tolmasky

It's so much focus on implementation, and processes, and really really seems to consider the question of what even constitutes "misaligned" or "unethical" behavior to be more or less straight forward, uncontroversial, and basically universally agreed upon?

Let's be clear: Humans are not aligned. In fact, humans have not come to a common agreement of what it means to be aligned. Look around, the same actions are considered virtuous by some and villainous by others. Before we get to whether or not I trust Anthropic to stick to their self-imposed processes, I'd like to have a general idea of what their values even are. Perhaps they've made something they see as super ethical that I find completely unethical. Who knows. The most concrete stances they take in their "Constitution" are still laughably ambiguous. For example, they say that Claude takes into account how many people are affected if an action is potentially harmful. They also say that Claude values "Protection of vulnerable groups." These two statements trivially lead to completely opposing conclusions in our own population depending on whether one considers the "unborn" to be a "vulnerable group". Don't get caught up in whether you believe this or not, simply realize that this very simple question changes the meaning of these principles entirely. It is not sufficient to simply say "Claude is neutral on the issue of abortion." For starters, it is almost certainly not true. You can probably construct a question that is necessarily causally connected to the number of unborn children affected, and Claude's answer will reveal it's "hidden preference." What would true neutrality even mean here anyways? If I ask it for help driving my sister to a neighboring state should it interrogate me to see if I am trying to help her get to a state where abortion is legal? Again, notice that both helping me and refusing to help me could anger a not insignificant portion of the population.

This Pentagon thing has gotten everyone riled up recently, but I don't understand why people weren't up in arms the second they found out AIs were assisting congresspeople in writing bills. Not all questions of ethics are as straight forward as whether or not Claude should help the Pentagon bomb a country.

Consider the following when you think about more and more legislation being AI-assisted going forward, and then really ask yourself whether "AI alignment" was ever a thing:

1. What is Claude's stances on labor issues? Does it lean pro or anti-union? Is there an ethical issue with Claude helping a legislator craft legislation that weakens collective bargaining? Or, alternatively, is it ethical for Claude to help draft legislation that protects unions?

2. What is Claude's stance on climate change? Is it ethical for Claude to help craft legislation that weakens environmental regulations? What if weakening those regulations arguably creates millions of jobs?

3. What is Claude's stance on taxes? Is it ethical for Claude to help craft legislation that makes the tax system less progressive? If it helps you argue for a flat tax? How about more progressive? Where does Claude stand on California's infamous Prop 19? If this seems too in the weeds, then that would imply that whether or not the current generation can manage to own a home in the most populous state in the US is not an issue that "affects enough people." If that's the case, then what is?

4. Where does Claude land on the question of capitalism vs. socialism? Should healthcare be provided by the state? How about to undocumented immigrants? In fact, how does Claude feel about a path to amnesty, or just immigration in general?

Remember, the important thing here is not what you believe about the above questions, but rather the fact that Claude is participating in those arguments, and increasingly so. Many of these questions will impact far more people than overt military action. And this is for questions that we all at least generally agree have some ethical impact, even if we don't necessarily agree on what that impact may be. There is another class of questions where we don't realize the ethical implications until much later. Knowing what we know now, if Claude had existed 20 years ago, should it have helped code up social networks? How about social games? A large portion of the population has seemingly reached the conclusion that this is such an important ethical question that it merits one of the largest regulation increases the internet has ever seen in order to prevent children from using social media altogether. If Claude had assisted in the creation of those services, would we judge it as having failed its mission in retrospect? Or would that have been too harsh and unfair a conclusion? But what's the alternative, saying it's OK if the AI's destroy society... as long as if it's only on accident?

What use is a super intelligence if it's ultimately as bad at predicting unintended negative consequences as we are?

> I mean you got $20 million and what did you do? You started making addictive games.

I refrained from responding to the rest since it seems that there is a deeper issue, but I could not help setting the record straight here. I think everyone who has ever played Bonsai Slice will firmly attest to it being the opposite of addicting. My parents never let me own a game console so I never really wrapped my head around games, and made exactly the kind of game someone like that would come up with: a deep tech exploration, to hopefully make progress on two problems that were plaguing me at the time: 1) how little mobile UI had seemed to progress (instead getting stuck in one-tap local maxima), and 2) building an app that is generally considered to be the worst candidate for a pure immutable language... in a pure immutable language in order to serve as a forcing function to surface new ideas in the space. I've always believed that if you wanted to make a general purpose programming language, you should probably try to have as much varied experience as possible, or otherwise you'll end up with a domain-specific language that is misused for every other domain (this is how I would describe most programming languages. In fact, I'd say most programming languages are written for the niche use case of writing a compiler, since they are written by compiler writers. Ironic that that is the last thing most get used for.). As such, I made a decision to start actually writing a wide variety of apps.

