Google refuses all of that. Basically all of the tech companies do.

No way Google would do that. The potential liability for making mistakes would be enough to make them run screaming all by itself, even assuming there's no risk if the information is accurate -- which is at least not clear and probably not true for many of the things you listed.

Google is leery of even having such information, and definitely isn't going to be giving it out to anyone who doesn't show up with a subpoena or a warrant.

Slippery slope fallacy. Given that this is strictly opt-in, for children only, under parental control, there's absolutely no reason to think that any of the rest of that is soon to follow.

I find it interesting that their intention is clearly based on concerns about IP ownership, unrelated to quality. They're not banning "AI slop", they're banning code with unclear ownership. They clearly want to avoid a situation where they take a code contribution which the AI model owner might claim to own, or which owners of the AI's training material might claim to own. If those claims were upheld, it would mean the contributor didn't own the copyright and therefore couldn't license it under the relevant open source license.

I think this will ultimately prove to be needless caution. I think courts will end up deciding that AI-generated code is no different than code generated by any other sort of code generation tool, and that the copyright is held by the person using and directing the tool. But it's definitely not impossible that things will go a different way.

I disagree with your conclusion. We don't need "more", we need "better"... and a key aspect of the "better" we need is that we need more focus on identifying and addressing regulatory capture.

Anti-trust law is largely a bust. You can't point to any case in American history, for example (I'm guessing you're not American, but I reach for the history I know; if you have good counterexamples from elsewhere I'd love to hear about them), where there was an abusive monopoly and the government stepped in and fixed it. What happened in every single case is that market forces were already taking down the monopoly by the time the government took action. I don't think anti-trust law is bad, per se, I think it defines some useful boundaries and probably warns a lot of companies away from abusive behavior just by defining clearly what is abusive. But I think the effect is ultimately just to decrease the amount of abusive behavior that happens before the market takes the monopoly down, which would happen even without the laws.

The real exception, the case where you get long-running, entrenched monopolies that the market doesn't eventually undermine, is when government regulation creates or maintains the monopoly. And the more regulation you have, the easier it is for industry to get that regulation tweaked in its favor. We need regular and intense scrutiny of possible cases of regulatory capture, and then we need to refine the regulatory regime to prevent capture -- the best way to do that in most cases is simply not to give regulators the type of power that businesses can exploit. Note that I said "type", as in "kind", not as in "degree". I'm not arguing for weak regulations, I'm arguing for regulations that primarily focus on shining a bright light on corporate operations, rather than regulations that control or restrict.

The primary role of regulation in business operations should be to ensure that companies operate with honest and full disclosure, of their financials, of their labor practices, of their product quality and characteristics, etc. Note that it's also important to ensure that the regulatory burden of information disclosure doesn't become another form of capture, preventing the entry of small players who don't have the necessary expertise to fill out all of the necessary forms.

Nationalization is its own can of worms, unrelated to regulation. It's generally a bad idea, though. In theory it seems like a good idea to eliminate the profit-taking that private enterprise necessarily engages in, but the savings that provides are basically always swamped by inefficiency created by the reduced need to be competitive.

Obviously, there are exceptions; industries with natural monopolies, where control-oriented regulation or even nationalization make sense. Most of these are near-universally government-run already, such as road networks and air traffic control. Utilities, however, aren't really natural monopolies, even though they're generally treated as such. The last-mile delivery of power, water, etc., are natural monopolies (we don't want five water companies digging up our streets to install their own pipes), but the industries as a whole aren't, and shouldn't be treated that way.

Above all, I think market regulation should focus on enhancing transparency (which is what prevents bad behavior) and removing barriers to competition. Markets are the most powerful mechanism we've found for reducing poverty and increasing wealth. Harness them, don't hobble them.

Only in certain sorts of games. Your point is that this particular case is one of those games, which makes sense to me. My point was more about the higher-level question of whether AI acts this way because it emulates humans or because this sort of behavior is inherent to cooperating/competing entities trying to maximize advantage; I lean toward the latter explanation though it's obviously hard to untangle.

