There’s a good amount of research going into reducing the compute needed for training and inference, as well as a ton of R&D going into making far more energy efficient hardware for training and inference
Just like how 3D rendering has gone from dedicated $40,000 workstations and render farms to something that’s just done for funsies on your phone, the capabilities of these really powerful models will eventually be squished onto the cheapest, lowest power mass market computers of the day
The biggest long term challenge will be the training data and licensing of outputs. If AI outputs are stuck in a legal state where you simply can’t use them commercially, the whole industry will collapse and return to the most ignored corners of university computer science programs. If models aren’t required to get licensing for all training data we’ll probably just keep seeing companies hoovering up data in the most unethical possible ways to train their big models
There’s a good amount of research going into reducing the compute needed for training and inference, as well as a ton of R&D going into making far more energy efficient hardware for training and inference
Just like how 3D rendering has gone from dedicated $40,000 workstations and render farms to something that’s just done for funsies on your phone, the capabilities of these really powerful models will eventually be squished onto the cheapest, lowest power mass market computers of the day
The biggest long term challenge will be the training data and licensing of outputs. If AI outputs are stuck in a legal state where you simply can’t use them commercially, the whole industry will collapse and return to the most ignored corners of university computer science programs. If models aren’t required to get licensing for all training data we’ll probably just keep seeing companies hoovering up data in the most unethical possible ways to train their big models