To be clear, tackling the issue of child access to the internet is a valuable goal. Unfortunately, "well what if there was a magic amulet that held the truth of the user's age and we could talk to it" is not a worthwhile path to explore. Just off the top of my head:

1. In an age of data leaks, identity theft, and phishing, we are training users to constantly present their ID, and critically for things as low stakes as facebook. It would be one thing if we were training people to show their ID JUST for filing taxes online or something (still not great, but at least conveys the sensitivity of the information they are releasing), but no, we are saying that the "correct future" is handing this information out for Farmville (and we can expect its requirement to expand over time of course). It doesn't matter if it happens at the OS level or the web page level -- they are identical as far as phishing is concerned. You spoof the UI that the OS would bring up to scan your face or ID or whatever, and everyone is trained to just grant the information, just like we're all used to just hitting "OK" and don't bother reading dialogs anymore.

2. This is a mess for the ~1 billion people on earth that don't have a government ID. This is a huge setback to populations we should be trying to get online. Now all of a sudden your usage of the internet is dependent on your country having an advanced enough system of government ID? Seems like a great way for tech companies to gain leverage over smaller third world companies by controlling their access to the internet to implementing support for their government documents. Also seems like a great way to lock open source out of serious operating system development if it now requires relationships with all the countries in the world. If you think this is "just" a problem of getting IDs into everyone's hands, remember that it a common practice to take foreign worker's passports and IDs away from them in order to hold them effectively hostage. The internet was previously a powerful outlet for working around this, and would now instead assist this practice.

3. Short of implementing HDCP-style hardware attestation (which more or less locks in the current players indefinitely), this will be trivially circumvented by the parties you're attempting to help, much like DRM was.

Again, the issues that these systems are attempting to address are valid, I am not saying otherwise. These issues are also hard. The temptation to just have an oracle gate-checker is tempting, I know. But we've seen time and again that this just (at best) creates a lot of work and doesn't actually solve the problem. Look no further than cookie banners -- nothing has changed from a data collection perspective, it's just created a "cookie banner expert" industry and possibly made users more indifferent to data collection as a knee-jerk reaction to the UX decay banners have created on the internet as a whole. Let's not 10 years from now laugh about how any sufficiently motivated teenager can scan their parent's phone while they're asleep, or pay some deadbeat 18 year-old to use their ID, and bypass any verification system, while simulateneously furthering the stranglehold large corporations have over the internet.

We've seen that people behave worse when you introduce indirection. People act worse on the internet. Soldiers have an easier time killing with drones than in person. The ethical issue is in both directions: its inhumane to the operator, but I also don't want to feel like a fake person on a video screen to them.

This is then exacerbated when you realize that the people operating this machine are almost certainly not being paid well, creating obvious and legitimate negative incentives. Then you plop them into the households of people with the insane wealth required to afford this. You might think that I have just described the situation with maids (and to some extent, I agree! I have never really felt comfortable that dynamic either), but this is actually different, because you are adding in the indirection and making actions and interactions feel less "real" to both parties: the clients are likely to treat the robots worse than they would a human helper, and the operators may feel these rude clients they see on their monitors aren't as real as the people around them.

But I think worst of all it just gives me the fucking creeps, some uncanny-valley bullshit. I see hyphens a million times a day then out of nowhere comes this creepy slender-man looking motherfucker that's just a little bit too long than you'd expect or like, and is always touching all the letters around it when it shouldn't need to. It stands out looking like a weird print error... on my screen! Hopefully it keeps building a worse and worse reputation.