The fact that these LLMs aren't smart enough to understand that such collusion and backstabbing is ultimately counterproductive doesn't affect my point. In fact, humans also don't generally understand that such collusion and backstabbing is counterproductive; the research results are subtle and non-obvious. The emergent effect we see in large-scale, long-running enterprises is that they settle into patterns of fairly forthright, honest behavior because that's what keeps the cash flowing decade after decade, but these kinds of shenanigans are common among humans who don't have that long-run institutional experience.

Increasing a country's human capital, its supply of skilled labor, is never a bad thing. It increases the country's capacity and lowers its costs. This isn't always great for people already in the profession in question (but it's not always bad for them, either).

Yeah, the wage spike will end and wages for those professions will go back to normal, maybe even dip somewhat, but supply induces demand. If the skilled labor portion of construction costs less, either in wage reductions or in availability increases, that motivates more construction -- and if we can get our collective heads out and fix the various regulatory obstacles blocking housing construction around the country, and especially in the cities, making use of that new supply of skilled labor for a housing construction boom would be great for everyone.

This is article is just someone looking to find clouds amongst the silver linings, and invent them when they're hard to find.

Alternatively, this is just the normal emergent behavior you can and should expect any time you have groups of entities competing and cooperating.

We didn't evolve to work the way we do because something made us evil, we evolved to work the way we do because it was effective. And that includes both the cooperation and the backstabbing. Being always and thoroughly evil to everyone quickly gets you excluded from everything. Being a perfectly-honest naif means you quickly get taken advantage of by someone slightly less-honest. So the most effective strategy is a well-chosen mixture of cooperation and betrayal, while trying to convince society that honesty and cooperation are the only correct values.

Heh. I still occasionally find people using Y2K as an example of "ludicrous sky-is-falling" claims that turn out to be a big nothingburger. That and the hole in the ozone layer and the dangers of acid rain.

What all of those things really are is examples demonstrating that when we decide we care about large-scale problems enough to invest in solving them, we actually can. I'm really hoping that climate change turns out to be another one that we solve before it becomes really bad (it's a little bad now, and definitely going to get somewhat worse, but not really bad), but it's not looking great.

To be precise, it turned Senate elections into a different kind of popularity contest. Previously they were a popularity contest within the state legislatures, whereas the 17A made them a popularity contest within the states. Well, in some states. 29 of the 45 states used some sort of direct election of senators by the time the 17A was proposed.

An interesting (to me) footnote to the 17A is that my state, Utah, was the only one to reject it. Six states ignored it, all in the Southeast. It was ratified in barely over a year, the seventh-quickest ratification thus far.

True... but only up to a point.

Each of the "automatic" background tasks you're performing does consume a bit of your overall capacity. Professions that have to deal with this explicitly because failing to manage it is dangerous call it "task loading". Take two examples: basic, automatic skills of divers and pilots.

Managing buoyancy is a skill that requires considerable practice for most novice divers. You have to regulate your breathing to maintain an average lung volume that keeps you at neutral buoyancy, not rising or falling (much) in the water column, but your breathing also has to be sufficient to supply your body's needs, and you also have to notice when your situation is changing so that you can't regulate with lung volume alone and then adjust the amount of air in your buoyancy compensation device (BCD).

New divers often find buoyancy difficult even with full attention. But with practice it becomes first easy, and then fully automatic, even when in the middle of the water column with no ready reference points. I have ~400 logged dives and I no longer even notice it, my perception is that I just "naturally" stay neutral except when I wish to rise or descend, and then it just happens. I find my hands occasionally reaching to adjust my BCD, but I don't think about it.

Pilots similarly have to learn to maintain level flight, using both direct observation out the window and instruments. It's hard and requires full attention for quite a while, but then becomes natural and automatic.

Once these skills are developed, they feel completely automatic, but they're really not, which becomes very apparent when you start piling up multiple "automatic" tasks, and possibly add a non-automatic task on top of them. As you add more "automatic" tasks, their quality degrades, and when you throw in a non-automatic task you find that both the automatic tasks are executed poorly and the non-automatic task is more difficult.

