sfink

Yes, if you put stupid words in my mouth, you can argue that I am saying stupid things.

Of course AI-less wars have always had innocent victims. I'm not even arguing whether using AI will produce more or fewer such victims, and at the moment I don't see how anyone could convincingly argue either way. The use of AI is probably a pretty minor driver behind the number of inappropriate targets.

I do argue that people (one or more humans) must be responsible for these decisions, and furthermore having a human in the loop is a necessary prerequisite for such responsibility. (A human is still responsible if there is no human in the loop of individual decisions; the people who allowed that usage were in the loop of making that decision, and thereby have responsibility.)

If we start agreeing that we can just blame the AI for any unintended casualties, then we lose most incentives to reduce such casualties. Even if an AI makes "better" decisions than a human, wars are always resource-constrained and decisions will be made based on responsibility and blame. For example, the decision might be "the AI needs too much data to make accurate decisions, gathering those data is expensive, let's give it less." We cannot allow the "cheapest" decision when push comes to shove to be the one where innocent people are killed because the perpetrators have externalized the responsibility.

In the case under discussion, we're not at the point where anyone is claiming that an AI acting alone will kill fewer unintended people than an AI+a human. The people who develop the technology are specifically saying it is not ready yet. In large part, that comes down to responsibility again -- if Anthropic were to promote that use, then they would be part of the human chain that is responsible for those deaths.

In addition, using AI for autonomous killing is ridiculously dangerous and ill-advised at the present time. They have common and unpredictable failure modes. They have biases that could change something from 99% accurate to way less due to a change that should have no impact on the decisions. When using them for programming, they'll do things like directly go against instructions (whether or not they apologize afterwards) and getting stuck in loops. What are the analogous failure modes when you give them control over lethal weapons? They are not ready.

> > The question is about the US government using its influence in a competitive environment to hamstring one company.

> No company is owed contract from the US government.

The discussion is not about the contract. It is about the supply-chain risk designation. The contract issue is separate (and real! The DoD was trying to retroactively change the terms of an existing signed contract). The DoD is free to not sign any contracts with Anthropic. It is not right, and illegal in my and many people's opinions, to misuse a tool intended for preventing harm from foreign adversaries. It was retaliation, plain and simple.

> We are getting into the territory of "70% of traffic crashes are blamed on sober drivers, therefore driving sober is a crime".

You are, for some reason. I was not. I'm only countering an argument that the particular tragedy that made a splash in the news cannot be discounted by saying "sure it's awful, but irrelevant because AI was not involved." AI was involved. We don't have to imagine hypothetical scenarios where AI was involved and something bad happened. And I'm not saying AI was the cause, or the only cause, of that tragedy.

>> My supposedly very vivid imagination translated this to reckless decisionmaking that led to a large number of innocent civilian casualties, aka war crimes.

> As I said, extensively, the reality of warmaking will inevitably lead to innocent civilian casualties.

The word "reckless" was critical to what I said, because it's part of what makes something a war crime.

Ok, I give up. You're equating war crimes and any civilian casualties. Or wait, then you go on to equate war crimes with any suffering of innocents. Why? Why are you so set on dismissing the possibility of a war crime being a thing that happens, and has an actual definition that goes well beyond innocent people suffering? You seem motivated to steamroll any obstacle to using maximal force regardless of the consequences. Why? Hegseth, is that you?

Never mind, I don't think I need to know your answer. You are a rando on the internet to me.

It did not say "I assume cheating". It said you're probably cheating, as in of you look at the population of students who are using such constructs, over 50% of them are using them because they are cheating. Given that accusations of cheating were not made when such constructs were found, it seems like that probability was not factored in when deciding whether or not to advise someone.

Heh, dumb idea of the day: do a "flipped classroom" in a different sense -- instead of a student taking a quiz or exam, you provide a clueless AI and the student has to train it on the subject at hand, and then the AI takes the test.

Lest anyone think this is a brilliant idea: an LLM would likely be much better at doing the ~~RLHF~~ RLMF, so you haven't actually gained anything. But I feel like there may still be the kernel of a good idea somewhere in here.

Perhaps your assignment is to iteratively train the AI, which is pre-prompted to seek out every reasonably possible failure mode when it generates solutions. So the learning mechanism is to train an adversarial AI on the subject matter as a way of learning it yourself. This does not solve the problem of cheating with an LLM, it's an alternative pedagogical approach.

> The American society is just as corrupt and screwed up as any other country

That's a weird phrase. It hints at an assumption or widespread belief that we aren't as corrupt or screwed up. Which I get; I was born and raised in the US, and I got a full dose of American exceptionalism too.

Corruption in particular is a tricky thing to define. Currently, it is patently obvious that we are quite a bit more corrupt than the median. The recent big spike is new, but if you pay attention to the news, you'll see that over and over again we're being presented with evidence that this is not a new thing. This is a longstanding thing and the only thing that's new is the degree of exposure. We may have been this corrupt in the past; there's some pretty mindblowing stuff in the history books. But I don't recall this level of naked corruption at any point in the country's history.

