Hacker News Reader

Comment by laserbeam | Hacker News Reader

I know the full text is not linked in this article, yet, the kind of regulation mentioned here sounds like the kind of regulation I'd love to see from AI companies.

- being required to disclose energy use

- not being allowed to scrape CCTV cameras for training

- not being allowed to use AI for things like social scoring

The things mentioned in this post feel like restrictions on the possible ways people could abuse others with AI. Even though I may not agree with all the items on the list (there are more than 3 hidden in the body of the article), the direction feels in service of me, as a citizen. That's great!

Contrast that to US regulation posted a few months ago which basically said: you have to be one of these 5 companies to develop AI because we trust these 5 companies.

View on HN

Why is energy use of these models in particular of interest to you, vs everything else a tech company does?

How do you even define it? If you start with a pre trained model, must you include the energy that went into the initial training, even though that didn’t occur at your company?

Energy use for AI is interesting due to exponential growth. Most energy spent by these companies falls under existing categories like office HVAC which isn’t going to change much globally.

However AI as an emerging technology could quickly go from irrelevance to significant in a fairly short period and some lead time before the actual impact is useful.

Would knowing the energy use of an ai product influence you to use it less than if you didn’t know? Or is it simply a curiosity?

If the latter, this regulation seems useless and performative.

My point is not about users but the electric grid. You need to start planning capacity long before the actual demand shows up.

Energy use is obviously related to carbon emissions, which as you may have heard we would like to reduce.

It's related, but consider that tech companies are close to the forefront of sourcing renewable energy compared to other industries, and AI has the potential to supplant higher-emissions activity, so it may be a false economy if regulation has the effect of constraining AI development in the name of saving energy.

Also - should the energy used to create the training data be taken into account?

Presumably if a drone is flown over a terrain to collect data that activity should count; but what about the energy used to create make up for models etc?

> - not being allowed to use AI for things like social scoring

How bout not allowing social scoring at all?

2 points by normalaccess 59 minutes ago | prev | next | edit | delete [–]

I have a few questions about this sort of legislation.

How do you determine what "AI" even is in this context? I feel like we have a Loki's Wager situation on our hands where either the context is so restricted to be functionally meaningless or so broad that simple things like basic math processing would be covered.

Although it sounds crazy I fell like certain kinds of math (like matrix multiplication, gradient decent, and simulated annealing) will soon be outlawed unless you are authorized.

As far as the other stuff I'm not sure how effective it would be... Like CCTV data, you could just simulate virtually identical visual data using tings like the unreal engine. But I do get it. Things like social scoring? That cat is already out of the bag and using machine learning and AI is only a value add to the system.

At the end of the day AI is just /fancy/ math and any problems that fall into the "P VS NP" category use the same computational techniques.

https://en.wikipedia.org/wiki/Loki's_Wager https://en.wikipedia.org/wiki/Gradient_descent https://en.wikipedia.org/wiki/Matrix_multiplication https://en.wikipedia.org/wiki/Simulated_annealing https://en.wikipedia.org/wiki/P_versus_NP_problem https://www.wired.co.uk/article/china-social-credit-system-e...

EDIT: Spelling

> I have a few questions about this sort of legislation. How do you determine what "AI" even is in this context?

Instead of linking 4 wikipedia articles you... could find the text of the legislation: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A...

> Although it sounds crazy I fell like certain kinds of math (like matrix multiplication, gradient decent, and simulated annealing) will soon be outlawed unless you are authorized.

Yes, you sound crazy. That is, you blindly parrot the industry that started spreading FUD about this before the AI Act was little more than a twinkling in its authors' eyes. This legislation doesn't outlaw a single thing.

Here is another link: The truth about the EU AI Act and foundation models, or why you should not rely on ChatGPT summaries for important texts https://softwarecrisis.dev/letters/the-truth-about-the-eu-ac...

> Like CCTV data, you could just simulate virtually identical visual data using tings like the unreal engine. But I do get it. Things like social scoring? That cat is already out of the bag and using machine learning and AI is only a value add to the system.

"You could just", "it's only", "bad things are already happening, why should we bother with laws and regulations, the cat is already out of the bag"

Reply on news.ycombinator.com