What I'm saying is that we cannot observe deep structure. All we can observe is the surface structure. If we have access to the author or speaker we can ask clarifying questions, etc., but even that just produces more examples of surface structure.
But even here, in the Chomsky sense, LLMs clearly exhibit deep structure because they can write at length in an internally consistent manner. Importantly, early generations, GPT-2 and even GPT-3, did not definitively have this property; roughly, an object that was green at the beginning of a paragraph might not still be green at the end of the paragraph. This was strong evidence for lack of a world model.
Current LLMs do not show this behavior. We cannot prove that LLMs have a world model, in fact, their architecture seems to rule it out, but looking at it from a linguistic standpoint, they produce language in a manner as if to reflect a world view. That is, we cannot easily falsify the statement "LLMs somehow represent a world model"; and current examples of "disproving" their world view are so convoluted that even humans do not appear to (observationally) have a world view either.
I'm not making a claim as to LLMs having genuine deep structure or consciousness or anything like that. I'm claiming that we can't rule out current or future capabilities or make structural assumptions. Yes, they generalize from their training data, but unless you can make very specific claims about the kinds of things that they _cannot_ do, I can't take this statement as particularly compelling.
Oh, now I see where we have an actual difference of opinion. I don't think you can deny that even Finnegan's wake proceeds one token at a time; your interpretation of it may require more context or out-of-order interpretation, but that's just as true when observing text in German or Japanese, which have word ordering constraints that are alien to English speakers. How it was written is irrelevant; all we can observe is how it was presented. Of course we can observe each other's inner life, but we do so one token at a time, even if the process of producing each token is done (internally or actively) via a backtracking or zeitgeist approach.
You seem to believe, on a more fundamental level, that LLMs are simply not capable of producing text that has deeper connections to itself or represents abstract thoughts. In my opinion, 99% of text written by humans does not show this, just as 99% of text produced by LLMs does not show this, but both have the capability, and I don't believe that LLMs are constrained in such a way that they can never do this.
I dunno, man, I looked at that text and I see one word after another.
Obviously language and the connection to human thought is more subtle than this; I think we all have a rich inner life. Just from an external perspective we can't observe it; all we can see is the token/phoneme stream. I'm just saying that it's a mistake to try to criticize LLMs on this basis because it's hard to see how the same criticism would not apply to any system (like humans) that generate language.
At least "glorified autocomplete" is technically accurate, even if vastly underestimating the capability of LLMs. It's just trying to make something very impressive sound trivial.
From an external standpoint, talking to another human, it's like the other human says one word and then says the next word. That's just how language works. Humans look like "glorified autocomplete" from this perspective.
I mean, looking at the time evolution of the state of the universe, one could say that all of physics and creation is "glorified autocomplete" to posit a next state of the universe given current and past state.
> Contrary to how it may seem when we observe its output, an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot.
I think this metaphor is so strained as to not be useful. I think key here is that the authors say "without any reference to meaning", which is a heavily loaded term, that does definitely apply to parrots, but does not apply when you apply it to immense bodies of text.
Namely that language embeds meaning in language. A sentence being written by a human (as a starting point) is designed to have consistent meaning. While it is possible to write syntactically correct meaningless text, that is not what most of human language has done; the meaning cannot be removed from the text.
This I think is clarifying, from the same paragraph in the text:
> ... the training data never included sharing thoughts with a listener, nor does the machine have the ability to do that.
That's just facially incorrect. The training data is entirely about sharing thoughts with a listener. Else why is the text being written?
The Northgate was a great keyboard, and a company called Creative Vision Technologies briefly owned and sold keyboards under the "Avant" brand. I had one of those, it finally succumbed to some sort of electrical issues so I let it go. You can still by old Northgates and Avants on eBay, but ~$300 a pop.
No keyboard since has ever matched that level of responsiveness and tactile feel. The downside is just that it was so loud. Now that everything is open floorplans I wouldn't use one even if I still had one.
I feel strongly that this conspiratorial mind-reading approach to this sort of issue is just counterproductive.
What all the governments (and non-governments, frankly, there are many supporters of these things) are asking for is excluding minors from certain websites and services.
The problem is that this translates to age verification, which translates to identify verification, which incidentally gives states and other actors a variety of other tools they can use for anti-civil-liberties purposes.
In the end their motives are just irrelevant unless there is a clear way to exclude minors from certain services without going down the chain towards identity verification. Such a way does not exist, so we have to fight it here, at the point where the basic ask emerges.
