Random #AI thought of the day:
The amount of computing power made available to the average user nowadays to run inference for these #LLMs probably far exceeds the amount of compute you would need to (for example) render a Blender video, or run post-production on 4K-quality video, which would probably be more useful to most average users than what LLMs provide — and they provide all that computing power for free to literally hundreds of millions of users.
I mean, come on, you stupid despot-kings of Silicon Valley, if you are going to provide that much free computing power to just anyone, can you at least let me do like Adobe Premier or After Affects, i.e. post-production of my home videos, directly on Google Drive without extorting me for monthly subscription fees?
On the one hand, it’s incredible that so much computing power is available to ordinary people now, on the other hand, it is even more incredible how much those computers (and the energy needed to power them) are wasted on something so frivolous as drawing a Pikachu with big boobs, or as dangerous talking with a companion that will always agree with you even if you suggest committing suicide.
But I guess that is why this tech financial bubble must eventually explode. No one can reasonably provide that much computing power for something so useless for free for very long, it is just not economically feasible.
#tech #ai #AIbubble #LLM #SiliconValley
reshared this
your auntifa liza 🇵🇷 🦛 🦦 reshared this.
wasabi brain
in reply to Ramin Honary • • •Ramin Honary
in reply to wasabi brain • • •@virtualinanity I think some of the GPUs can be re-sold for video rendering or gaming, but they aren’t purpose-built for that. Some of the new chips NVidia are designing are really only useful for neural network inference, so if too many of those get built I can’t see any way to recover the costs of that investment.
HoldMyType
in reply to Ramin Honary • • •Ramin Honary
in reply to HoldMyType • • •@xameer for I while I mostly did image recognition for quality control systems. I recently started learning more about how to do Retrieval-Augmented Generation (RAG) using small LLMs like the 8-billion parameter LLaMa. I know enough to generate language embedding for your corpus and do a similarity search.
But let’s not conflate actually useful AI with what Silicon Valley is peddling to all the big governments of the world right now. They seem to have this strong religious conviction that scaling these things up indefinitely will create a general intelligence that will then explain to them how to create a smaller, faster, more efficient AGI and they won’t have to worry about energy costs any more after that — that is, at least what guys like Sam Altman and Peter Theil seem to be trying to convince everyone will happen.
HoldMyType
in reply to Ramin Honary • • •was it something like pict-rs * and ollama?
Ramin Honary
in reply to HoldMyType • • •@xameer I was not the data scientist on the project so I don’t recall the exact algorithms that were used, but there were several used for image recognition, there were a variety of CNNs we used, they were all programmed as scripts for PyTorch with whatever hyperparameters the data scientists thought were best for the application. Some of these algorithms were as old as AlexNet or ResNet, but these still work well so we used them. Some of them weren’t even deep neural networks, they were just ordinary convolutions using OpenCV for box drawing, I recall using OpenCV for one application I made for cleaning-up data sets and doing data augmentation.
For RAG, the only ones with which I experimented were LLaMa and DeepSeek, and yes, I used both through the Ollama library.
plan-A
in reply to Ramin Honary • •Ramin Honary likes this.
Ramin Honary
in reply to plan-A • • •plan-A likes this.
plan-A
in reply to Ramin Honary • — (Proud Eskimo!) •Also I use just my pc, I thank you for your kind formulated answer friend ; )
Ramin Honary likes this.
dogzilla
in reply to Ramin Honary • • •