Date: February 12th, 2024 10:23 AM Author: Stirring brass office
naw SalesGPT with hot sexy female or chad voices/Zoom avatars will be way better at manipulating simps/cat ladies and getting them to hand over credit card than even the best salescuck
Date: February 12th, 2024 11:37 AM Author: doobsian electric becky
They can’t really stop it simply from a national competition perspective. UBI or something like it is necessary.highly capable AI will proliferate no matter what
Date: February 12th, 2024 10:37 AM Author: drunken tattoo
that AI commercial last night was depressing. They set it up with all these plucky "THEY said I could never x y z" but instead of an inspirational "I showed them" moment, you just get people talking into their phone saying "AI do x y z for me"
Date: February 12th, 2024 11:47 AM Author: doobsian electric becky
I don’t think people have even accepted the likely job losses from current tech. Customer service jobs are clearly going to be gutted. Coding and other white collar jobs can see significant productivity increases from LLMs. It’s clear to me when I play around with GPT vision that these models are already very versatile when understanding visual scenes. We don’t have widespread humanoid robots because there was no way to use them. They were too inflexible outside of factory work. That’s basically over with. People are telling themselves a story that none of this matters because self-driving didn’t materialize to the extent predicted, but that’s not very convincing because models need substantial common sense and versatility to handle all the edge cases. That seems to be changing very quickly and it hasn’t yet been integrated with driving capabilities yet.
Date: February 12th, 2024 10:00 PM Author: Stirring brass office
interesting analysis but what are the computational resources needed to have chatgpt-like analysis of visual scenes being done on a *local* basis and done on gigabytes of visual/LIDAR data near-instantaneously? you can't be relying on the cloud for self driving due to network unreliability and speed issues, so you'd need the computational capacity available in the car. and it needs to be happening within milliseconds due to safety issues. right now chatgpt takes forever to analyze a simple static image. I know I probably sound like a guy saying this internet thing will never displace faxes, but it seems like the time horizon for widescale implementation of this tech in a self-driving application could be decades out just due to technical limitations
Date: February 12th, 2024 10:31 PM Author: doobsian electric becky
I think it’s plausible a single GPU could do it, especially with increased memory compared to state of the art cards. A single forward pass of a large model doesn’t take long assuming you can fit all the weights on the card. Running a GPT-4 sized model locally isn’t realistic, but 1) model size keeps being driven down. There are now GPT 3.5 equivalent models that are rather small. This hasn’t hit a limit yet. 2) GPT vision has lots of weight data that is unnecessary for driving a car. Spatial information in particular is highly redundant and compressible compared to language, which is why image diffusion models can be so small and retain so much information about image structure. Language models have to be larger because language is already a highly compressed representation of the world. I think it’s very likely that models with robust spatial and motor processing could be small and run locally.