@davidgerard Hard to imagine someone is going to keep a service online if it's not profitable.

@davidgerard To quote @piater:

> If it works, we don't call it AI

@davidgerard The 'decentralized tech' field of research is still recovering today, a decade later, from the immense damage and funding vacuum caused by the blockchain hype - so I wouldn't expect the ML field to fare any better... though I guess at least the ML field has a longer history of such disasters, what with the "AI winters"

@davidgerard

Part of the problem is that AI is such a meaningless marketing term that you can always find something under that umbrella that is genuinely useful and will stick around. Mozilla’s little translation models are amazing. They run on my phone, are about 60 MiB for each language pair, and can be trained on a single desktop from ethically sourced training data. The same with a bunch of the computer vision models.

And, as with other ML systems, they work really well in setting where the value of a correct answer is significantly higher than the cost of a wrong answer. I wouldn’t use the Mozilla models for translating a contract or a technical reference. I wouldn’t use them for a menu if I had a severe allergic reaction, but I would use them for translating random things as a tourist or getting a rough idea of what a web page says. I wouldn’t use the OCR models and discard the original image, but using them to add a searchable layer to a PDF is great.

@davidgerard I like meeting people who say they find AI useful, because then I can ask them how much they'd personally pay for access to it. The highest I've found so far is £10 per month. The most common answer is 'oh no, I wouldn't *pay* for it'.

@davidgerard Plus all the electronic waste that will be produced by the giant server farms. They will go through hardware fast. The whole thing is bogus.

@davidgerard If anything about AI is here to stay, it’s that it has outed people who don’t give a shit about other people’s intellectual property or rights, who never wanted to treat other people with respect anyway, who don’t give a shit about truthful representation in media, are nazis and a load of other people either incapable or unwilling of realising what in general makes humans great.

For that I am thankful. 🙏

@davidgerard

A necessary but not sufficient condition for LLMs to be useful is for truth/accuracy not to matter too much in a given context.

This means one of two things: either checking truth/accuracy is easy (some simple code like unit tests) or it doesn't matter at all (spam emails)

For the first scenario, LLMs are kind of almost useful if you have them already set up and someone else is paying the bill.
The second case can be tested into the sun and we'll all be happier.

@davidgerard is there a valid argument to be made about any of the hardware built to run these things? Or was that all just scaling problems and not actually anything new? Or not transferrable to other problems?

@davidgerard you make a good point but the other side of it is this: AI today is just the beginning think of it as a thought train that can validate thoughts it has yet to prove it can pontificate or actualize at a groundbreaking level that can replicate human behavior however I have seen hacker threads showing how injecting prompts of commands for powershell and cmd show it can provide new security threats and security advancements that will be like an immune system to the body

@davidgerard They are actually pretty useful to actively harm the earth, avoid copyrights, and waste more of your customers' time until they realise they were actually talking with a slop generator machine.

@davidgerard Personally I think we should make a coordinated effort to erase the tech, as the primary industry that benefits from it is the scam industry.

@davidgerard “are we the baggies?” ;-)