Yeah sure, you can trust is with backdoors to encrypted communications but only for the good guys because they know how to take good care of such important mechanisms.

All of them really. What's typically open source is
1) the code used for training, but never the dataset for initial LLM and never the RLHF (reinforcement learning with human feedback) data used to make a text vomiting LLM into a useful question answering maching.
2) the resulting weights - these are totally uninterpretable.

So it's never fully replicable; even if you had the infra and were willing to burn electricity you don't have a way of going to 2) yourself.

AFAIK that's not just the Chinese but also open-source / weights Llama and Mistral.

This sounds dismissive, but I wouldn't read it as such.

"China doing something first, however, has never been a reliable indicator that the thing will prove durable, economic, or widely replicable. China is large enough to try almost everything."

This has always been true, you can for example read the history of improving iron / steel production during the Industrial Revolution. Either you had existing outfit and capital to try things. Or you raised capital and set-up an outfit to try things. If you had something viable, you made money, if not, the world (and hopefully you) moved on.

And in the end VCs work more-or-less on the same principle (which is why at some point someone was trying to do Uber for xyz).

This is a PR "thought leadership" BS article by Benjamin Riley, Cognitive Resonance, who "provides direct consulting support to organizations to improve understanding of how generative AI works."

This doesn't mean they're wrong but it's probably nothing terribly original (there is a reason why it's not on openreview.net as a submission into one of the relevant AI conferences).

What I don't get is this: let's say they believe that to get to "genuinely useful AI" (generating value in business, automating science) they need 1024x the compute they have now; which I think is ballpark with their spending.

But Moore's law says that number of transistors doubles every 2 years, so if we believe we can keep this going then in 4 years time they'll only need 256x the data centres and in 8 years it's only going to be 64.

So all the chips they put in will be next to worthless in 8 years and they'll only need a fraction of the other infrastructure. So all this massive spend is to be first because? Because whoever is first will be able to use their "genuinely useful AI" to out-innovate everyone to singularity?

The whole thing is a one-way bet on the empirical LLM scaling laws (https://arxiv.org/abs/2001.08361) scaling to something useful, which I don't buy.

This is great but there is still a long way to go. Renewable capacity is not really comparable to fossil fuel power station capacity because the coal / gas ones can run 24/7...

To get a better picture of where we are check out http://grid.iamkate.com/ . Basically in the last year UK electricity was 19% from renewable sources with fossil fuels at 48%.

Have you looked into https://www.winlink.org/conten...

... and the ArcaBoard is really real.

Are we allowed to say that out loud?

I've installed it to edit my CV. Most features are there, and I could open and edit a new blank doc. You need a Microsoft account to sign on, but an Office 365 subscription isn't needed. Who going to be worried about tracked changes whilst using a tablet?

It is largely down to terrible marketing and poor public relations. We all need to have someone to hate.

Tots agree that this could work for corporations. Especially if a certain Canadian firm gets a licence.

http://soylentnews.org/article.pl?sid=15/03/01/1949210 Are you sure?

Looks like they've got it covered. https://www.openhub.net/p/crouton/commits/367012555