Like, a good amount of local models are liable to invent the results of their tool call itself and inject that into the context, and then the next turn is gonna be very confused from getting two different results from running the command a single time
qwen planned and built a thing over night (technically the planning was done while I was at work but still) and it's very scuffed initializing 50K neurons takes 100+ seconds single-threaded, unoptimized code but it does, seemingly, literally work which is pretty cool for something I just told my GPU to do when I went to bed
Samu ／人◕ ‿‿ ◕人＼
dang thats a while but yea sounds like optimization could happen
S C
I'm gonna have to make it redo the whole thing and make it at least able to use the GPU for stuff cause I mean, no matter how optimized, doing stuff sequentially on the CPU is gonna be fucking terrible for any amount of Stuff worth using Even though this is just a stupid project overall anyway
moon
dang i asked a company if i could have a discount coupon for my first order with them and they said NO i told them how excited i was about trying their product too lame
moon
samu what if we had threads that simulated cerebral hemispheric communication so like the dual thread but with a corpus callosum too you could post in the left hemisphere or right hemisphere and a corpus callosal LLM would translate between the two and you could pop over to the other side to see how your posts got translated or visit the CC and see the translation process
i dunno it sounded cooler in my head but it sounds kinda lame here
>>1188310 that's kinda what existing ML does like e.g. optical character recognition it looks for a specific set of features learned through training the tricky part is dealing with bigger, noisier, more generalized inputs need a bunch of layers up front that sift out the signal from the noise
Samu ／人◕ ‿‿ ◕人＼
LLMs are great in this respect because they have so many layers that can sift through typos and euphemisms and common expressions and decipher the intent and then the reasoning can go to work
moon
connect to me?
Samu ／人◕ ‿‿ ◕人＼
connectome yes sir
Samu ／人◕ ‿‿ ◕人＼
if they can ever figure out how to train a spiking model on like video/audio feeds then we're in deep trouble