As for screwed up, we've held the crown on that for decades at this point. We've exported that trait throughout the world, so it's getting a little harder to tell. Also, there's plenty of homegrown screwed-upness for our variety to interbreed with.

But this is just an American talking to an American. To the rest of the world, the corruption and decay is and has been extremely obvious. Talk to Europeans. They see us as a third world country who happens to still have the biggest military and hold the world's reserve currency. (For now.) But the "third world" part is not an exaggeration, from the people I've talked to. And I can't really argue? Fascism is risen, we're ok with secret police disappearing people from our streets, we're ok with our public money being used to pay people to vote for a particular party. (If you object to "ok", protesting and writing strongly-worded messages on social media doesn't count. I confess that that's all I do.)

> Communism is better than capitalism.

Here's where I disagree. "Structure X is better than structure Y" conveys very little information and is not a useful framing. It's a real example of "ideas are worthless, execution is everything". Capitalism is fine. Free trade is powerful; it's capable of and has accomplished an enormous amount of good. It just needs to be protected and managed; free trade being good does not mean the freer the better, for any definition of "free" that someone wants to come up with. Communism is fine too. It's not even incompatible with competition. Hell, the actual systems in use today aren't even one or the other, the pure forms don't work and aren't acceptable to any population.

But perhaps you mean late stage capitalism and late stage communism. Those both suck, from what I can tell.

The main pic makes absolutely zero sense. The sides are tilted upwards, so clearly they were moving towards each other. But there's a big gap, so clearly they're moving away from each other. I was mentally sketching out a scenario where it was first one then the other, until I realized it was just AI slop.

Oh gods, I don't give it write access to my actual DB.

For my small-scale sqlite dB, it gets read access, and I encourage it to test modifications by copying it somewhere and writing into that.

Scale-dependent, but I hope to not have to work at a scale where it gets write access to the production DB. That just seems like asking for trouble.

> Especially in a time of war.

The US has not declared war on Iran. No wartime legal mechanism is at play, and cannot be until Congress decides to declare war.

Ok, I'll be precise: the main dispute was over fully autonomous decisions to use lethal force. Because if the AI chose incorrectly, the military would be attacking targets that it shouldn't. Whether that is a war crime hinges upon whether it was reckless. Anthropic's position was that AI is not capable of making such decisions, which is another way of saying that giving that decision wholly to an AI is reckless. I can stop using the term "war crimes" if it bothers you so much, but I'm failing to see a distinction between what Anthropic was fearing and the definition of what a war crime is. These are fears, so this isn't about whether one specific action was a war crime or not. This is just naming what Anthropic was worried about and why.

> Anthropic wanted them to be the last word in every decision, and DOD could not allow that

Cite one source for Anthropic demanding to be in the loop. This was about the terms of a contract. If I'm agreeing to sell you a medicine, under the condition that you do not use it to euthanize homeless people, that does not mean I am demanding to follow you around and make you ask me whether each use is ok.

> Nobody is destroying anybody's business - pretending that Anthropic could never find any users for their AI outside DOD is plain silly.

Strawman much? The question is about the US government using its influence in a competitive environment to hamstring one company. It doesn't matter whether there are other customers. It doesn't matter whether the company can survive the attack. What matters is whether it is legal and ethical for our public resources to be used for retaliation in this way.

> US Society has mechanisms for preventing and investigating war crimes....I doubt that "give control over it to some random guy that makes AI models"

This is an invented claim. Cite a source.

>> I quit using ChatGPT entirely over this incident > > And this is supposed to be an argument because....?

The argument was the first part of the post. This part was me giving my perspective on things, in this case it was about how I don't find Anthropic's position to be good, just that OpenAI's is far worse. My post was not for you exclusively; based on what you have written, I have concluded that there is no value in debating you or attempting changing your mind.

>> They agreed to a contract where the DoD could and in practice is using their AI for committing war crimes. > > You have a very vivid imagination. Not trying to use it to state things that could be misinterpreted as statements of fact would be nice.

Ok, fair. Sticking to facts: the crucial part of the negotiation was about whether the DoD can use Anthropic's products for autonomous decisions on lethal force. Since then, it has come out that AI was used in the decision to target an elementary school for girls that killed 120 schoolchildren. (The Palantir Maven Smart System, fwiw, not Claude.) The Pentagon investigated and found that although a human was theoretically in the decision loop, in practice the oversight was minimal and ineffective. "No one remaining on those teams [the teams tasked with minimizing civilian harm] reviewed the strike plans in advance of the attack on the school, according to Bloomberg." Those teams had been gutted, with Hegseth explicitly stating that they were getting in the way.

My supposedly very vivid imagination translated this to reckless decisionmaking that led to a large number of innocent civilian casualties, aka war crimes.

Anthropic agreed to uses that involve a human decision, so this is a non sequitur.

The whole conflict is specifically about uses that do not involve a human decision. The DoD agreed to a contract that said it wouldn't use Anthropic for that, and then tried to change the terms.