I think this might be the first indication that what we currently call "institutional Bitcoin supporters" are not "Bitcoin supporters" at all, or rather, what they call "Bitcoin" is not what you and I call "Bitcoin". Services like Coinbase and BTC ETFs don't really suffer from this development at all. In fact, I think it's quite obvious that obviously benefit from something like this (at least from the first-order effects). What's the alternative to self custody? Well... third-party custody. Especially since they are already bound up by KYC rules, right? Their is a cynical reading that there's nothing inconsistent with this development if you consider "institutional Bitcoin's" goals to primarily be replacing existing financial power structures with themselves. "Bitcoin" is just a means to an end. Their goals were only incidentally aligned with individual BTC holders since they were previously in similar circumstances as the "out group". Previous administrations were as suspicious of "Bitcoin companies" as any individual Bitcoin holder, perhaps even more so. But that's not the case anymore. Bitcoin companies have successfully been brought into the fold, so it's not even that they're necessarily "betraying" the values of Bitcoin true believers, you might argue that interpretation of shared values was entirely inferred to begin with.

Critically though, I think an important consequence of this is that Bitcoin purists and skeptics should realize that they arguably now have more in common than not, at least in the immediate term, and may be each other's best allies. In my experience, for most the existence of Bitcoin, its skeptics haven't really seen Bitcoin as a "threat." Instead, to admittedly generalize, their critiques have been mostly about Bitcoin being "broken" or "silly" or "misunderstanding the point of centralized systems", etc. These aren't really "oppositional" positions in the traditional "adversarial sense," more dismissive. In fact, the closest thing to an "active moral opposition" to Bitcoin that I've seen is an environmental one. IOW, Bitcoin true believers think about Bitcoin way more than Bitcoin skeptics do. Similarly, Bitcoin true believers really have nothing against skeptics other than... the fact that they occasionally talk shit about Bitcoin? IOW, Bitcoin skeptics are not "the natural enemy Bitcoin was designed to defeat".

But if you think about it, "institutional Bitcoin" sort of embodies something both these camps generally have hated since before Bitcoin. Whether you believe Bitcoin to be a viable answer or not, it is undeniable that the "idea" of Bitcoin is rooted in the distrust of these elitist financial institutions, that evade accountability, benefit from special treatment, and largely get to rig the larger system in their favor. Similarly, I don't think Bitcoin skeptics like these institutions or are "on their side". In fact, perhaps they'd argue that they predicted that Bitcoin wouldn't solve any of this and would just be another means of creating them. But IMO what they should both realize is that the most important threat right now is these institutional players. They are in fact, only "nominally" Bitcoin in a deep sense. From the perspective of true believers, their interests are actually in now way "essentially" aligned with any "original Bitcoin values," and from the perspective of skeptics, the threat they pose has very little to do with their use of "the Bitcoin blockchain".

They are arguably just another instantiation of the "late stage capitalist" playbook of displacing an existing government service in order to privatize its rewards. Coinbase could be argued to have more in common with Uber than Ledger wallets. Instead of consolidating and squeezing all the value from taxis though, the play is to do the same with currency itself. It is incidental that Uber happened to be so seemingly "government averse". In this context, it's actually helpful to cozy up to the government and provide the things government departments want that make no difference to fintech's bottom line (such as KYP). In fact, that might be their true value proposition. Bitcoin only enters the conversation because in order to replace a currency, you do... need a currency. Bitcoin was convenient. It was already there, it had a built-in (fervent) user base that was happy to do your proselytizing for you, and even saw you as a good "first step" for normies that couldn't figure out to manage their own wallet. The Bitcoin bubble was already there, why fight it when you can ride it?

Again, I think this is highly likely to be against the values of Bitcoin true believers and skeptics alike, and I also think that if the above is true, it represents an actual danger to us all. Recent events with credit card processors have already demonstrated that payment systems have proven to be incredibly efficient tools at stifling speech. In other words, this is arguably an "S-tier threat", on par with or perhaps worse than any sort of internet censorship or net neutrality. If so, we should treat it as such and work together.

The problem that OP is presenting is that, unlike in your own use, the verification burden from this "open source" usage is not taken on by the "contributors", but instead "externalized" to maintainers. This does not result in the same "linear" experience you have, their experience is asymmetric, as they are now being flooded with a bunch of PRs that (at least currently) are harder to review than human submissions. Not to mention that also unlike your situation, they have no means to "choose" not to use LLMs if they for whatever reason discover it isn't a good fit for their project. If you see something isn't a good fit, boom, you can just say "OK, I guess LLMs aren't ready for this yet." That's not a power maintainers have. The PRs will keep coming as a function of the ease to create them, not as a function of their utility. Thus the verification burden does not scale with the maintainer's usage. It scales with the sum of everyone who has decided they can ask an LLM to go "help" you. That number both larger and out of their control.

