your judgement is clouded by the novelty of the idea of AI.
@fourthplanet
Llamas are dangerous, so if you see one where people are swimming, you shout "look out, there are llamas!"
@H3adcrash
Noo thank you. I'mma just meander back to 1999 and start over.
@jonathanellis6097
Im still not entirely sure what i would actually use AI for. So much effort and hatdware power, to do what?
@Jcewazhere
We're on the cusp of another bubble burst. The entire stock market is massively over-valued thanks to unending stock buybacks and endless angel investors propping up failed companies. Maybe Deepseek will be the proverbial straw that breaks the back and pops the bubble. I played with Deepseek a bit on my 'gaming' laptop. i7, 3060 laptop version, and 16gb ddr4 ram. Had the 7b version up and running in about 10 minutes. Then turned off the wifi and asked it questions. It got most of them wrong, but convincingly so. It's scary how good it is at explaining its reasoning and gaslighting you. It remembered my first question in the second question on roughly the same topic. Just like any tool for some stuff it'd be great, for others make sure to double and triple check it or use a different tool.
@Level1Techs