So this was close on the 50%, although not that close on the 75%, but I checked and I did not hit either criteria in time, and I am definitely not going to be working on this over the weekend, so I want to ensure everyone has their mana back.

I do have a draft post with a lot of details, and I eventually will write up something, but other things kept seeming higher priority, and at this point there's no speed premium.

At this point I’m counting on a benevolent lord

@Volty OK but I am pretty sure Zvi is currently in the process of reading the whole damn thing...

I read/skimmed the trilogue mandate texts earlier this year. It's probably worth reading/skimming, but there's big chucks of it that you can skip over without even skimming.

I find it unlikely you'll go through the bulk of the thing for the simple reason that most of it isn't very interesting nor necessary to read.

If you look at the law (https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:52021PC0206), it's about 50K words. However most of that is in a lengthy preamble defining terms, background, stuff about the stakeholders, budgetary implications, etc.

If you only count the words of the text which has an actual legal effect, it's at about 20K. Even there, a bunch of it is eminently skippable stuff like how the supervisory board is governed.

@JensGoldberg

eminently skippable stuff like how the supervisory board is governed.

Based on recent events in the world of AI, I'd argue that the procedural details of exactly how a board is governed can be rather important.

@Ziddletwix also "defining terms" definitely has legal effect

This is maybe too outside-view, but in general I'm skeptical that reading the entire text of laws is ever the right way to understand them better, unless you're in a very specific set of people.

• Laws often contain legalese or parliament-ese that needs translation to layman context to properly understand the scope, the likelihood of it being enforced, etc

• Human beings don't really retain information much reading through texts like laws

There are people who have the context and will write summaries, finding ones you trust and relying on their take is usually a better approach.

@DanMan314

There are people who have the context and will write summaries, finding ones you trust and relying on their take is usually a better approach.

I think this is a compelling case for an actual layperson, who just wants to understand things for themselves, but Zvi chooses to spend his time writing 10K+ words/week about the development of AI, for others to read. I don't doubt that Zvi lacks some context that might help him understand all the complexities in a dense regulatory doc. But those experts would lack the context that Zvi himself brings (no one has all of the relevant context). Their summary is not just "the doc, but condensed", it reflects what they were looking for in the doc. Given how much Zvi writes about this topic, I have to assume he feels he brings some perspective that isn't sufficiently covered by other journalists/experts.

To me, the fundamental question here is simply "do the details of regulation matter, or is it just about the spirit/intent". If the latter, then a summary is probably sufficient. But if you think that the impact of a regulation can come from small details in its implementation (e.g. unintended consequences), I don't think there's much of a shortcut for reading the long boring bill. When AI regulation is crafted in the future, aides will not invent it from first principles, they will look to what already exists and adapt it.

@Ziddletwix Detailed summaries exist! I think this is the key point of disagreement:

To me, the fundamental question here is simply "do the details of regulation matter, or is it just about the spirit/intent".

My claim is that the "details" of regulation don't exist in their text. The recent spate of tech regulation out of the EU proves this - reading the body of the law would not have given you the ability to predict how they would be implemented by tech companies, then that implementation interpreted by courts, which parts would be challenged and enforced vs functionally ignored. I'm also not claiming the summaries make you omniscient on this regard, but they'll be much better than just the raw text.

@DanMan314 I really, really disagree here; they’re complementary goods. I have literally lost track of how many times a summary of a law, confidently cited by a government employee or similar, is in fact not accurate. Zvi having the ability to call BS on summaries that will be quoted for many years at him Is disproportionately valuable

Reading the whole thing is the only way to get a sense of the ratio of pro-acceleration momentum vs. pro-safety openness.

I think you will become quite pessimistic about the overall situation, the EU AI act will probably clearly display how badly countries already want AI growth (unlike the speeches which dogwhistled to AI safety, fooling everyone, while openly supporting acceleration).

Discussion of AI regulation has descended into full tribal epistemology, with both sides, but especially the e/acc side, making wild claims about the content of documents which are contradicted by the text.

Obviously, government regulation will shape the development of AI. Even if the spirit/flavor of the regulations feel obvious, unfortunately, often it's the boring details buried in the document that have the most impact (i.e. in other regulatory areas, would you agree that "the details matter"?). There's no shortcut for reading the doc—you can rely on the summaries of others, but they do not share your objectives or perspective.

If there's skepticism that this particular AI act will matter, I'd argue that it's very likely to serve as a blueprint/inspiration for any future regulations drafted in the US (a super common pattern—we often don't "lead the charge" w.r.t. regulation, but when we follow a year later congressional aides invariably start from the blueprint of what's already out there, they don't come up with this stuff from scratch).