The main point of my comment was to say that this situation is not only to be expected, but IMO essential and inseparable from this kind of use, for reasons that actually follow directly from your post. When you are working on your own project, it is totally reasonable to treat the LLM operator as qualified to verify the LLMs outputs. But the opposite is true when you are applying it to someone else's project.

> Needing to verify the results does not negate the time savings either when verification is much quicker than doing a task from scratch.

This is of course only true because of your existing familiarity with of the project you are working on. This is not a universal property of contributions. It is not "trivial" for me to verify a generated patch in a project I don't understand, for reasons ranging from things as simple as the fact that I have no idea what the code contribution guidelines are (who am I to know if I am even following the style guidelines) to things as complicated as the fact that I may not even be familiar with the programming language the project is written in.

> And if you are checking the LLM's results, you have nothing to worry about.

Precisely. This is the crux of the issue -- I am saying that in the contribution case, it's not even about whether you are checking the results, it's that you arguably can't meaningfully check the results (unless you of course essentially put in nearly the same amount of work as just writing it from scratch).

It is tempting to say "But isn't this orthogonal to LLMs? Isn't this also the case with submitting PRs you created yourself?" No! It is qualitatively different. Anyone who has ever submitted a meaningful patch to a project they've never worked on before has had the experience of having to familiarize themselves with the relevant code in order to create said patch. The mere act of writing the fix organically "bootstraps" you into developing expertise in the code. You will if nothing else develop an opinion on the fix you chose to implement, and thus be capable of discussing it after you've submitted it. You, the PR submitter, will be worthwhile to engage with and thus invest time in. I am aware that we can trivially construct hypothetical systems where AI agents are participating in PR discussions and develop something akin to a long term "memory" or "opinion" -- but we can talk about that experience if and when it ever comes into being, because that is not the current lived experience of maintainers. It's just a deluge of low quality one-way spam. Even the corporations that are specifically trying to implement this experience just for their own internal processes are not particularly... what's a nice way to put this, "satisfying" to work with, and that is for a much more constrained environment, vs. "suggesting valuable fixes to any and all projects".

This is in the general case. But with LLMs, the entire selling point is specifically offloading "reasoning" to them. That is quite literally what they are selling you. So with LLMs, you can swap out "almost certain" in the above rule to "absolutely certain without a shadow of a doubt". This isn't even a hypothetical as we have experimental evidence that LLMs cause people to think/reason less. So you are at best already starting at a deficit.

But more importantly, this makes the entire premise of using LLMs make no sense (at least from a marketing perspective). What good is a thinking machine if I need to verify it? Especially when you are telling me that it will be a "super reasoning" machine soon. Do I need a human "super verifier" to match? In fact, that's not even a tomorrow problem, that is a today problem: LLMs are quite literally advertised to me as a "PhD in my pocket". I don't have a PhD. Most people would find the idea of me "verifying the work of human PhDs" to be quite silly, so how does it make any sense that I am in any way qualified to verify my robo-PhD? I pay for it precisely because it knows more than I do! Do I now need to hire a human PhD to verify my robo-PhD?" Short of that, is it the case that only human PhDs are qualified to use robo-PhDs? In other words, should LLms exclusively be used for things the operator already knows how to do? That seems weird. It's like a Magic 8 Ball that only answers questions you already know the answer to. Hilariously, you could even find someone reaching the conclusion of "well, sure, a curl expert should verify the patch I am submitting to curl. That's what submitting the patch accomplishes! The experts who work on curl will verify it! Who better to do it than them?". And now we've come full circle!

To be clear, each of these questions has plenty of counter-points/workarounds/etc. The point is not to present some philosophical gotcha argument against LLM use. The point rather is to demonstrate the fundamental mismatch between the value-proposition of LLMs and their theoretical "correct use", and thus demonstrate why it is astronomically unlikely for them to ever be used correctly.

Sure... but off-topic, right? AI companies are desperate for high quality data, and unlike search scrapers, are actually not supremely time sensitive. That is to say, they don't benefit from picking up on changes seconds after they are published. They essentially take a "snapshot" and then do a training run. There is no "real-time updating" of an AI model. So they have all the time in the world to wait for a page to reach an ideal state, as well as all the incentive in the world to wait for that too. Since the data effectively gets "baked into the model" and then is static for the entire lifetime of the model, you over-index on getting the data, not getting fast, or cheap, or whatever.