Hacker News Reader

Comment by iambateman | Hacker News Reader

If the author sees this…could you go one step further, what policy specifically do you recommend?

It seems like having LLM providers not train on user data is a big part of it. But is using traditional ML models to do keyword analysis considered “AI” or “surveillance”?

The author…and this community in general…are much more prepared to make full recommendations about what AI surveillance policy should be. We should be super clear to try to enact good regulation without killing innovation in the process.

View on HN

Thanks (author here). I am working on a follow-up post (and likely posts) with specific recommendations.

While I agree with your take on the harms or AI surveillance, I will never agree that AI is beneficial, and there is a net negative outcome using AI. For example electricity prices, carbon release, hallucinations, cognitive decay...they all outweigh what benefit AI brings, which still is not clear.

Like nuclear fission, AI should never have been developed.

As well as crypto.

LLM providers should only be allowed to train on data in public domain or their models and outputs should interior the license of the training data.

It's ironic copyright is the law that protects against this kind of abuse. And this is of course why big "AI" companies are trying to weaken it by arguing models training is not derivative work.

Or by claiming that writing a prompt in 2 minutes is enough creative work to own copyright of the output despite the model being based on 10^12 hours of human work, give or take a few orders of magnitude.

Makes sense, have to deal with the cat being out of the bag though.

The groups that didn't train on public domain content would have an advantage if it's implemented as a rule moving forward at least for some time.

New models following this could create a gap.

I'm sure competition as has been seen from open-source models will be able to

It's simple, the current models and their usage is copyright infringement.

Just because everyone is doing it doesn't meant it's right or legal. Only that a lot of very rich companies deserve to get punished and pay the creators.

I was referring to the issue of the new models having to train different than the original ones.

Not arguing, debating about the legality of what the models have done.

Anthropic just paid a settlement. But they also bought a ton of book and scanned them, which might be more than other models. Maybe it's a sign of things to come.

$3000 per book is laughable given they can then use the book as much as they want and if they build a good enough model, it'll completely obviate the need for the book.

Copyright designed at a time when reproducing work in way which was not verbatim and not obviously modified to avoid detection (like synonym replacement) would require a lot of human work and be too costly to be done. Now it's automated. It fundamentally changes everything.

Human work is what's to be rewarded, according to the amount of quality.

From the TFA:

> That’s why we (at DuckDuckGo) started offering Duck.ai for protected chatbot conversations and optional, anonymous AI-assisted answers in our private search engine. In doing so, we’re demonstrating that privacy-respecting AI services are feasible.

I don't know if its a great idea, or just I wonder what does make it feasible, but there is a kind of implied recommendation here.

By "killing innovation" do you just mean: "we need to allow these companies to make money in possibly a predatory way, so they have the money to do... something else"? Or what is the precise concern here? What facet needs to be innovated upon?

The country who most effectively deploys AI models has a big advantage over countries who bury their head in the sand.

I believe that LLM’s will have the capability to fill in for human workers in many important ways. It’s like getting an economic infusion without the associated population growth required.

But we aren’t there yet, so innovation looks like continuing to build out how to efficiently use AI tools. Not necessarily better models, but longer memory, more critical reasoning, etc.

At the same time…there are winner-take-all dynamics and possibility to weaponize that are not good for society in the long-term.

So we need to both encourage innovation while making sure we don’t kill each other in the process.

the law could be as simple as requiring to blur faces and body silhouettes of all people inside each camera, prior to any further processing in the cloud, ensuring privacy of the CCTV footage.

Reply on news.ycombinator.com