[rvz]

> AI companies will continue to generate waste and CO2 emissions at a huge scale as they aggressively scrape all internet content they can find, externalizing costs onto the world’s digital infrastructure, and feed their hoard into GPU farms to generate their models. They might keep humans in the loop to help with tagging content, seeking out the cheapest markets with the weakest labor laws to build human sweatshops to feed the AI data monster.

Again, I've said this same thing months and yet the AI bros continue to deflect with more nonsense to justify burning the planet with their snake-oil garbage.

Drews points still stand and the Deep Learning industry has no methods of efficient methods of training, fine-tuning and inference and continues to burn down the planet no matter the amount of greenwashing they continue to project.

[danielbln]

What a load of snide hogwash. First of all, how are LLMs snake oil? There is actual usefulness to be had.

And I can run inference on my laptop, transferring capabilities to smaller models is a thing, quantization is a thing, optimization is a thing. And, DL has been feasible for barely a decade, LLMs for a few years.

Are you talking about crypto by any chance?

[rvz]

> What a load of snide hogwash. First of all, how are LLMs snake oil? There is actual usefulness to be had.

It is an energy wasting snake-oil burning the planet. [0] [1] Especially when they confidently hallucinate without any transparent reasoning or explanation other than the regurgitation that it has been trained to do, or more accurately they are stochastic parrots.

The issue is fundamental to LLMs and deep learning and researchers still don't know why other than tweaking parameters and fine tuning / re-training it with GPUs still incinerating the planet with no viable alternative to such wasteful methods.

> And I can run inference on my laptop, transferring capabilities to smaller models is a thing, quantization is a thing, optimization is a thing.

We are talking worst case for inference not 'smaller models' which still need to be trained or fine-tuned to exist in the first place and for improvements. For the so-called 'serious' cloud-based LLMs, they need to continuously serve every inference and that requires a fleet of GPUs to serve lots of users as the parameter count of the model gets larger.

> And, DL has been feasible for barely a decade, LLMs for a few years.

Neural networks which are fundamental to LLMs have been around for decades and are still unexplainable black boxes which are incapable of transparent reasoning other than regurgitating responses that it was trained on. Unacceptable and useless for a wave of use-cases that require explainability.

> Are you talking about crypto by any chance?

Crypto already has viable alternatives to its energy wasting problem [2] available today right now. Deep Learning still does not.

[0] https://gizmodo.com/chatgpt-ai-water-185000-gallons-training...

[1] https://www.independent.co.uk/tech/chatgpt-data-centre-water...

[2] https://consensys.net/blog/press-release/ethereum-blockchain...