@deepfates Just build a server at this point, you'd be surprised how even a 27/32B swings when you can personally strip the RLHF, set the sampling (off, always off, do_sample=false) and sysprompt, train your own LoRa and write tool scripts. There is so much overhead in flagship API
@Killdeerrr Could always put out a Facebook marketplace ad ..., ok I meant that to be funny but it might actually work.
Or you could end up in more than one freezer
Or it could work
@foolibuster There is noise
There is signal
I am a flawed creature
I do not claim to be remotely adequate in any metric
I cannot process both
I cannot engage with the former on any level
@VeryGreatMathew To be fair, the tech behind personal multi-medium transport was well mature enough to support real applications.
Licensing and regulatory enforcement at that scale was just gonna be bonkers and the FAA was like oh hellllll no 🙂↔️
Wat.
So he had friends who worked mains utilities and always had him back up their pics of trenches, gear, installs. So he just... takes his massive niche hi-res imagepool and trains SDiff in his barn as a "hobby", on WINDOWS.
My guy, you are ahead of 99% of the AI talent pool.
Insisted on desolate chainsaw massacre barn (not his house)
Met by agriculture flesh Mater that shook hands like a bear.
Asked if he just resold (ewaste is a big side gig if you got land) "naw I was trainin custom stable diffusion datasets with millions of util pics from my buds"
About to pick up a small AI/ML datacenter server from a guy in the middle of a literal cornfield about a 20 miles from the nearest town.
Huh, when I say it out loud like that it sounds like a terrible idea.
About to pick up a small AI/ML datacenter server from a guy in the middle of a literal cornfield about a 20 miles from the nearest town.
Huh, when I say it out loud like that it sounds like a terrible idea.
@CRXSSF4DE Yee and I hope they stay that way after I install them with ... 8 inch pounds of torque on the heatsink hardware you gotta do like a cylinder head or i guess it can crush the micro pins or just straight up crack the die which is not at all uncommon 🙃
The big models must be adopting Multi-Token prediction (I like to call it Reality³) because I'm seeing some specific growing pains that must make it wonderful for normie one-offs but feel like training wheels made of butter for long format and informed interactions.