The article talks about the possibilities of malicious cloning of these tokens by third parties, but fails to identify the much more common use case, and one that makes this scheme useless for age verification.
It's one thing to be concerned about someone stealing my credential, but another to prevent the transfer of these credentials, especially if they are limited use credentials.
The entire point of age verification systems is to prevent minors from accessing certain resources. I think we all know that this is basically impossible; but what these various governments and social media companies want to do is to make it high friction to do so.
The highest friction version of this is that the credential ties to a real world identity somehow; maybe locked behind legal barriers, etc., but if a minor is caught using someone's credential, then the person whose credential they are using can be investigated, and, if necessary, charged with a crime roughly equivalent to providing alcohol to a minor. Without the possibility of real world enforcement, none of these identity solutions can possibly work.
Keep dreaming of a technological solution -- there is none that does not lead to the world that FIRE is warning about, except to accept that we can only make a solution "good enough" and leave it at that, without expanding into full on identity verification. The solution here is likely to just try to provide better abilities for parents to monitor and limit their children's use of the internet. Let individual parents decide on the level of harm that they are willing to accept, and accept that there will be ways to work around this even if parents are vigilant, but just try to reduce it on the margins.
As another happy Enpass user I don't think it is significantly different. The exception being that the KeePass frontends are really just as expensive as Enpass is now.
I also got a good deal on the lifetime pro membership before they moved to more of an annual model, which factors into my decision.
Which Keepass frontend do you use (if you use one on mobile)? Keepassium and Strongbox seem to be the ones that people talk about, but they are pricey too. I don't know anything about AuthPass, but I'm reluctant to use a free product; I'd rather use an inexpensive one, just to hopefully thwart supply chain attacks on that front.
It is the easiest thing in the world to make it illegal. But while making the sale of alcohol to minors has some teeth, as above, the sale of these "tokens" is so low friction as to be laughable.
This is another of the many many solutions in the "the age limitations that exist currently are easily circumvented" class. While it may be effective in some limited ways, it's not clear that it's better than say parental controls. And it is obviously deficient compared to identity verification solutions that do not preserve anonymity. You can't pretend that this solution is "just as good" when it isn't. The identity verification solution is bad not because it is ineffective at its direct aims but because it has disastrous side effects.
I'll try one more time -- how many people, that is, individual human beings, do you believe have this capacity and the desire to use it? Is that number, like, 5? Is it 50? Or, more likely, is it 50,000? Of those people, how many successfully do anything with this?
None of the pattern you describe is at all effective in a democracy unless you can get people to support your candidate or support your issues or whatever. You can't just buy that. There's no direct path to spending money to make people support you or believe in you or excuse your actions. There is no formula or process that works. All of this hinges on the individual, and I'm extremely unconvinced that personal wealth is an interesting axis. Where personal wealth intersects with power is another matter, but that powerful people can wield power is not an interesting or astute observation.
There is no "done correctly". Intrinsically you simply cannot do this, and you have to accept this as almost axiomatic.
If the online age token is not attached to a human identity in a strong way, then it can be transferred to others without friction. The only way we can make it sticky is a mix of technological solutions (how we tie to identity) and real-world systems -- having the ability to prosecute misuse and to link to a physical identity.
> I think lots of people around here would agree that "the rich" are doing nefarious things
I'm sorry, while it may be true that many people around here believe this, you have lost me in particular at this line.
It may be true that some rich people, like people from all walks of life, are occasionally doing nefarious things. I think generalizing this to an entire group, "The Rich", sounds as off-kilter as people talking about crystal energy or pyramid aliens or bigfoot to my ears.
To put another way, the supply of rich people who desperately want to change the world to support their desires vastly exceeds the ability of the world to meet that demand. For every rich person who successfully bankrolls and gets a person elected there are 10,000 or 100,000 who fail to do so at early and late stages of this process, and imagining that "The Rich" have access to some facility to change the world is just untenable.
Adults can be prosecuted for selling alcohol to minors (I mean, they can be prosecuted for selling drugs to other adults even). Because there is a physical handoff, there is risk in doing so, and thus there is high friction, which we deem to be sufficient to curtail the risk.
No such friction or accountability can possibly exist when we're talking about a bunch of yubikeys, unless we attach identity to them (THIS IS A BAD IDEA).
But it's not a question of poor execution -- there is simply no way to execute this. There is no way to achieve the goal (age-restricting websites) without identity verification. There are any number of half solutions that will solve 80% of the problem, but to move the needle past that requires identity verification. Even then, as the article points out, we only move to 90%.