For divers, this becomes really apparent when, having mastered buoyancy, they begin trying to do other things like executing search patterns, or building or welding something, etc. Even something as simple as looking at a compass to determine heading while navigating can result in divers suddenly discovering that their "automatic" buoyancy control has disappeared and they are rising or sinking (usually rising). To some degree this can be mitigated with more practice/experience of the "automatic" skills, but it never goes away. One solution when performance of a complex skill is required (e.g. welding) is to eliminate buoyancy control as a requirement, for example by going to the bottom and adjusting the BCD so the diver is strongly negative and can't float up even with lungs filled to maximum.

Pilots address the problem by defining a strict hierarchy to three important automatic or semi-automatic skills. The phrase is "aviate, navigate, communicate", and the idea is that when task loading begins to overwhelm the pilot they focus entirely on flying the plane first, and only when they are confident they have that fully under control do they pay attention to course and heading, and only when that is fully under control do they pay attention to the radio. Of course, when not overloaded they do all three things at once, all the time, and don't really think about any of them, free to watch the scenery or whatever. But the automatic skills break under load, and when they break the pilot has to consciously "take over", and to prioritize conscious focus on the right tasks first.

Driving is another good example. It feels completely automatic... but I know firsthand that it's not. One way I know this is by what happens when I listen to audiobooks. I cannot listen to a book unless I'm doing something else. The stream of information isn't sufficiently dense to occupy my mind, so my mind wanders and then I suddenly realize that I haven't heard the last five minutes. But when I'm driving, that doesn't happen. It's clear that although driving feels automatic it does consume enough of my attention that I no longer have enough excess capacity to begin daydreaming. Driving + (good) book is close to my limit. I also find that doing non-word puzzles (e.g. Sudoku) is enough to keep my mind from wandering away from a book. Word puzzles don't work because I can't do two language-oriented tasks at the same time (some people can).

Anyway, the point is that (a) I hope there's more to this research than appears in the summary because it really does seem like it's well-understood and (b) background tasks feel completely automatic, but they're not. They still consume attention, just less, and subconciously.

Thanks! That makes sense. So seven-round AES is now 8 bits more broken than it already was, which is interesting but doesn't erode the safety margin.

Fascinating. I'll have to read more about the railroad bonds. That notion that the bonds could only be issued up to market cap, which could be expanded by issuing stock to the very people being paid by the bond sale proceeds, and the bonds in question all being government-backed, makes the current AI circular investment situation look downright ordinary.

That would be nice, but no, at least not unless NSA or similar has some non-public knowledge that enables them to model the confusion and diffusion precisely enough to estimate bounds. That's not impossible, but it doesn't seem very likely.

According to NIST and the AES designers, the round count was selected by breaking reduced-round variants and then choosing a high enough round count to provide a "safety margin". How much safety margin was needed was a subject of much debate, and there was some criticism that AES's margin wasn't large enough. AES-128's 10 rounds give about a 30% safety margin, and AES-256's 14 rounds give about a 40% margin. The article doesn't say what the "weakened" variant here was, so we don't know if those margins have been eroded.

Yeah, I think what we're seeing here is that attackers are resource-constrained, and that the hard part of creating an exploit isn't finding the vulns (there are tons of them), but figuring out how to turn the vulns into a working exploit chain. Merely having more fodder for crafting exploit chains doesn't necessarily make them easier to build... and keep in mind that the vast majority of attackers are glorified script kiddies. So the small number of really capable black hats may not have seen much benefit from AI yet. Probably not none but not enough to make a measurable difference.

If this theory is correct, when AI gets good at piecing together robust exploit chains, then we'll see the explosion in attacks, because all those glorified script kiddies will be able to get high-quality exploit chains cheaply and easily.

Note that none of this is applicable to nation-state attackers. They have plenty of smart, capable people, and plenty of money to buy chains. They weren't resource constrained before AI, and aren't now either. But they don't generally engage in mass attacks, so they don't register in the statistics.