Comment on Lemmy be like

<- View Parent
FauxLiving@lemmy.world ⁨3⁩ ⁨days⁩ ago

I firmly believe we won’t get most of the interesting, “good” AI until after this current AI bubble bursts and goes down in flames.

I can’t imagine that you read much about AI outside of web sources or news media then. The exciting uses of AI is not LLMs and diffusion models, though that is all the public talks about when they talk about ‘AI’.

For example, we have been trying to find a way to predict protein folding for decades. Using machine learning, a team was able to train a model (en.wikipedia.org/wiki/AlphaFold) to predict the structure of proteins with high accuracy. Other scientists have used similar techniques to train a diffusion model that will generate a string of amino acids which will fold into a structure with the specified properties (like how image description prompts are used in an image generator).

This is particularly important because, thanks to mRNA technology, we can write arbitrary sequences of mRNA which will co-opt our cells to produce said protein.


Robotics is undergoing similar revolutionary changes. Here is a state of the art robot made by Boston Dynamics using a human programmed feedback control loop: www.youtube.com/watch?v=cNZPRsrwumQ

Here is a Boston Dynamics robot “using reinforcement learning with references from human motion capture and animation.”: www.youtube.com/watch?v=I44_zbEwz_w


Object detection, image processing, logistics, speech recognition, etc. These are all things that required tens of thousands of hours of science and engineering time to develop the software for, and the software wasn’t great. Now, freshman at college can train a computer vision network that outperforms these tools using free tools and a graphics card which will outperform the human-created software.

AI isn’t LLMs and image generators, those may as well be toys. I’m sure eventually LLMs and image generation will be good, but the only reason it seems amazing is because it is a novel capability that computers have not had before. But the actual impact on the real world will be minimal outside of specific fields.

source
Sort:hotnewtop