The most prestigious law school admissions discussion board in the world.
|
 |
 
|
So can any of you NERDS explain how you "train" a local AI? You just buy expensi
Poast new message in this thread
Date: February 21st, 2026 10:07 PM Author: 180 ocher brunch tattoo Expensive ass graphics card and then what? Download some shit and it starts training AI? (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49685993) |
Date: February 21st, 2026 10:20 PM Author: Odious cruel-hearted lodge national security agency no its more complicated than that. you don't do full training yourself at this point. like pre-training + training across billions of tokens from scratch. that's something frontier labs do it takes distributed training tons of gpu hours millions of dollars. you can train your own small models. but most of what peole do today is fine-tune big models. I wouldn't do the training locally either. Those GPUs aren't optimized for that you use a cloud server gpu, you pay for the VM, deploy it and they charge you per hour when its running. Train it on the cloud then basically you extract the base model directory + weights to your local system. Depends how big and complex the system is whether your local computer can run it. I don't know as much about that part I have never ran anything directly on a local machine. I can tell you how to set up cloud servers, how set up a cli runner on it that plugs in to hugging face transformers etc, but I don't know exactly how to import it locally it would take actual experimentation for me to figure it out (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686010) |
 |
Date: February 21st, 2026 11:04 PM Author: bossy pocket flask (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686073) |
|
Date: February 22nd, 2026 1:27 AM Author: 180 ocher brunch tattoo And do what with it (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686201) |
|
Date: February 22nd, 2026 1:30 AM Author: Thirsty Mood There are a lot of reasons people need a local llm (but used in parallel with cloud AI for reasoning)
Do you think the guy who is working on the next Coke product or designing the next iPhone is chatting with Claude or using MPCs? Now imagine an enterprise org with millions of people's SSN/DOB and ACH info |
|
Date: February 22nd, 2026 1:31 AM Author: 180 ocher brunch tattoo Ok I'm imagining. Now what (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686208) |
|
Date: February 22nd, 2026 1:32 AM Author: Thirsty Mood Start using Claude and ask the same questions you're asking here (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686211) |
|
Date: February 22nd, 2026 1:33 AM Author: 180 ocher brunch tattoo No (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686213) |
|
Date: February 21st, 2026 10:20 PM Author: Umber toaster indirect expression Use Axolotl for training. Train a qlora in 8bit and let your dataset be formatted as plaintext, you also need to establish a format (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686012) |
|
Date: February 21st, 2026 10:31 PM Author: Odious cruel-hearted lodge national security agency I have never used that. Isn't it supposed to be config file based controls, no need to write a runner file w python, torch etc? Is it good? Or are there limitations? (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686022) |
|
Date: February 22nd, 2026 1:33 AM Author: Lime chapel Yeah I think that's where you get Bitcoin too (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686212) |
|
Date: February 22nd, 2026 1:34 AM Author: 180 ocher brunch tattoo You'd think AI would be DEVVING thousands of entirely new groundbreaking L1 block chains every day. But no one is doing this. They're focusing on personal assistant agents (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686215) |
|
Date: February 22nd, 2026 1:37 AM Author: Lime chapel Not flame, the ban on porn usage is a dumb artificial throttle. If you want true innovation you need to have cumming as an end goal. (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49686219) |
|
Date: February 25th, 2026 2:59 PM Author: Umber toaster indirect expression (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49694728) |
|
Date: February 25th, 2026 3:00 PM Author: bossy pocket flask (http://www.autoadmit.com/thread.php?thread_id=5837098&forum_id=2#49694732) |