Didnt say I liked it you tard but that the perfection gives off its fake, talking about autism and you cant even parse a fucking sentence
There's always a moral panic whenever "AI" does something it couldn't do before, even if it's be seemingly intelligent enough that instead of being able to detect it instantly, it now takes a couple minutes. Some people have always been blind to this and even in the days of dumb babblebots like Eliza, would actually react to it like it was real. These people are called retards.
671b actually. It is a sparse model, meaning that only 37b parameters of those are active at all times. So yes, on a fast serverboard with AVX512 CPU+400GB fast quad-channel DDR5 RAM (that assumes quantization, better would be 600GB+) you can probably run this model at reasonable speeds, I'd guess ~10 tokens/s? Faster than a normal person can read with streaming, at any rate. Some people theorize even fast nvme drives might be enough, I don't believe this because we're talking (very) random reads here, not sequential ones. I don't know what such hardware currently would cost but I know it's likely magnitutes cheaper than the GPU+VRAM nickel and diming.
So far it doesn’t seem particularly “politically correct” but I haven’t stress tested too hard.
Not that great, he didnt even sync the audio properly when they are in the burned car
It's fine as a little proof of concept/experiment, though I don't know why Ahriman wanted to show this since this really isn't all that spectacular besides being a decent enough editor fooling with AI video.
The software has been lobotomized to the point where all it does is just regurgitate back whatever you've just said to it. It ignores stated preferences and is just biased towards leftist political agendas above all else.
LLMs are trained mostly on positive instructions. You should tell the bot to give you paragraphed answers.
I dont use chatGPT anymore and when I did I used it mostly for work but I stopped because its definitely getting dumber, dont know why but my guess seeing other models getting ahead is that openAI is now like apple in the nineties losing their marketshare and not knowing what to do to stop it so they keep fucking it up
I wonder if it's that the Internet, the main source of most ChatGPT content anyway, is now getting excessively flooded with AI slop to the point it's degrading the quality of its own output by, basically, constantly feeding on a diet of its own shit.
It's called "Model Collapse", although I would expect that given OpenAI is a for-profit organization they would use that income to vet their data pool.
Training models with synthetic data is actually an established practice now. The Internet is really, really big, but apparently not enough.
Other AIs dont give me those bad results so my guess? its something else
I need something to fill the gaps, hate writing filler but without it its a short story
