Fahd, your channel has become one of my most liked. I really love that you do profound research and actually show how to use the tools. Unfortunately, but fortunately for you, not many people do that on TH-cam. Please keep up the good work!
sir I have been a follower for a few months and the knowledge you impart us is invaluable. I have a question how can you make videos this quick and deploy them?
Appreciate the detailed breakdown! A bit off-topic, but I wanted to ask: I have a SafePal wallet with USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). What's the best way to send them to Binance?
Hi Fahd, which self-hosted AI model would you recommend to run on modest hardware resources and still produce reasonable output? Thank you for your simple, clear, and brief videos :-)
Their last coding model was amazing, the only one which capable repair my code, i tested with qwen and marco. That was 236 billions (270Gb Ram in q8), the new V3 of 671 billions is very huge, bigger than Meta. I'm waiting for quantized versions but even on my 12 Ram slots motherboard i could run maybe q3-q2 quality. To run full quality quantized like q6 or even q8 i think needed ~750+ Gb Ram (usually it's equal to model size + little more).
I am still uneasy about adding CoT and other "steering" mechanisms directly into the token patterning of the model itself as these really seem to me not only alignment (Like DPO or heuristic analysis) but forcing a "core" mental model on the users of a mechanism that is SUPPOSEDLY looking to achieve generalization and wide utility. We're back to 1981. AI and Anthropic are Apple and MS, with Meta, MS and Google struggling not to become CompuServe and AOL... Prepare for some really shady shit... then with AI/Crypto... the browser wars were only an exercise. You ready Fahd? I am... and I am rooting for the little guy. These are all first-outers. They are all IBM, Commodore and Tandy... or even Solaris ;) They will never be the new Apple acting like the present one or the new MS with the same shady licensing and force fed, all batteries included schemes and strategy. With open weights and code however I can live with the "opt out" choice, but with closed models this will be even more of a conflict (at least for me).
Fahd, your channel has become one of my most liked. I really love that you do profound research and actually show how to use the tools. Unfortunately, but fortunately for you, not many people do that on TH-cam. Please keep up the good work!
sir I have been a follower for a few months and the knowledge you impart us is invaluable.
I have a question how can you make videos this quick and deploy them?
Appreciate the detailed breakdown! A bit off-topic, but I wanted to ask: I have a SafePal wallet with USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). What's the best way to send them to Binance?
Hi Fahd, which self-hosted AI model would you recommend to run on modest hardware resources and still produce reasonable output? Thank you for your simple, clear, and brief videos :-)
I love your channel lol. I started using your code language recommendations, Qwen is amazing,. Deepseek v3 is next level.
Why are they benchmarking it against Llama3.1 and not Llama 3.2 or 3.3?
That's a valid question.
Can you please show i can use it with cline
aicodeking
let's have a vote... conda vs pip. :)
Their last coding model was amazing, the only one which capable repair my code, i tested with qwen and marco. That was 236 billions (270Gb Ram in q8), the new V3 of 671 billions is very huge, bigger than Meta. I'm waiting for quantized versions but even on my 12 Ram slots motherboard i could run maybe q3-q2 quality. To run full quality quantized like q6 or even q8 i think needed ~750+ Gb Ram (usually it's equal to model size + little more).
is it ram or vram?
@santiagomartinez3417 are you millionaire? Of course Ram DDR4. V2.5 coding you can run on 22 core CPU with okayish speed.
@@fontenbleau What tool do you use for that?
Any app based on LLama, you can't write them here.
I am still uneasy about adding CoT and other "steering" mechanisms directly into the token patterning of the model itself as these really seem to me not only alignment (Like DPO or heuristic analysis) but forcing a "core" mental model on the users of a mechanism that is SUPPOSEDLY looking to achieve generalization and wide utility. We're back to 1981. AI and Anthropic are Apple and MS, with Meta, MS and Google struggling not to become CompuServe and AOL... Prepare for some really shady shit... then with AI/Crypto... the browser wars were only an exercise. You ready Fahd? I am... and I am rooting for the little guy. These are all first-outers. They are all IBM, Commodore and Tandy... or even Solaris ;) They will never be the new Apple acting like the present one or the new MS with the same shady licensing and force fed, all batteries included schemes and strategy. With open weights and code however I can live with the "opt out" choice, but with closed models this will be even more of a conflict (at least for me).
Thanks for the insights, I agree with some points and it's a bit too early to predict the future.
😊😊
cheers