They're talking about the AI.

I'm not sure if this is what the commenter meant, but I am vanishingly unlikely to own the AI. Even if I write the AI I'm unlikely to own it. The training costs are too large and even if I did train a working model, AI can be duplicated, and there's little reason to use anything but the best. The scary thing about AI to me is finally turning intellectual labor into a pure process of capital. You put more energy and capital in, you get more out, no need or room for humans anywhere in that loop. Now of course we're a long way away from that. There will be room for humans for a long time, but it's scary how much additional power it will give to capital. How much less the interests of ordinary people will mater

That's actually a great thing in the long run. Money will be spent on compute, compute will generate data by generative models + validation models, then data will be compacted into a new model. Models and data can be copied, money can't.

If it works with just electricity and doesn't require manual human work it is a game changer. No longer limited by human resources, we can scale research in any field and improve everyone's life much faster.

AlphaGo is an example of such an approach. I don't think models will be locked down, they will probably be like go bots in recent years, about 50% of them open sourced. As long as there is an open dataset, the models can be replicated.

People start out with a brain, they don't tend to start out with money, electricity and in the absence of and incentive it's hard to imagine society doing much for the downtrodden. I think empathy exists and it can be strong, but we've had racism, bigotry, and genocide all while the people we were killing had every potential to be another Einstein.

Also why would you think the weights for models will be open source. That isn't even true now for the most complicated models (gpt3 and the like) and certainly isn't for all the models running right now in production using proprietary datasets. As long as keeping it secret makes it more likely that you (or your models) can generate the next better model it doesn't seem very likely we're converge on them all being public

> Now of course we're a long way away from that.

I am no longer confident of that.

The far future where we can ask an AI something like make me money, or improve humanity seems pretty far off still. Not for sure. Change can accelerate, but probably under 2% over my lifetime (which is still way higher than I would have put it a few years ago). Until we get there the question will be how humans can better leverage the tools. We won't be in quite as dire problems because the limiting factor will probably still be humans minds to do the increasingly small set of things the AI is bad at

This confuses AI training with inference. Training a machine learning model is expensive, complex, and requires data. Once trained, inference is cheap in terms of compute. A trained model is trivial to leak so hopefully nobody is putting much stock in their models staying secret.

Not to mention there are techniques by which training can be avoided entirely like transfer learning and others.

Inference is not cheap -- SOTA LLMs require 100s of gigabytes of VRAM for inference.

Apple has made a sample project available for Stable Diffusion that can run on both macOS and iOS.[1] The sample code is here.[2]

I'll concede that Stable Diffusion may or may not meet the threshold for SOTA but still think this is indicative of inference eventually becoming supported for any compelling LLM on consumer-grade hardware. The possibilities for creative tools are just too compelling.

[1] https://machinelearning.apple.com/research/stable-diffusion-...

[2] https://github.com/apple/ml-stable-diffusion

I know, I was just using their metaphor.