This is underselling it -- an adult can sell these yubikeys untraceably, pocket the cash, and never think about it again in their life.
If we tie the person's identity to the token (THIS IS A BAD IDEA) then if underage use is detected, the adult who sold the token can be prosecuted for selling adult tokens to minors.
Yes, it screams "conspiracy theories" because it is literally a theory that involves malign conspiracies. And yes, this is repellent to many people.
It is most repellent, I think, to people who genuinely hold the belief ("I want to prevent minors from using websites that are generally agreed to be harmful to minors"). When you tell these people "no, you actually want a totalitarian state that controls what adults are allowed to believe", they think you're crazy, because they don't believe this.
That this is an inevitable consequence of the solution to the problem that they want to solve is a question of tradeoffs that people are not generally aware of, and I think it's way more important for people to be aware of those tradeoffs without being told that it is the illuminati and the freemasons and George Soros and Fox News trying to Orwell their way into a global police state.
If you mean this to say "this is probably the best we as a society can do on balance from a 'worse-is-better' approach" then I'm pretty much in agreement.
But obviously this doesn't "solve the problem". It's another bandaid with an extensive list of failure modes and tradeoffs. It falls into the class of "the age limitations that exist currently are easily circumvented" type of solution.
In my opinion it is fine to leave it there and accept the tradeoffs. We could mandate better website marking, and mandate better device or app-level mechanisms, and improve monitoring and restricting tools, or we could do even less and keep it more or less heterogenous.
But I do not agree that it is "moatism" to talk about it on the website side. There is a real and genuine desire to actually have the kinds of age restrictions that are only possible with strong user identity broadly deployed. Refusing to engage because of imputation of malign motives on the other party's part is not going to persuade anyone, especially if they do not personally have those malign motives.
I'm maybe a bit of an outlier here in that I do think that this is a genuine grassroots good faith effort to "protect the children" that does not have sinister ulterior motives. I know plenty of parents who have expressed enthusiasm for the idea of age-restricting websites.
"Why now" I think is pretty obvious -- the age limitations that exist currently are easily circumvented, but have given enough of a plausible deniability aspect that politicians have been able to skate by. There has been increasing research and media dedicated to the idea that there are aspects of the internet which we should be shielding children from. While many of this research is dubious, there's a rising moral panic around it.
The core of the problem is that there is no possible implementation of age verification that does not also require identity verification. In this I am in strong agreement with the article, but the use of paranoid and dramatic language as in this article only alienates people who find the conspiratorial tone to be reverse polarizing.
Not a doctor, but the overdiagnosis concern is at the intersection of three phenomena:
1. Imaging is expensive, just in dollars and time, even without analysis
2. Imaging is not without impact -- CT scans, especially full body scans, expose the body to ionizing radiation
3. Imaging is time-consuming
The net result of these means that full body scans are difficult to interpret. If a doctor given a patient complaint suspects a condition that is sufficiently non-specific that a full-body scan is required, then the scan will be interpreted through the lens of the known progress of the differential diagnosis. And typically these scans must be done without a healthy baseline, so minor findings in this context might have significant diagnostic power when combined with history or other findings.
But on a healthy patient, minor findings are very likely to be noise, because we don't have a great deal of experience with scans of healthy people, for the reasons above.
This technology, if it pans out, gives a way of inverting 1, 2, and 3. If every healthy doctor visit includes one of these scans, then the medical field gets experience interpreting them, and more importantly, when new symptoms occur, previous scans can be compared to determine whether a particular finding in the current scan is new or has changed.
But even here, in the Chomsky sense, LLMs clearly exhibit deep structure because they can write at length in an internally consistent manner. Importantly, early generations, GPT-2 and even GPT-3, did not definitively have this property; roughly, an object that was green at the beginning of a paragraph might not still be green at the end of the paragraph. This was strong evidence for lack of a world model.
Current LLMs do not show this behavior. We cannot prove that LLMs have a world model, in fact, their architecture seems to rule it out, but looking at it from a linguistic standpoint, they produce language in a manner as if to reflect a world view. That is, we cannot easily falsify the statement "LLMs somehow represent a world model"; and current examples of "disproving" their world view are so convoluted that even humans do not appear to (observationally) have a world view either.
I'm not making a claim as to LLMs having genuine deep structure or consciousness or anything like that. I'm claiming that we can't rule out current or future capabilities or make structural assumptions. Yes, they generalize from their training data, but unless you can make very specific claims about the kinds of things that they _cannot_ do, I can't take this statement as particularly compelling.