Incidentally, the DoD used Claude in its decision that resulted in killing 120 children at an elementary school for girls. The Pentagon concluded that there was human review (which would mean that it fell within the contract's terms) but that the review was hollowed out (which could mean that it was a violation of the contract).

You're skipping the supply chain risk designation, which is an extraordinary and punitive measure. Without that part, you'd be correct that this is a nothingburger. "Government decides to not renew contract with corporation X and goes with Y instead" is not headline news.

Then the Pentagon shouldn't sign a contract with them. (Or in this case, they shouldn't try and fail to renegotiate the terms of an already signed contract, and then retaliate in order to get their way.)

Not the OP, but personally I'd say both: unequal enforcement by the administration, and an incorrect and politically motivated decision by the court.

...I suggest you read about what actually happened:

Anthropic: "We'll sell you access to our AI as long as you don't use it for war crimes or mass surveillance."

DoD: "Ok, deal."

...time passes...

DoD: "Wait, no, we want to be able to use it for war crimes too."

Anthropic: "Too late, you agreed already, and we won't sign a contract for that. You can keep using it for everything we agreed to."

DoD: "Change the terms or we will destroy your business."

Anthropic: "No."

DoD then signs contracts with OpenAI and Google that permit them to commit war crimes and mass surveillance as long as they get legal cover, and proceeds to apply as much leverage as it can to destroy Anthropic's business.

----

And for the record, Anthropic agreed to all "yucky Army things" usage that was an extension of past yucky Army things, they just did not agree to a new class of yucky Army things: namely, fully autonomous killbots. Even there, it was only "not yet, it's not ready". Your whole comment is a gross mischaracterization.

I quit using ChatGPT entirely over this incident, and that's not because I find Anthropic's position to be saintly. They agreed to a contract where the DoD could and in practice is using their AI for committing war crimes.

So you renegotiate the contract. You don't get to retaliate.

Only it turns out that with sufficient corruption, you do.

You could use the same logic to defend the military declaring that Anthropic is full of pedophiles and therefore they cannot use their products as it would be contributing to pedophilia.

Anthropic offered a contract, with certain conditions, as do all contracts everywhere; that's their purpose. The military did not want to agree to those conditions. If they had stopped there, and refused to sign the contract, all would be good. (It's actually worse, they did sign such a contract, and then decided they wanted the contract to say something different than it actually did. "Pray I don't alter it any further.")

> >The Department reasonably feared that Anthropic might manipulate Claude’s design to prevent it from performing national-security functions that the Department deems contractually authorized [emphasis mine]

That is speculation, and you can't call it a "textbook designation" without discussing whether there's a basis for that "reasonable fear". Again, that argument also works for declaring you have a reasonable fear that giving money to Anthropic will support pedophilia. We do not know the specifics, but I at least have heard of zero evidence that Anthropic would sabotage something they signed a legal contract for, and yet I have an abundance of evidence that this administration will use whatever contortions are necessary to pressure and punish those who interfere with it getting what it wants. It all hinges on the word "reasonable", and based on the evidence that is public, this specific fear seems more ridiculous than reasonable to me.

If the government somehow had a way to force Anthropic to sign a contract that it did not want to sign, then this fear might become more reasonable. The twist is that this supply chain risk designation is exactly that. If Anthropic now capitulated, the accusation of supply chain risk (eg from Anthropic employees acting alone) would be justified. So the only way Anthropic can reasonably be considered a supply chain risk is because it is accused of being a supply chain risk.

This is a textbook example, yes, but it's a textbook example of corruption and judicial capture.

> Though they'd probably put the DoD on the cybersecurity whitelist today, the very idea of the claude whitelists for certain functionality already exists and is being used by them today.

And if they signed a contract permitting fully autonomous killbots, then using the whitelist mechanism would be a contract violation. I have some degree of faith that we'd know if such a contract were signed, because half the staff would quit. (As opposed to half the Google staff quitting after signing such a contract, which has been proven to be an incorrect expectation -- such a contract was signed, and I've heard of exactly one person quitting over it. There may be more, I don't know.)

Nope, not the only one. I mean, it's useful to look ahead to the next tier of problems, but I fully agree that there are plenty of problems to look at right now that don't require any predicting. And in my mind, the "next tier" is more about wealth distribution and societal malfunctioning than robot apocalypses; the robots will get their turn, but for a while they'll have to be content to wait. In the meantime, kids are acing tests on subjects they know nothing about, and we software engineers are writing code that we're not capable of debugging or even managing.

Any system is set up for and tuned around absorbing some amount of input at a time -- nutrients, say. If it gets too little, it starves and dies. If it gets too much, it breaks down and dies. A system can adapt over time, at a limited rate, and become capable of handling more or less.

Drinking a glass of water is beneficial and necessary. Drinking 100L in one go and you're dead.

The argument is that the rate of AI influx is too high for our human systems to metabolize, and so the system will break down. Not just operate in a degraded way, but break down and cause mass catastrophe. Even there, it doesn't necessarily mean the end of humanity, just 99.9% of us.

What the argument is not: AI is fundamentally poisonous. We are incapable of handling any level of AI. We will never be able to incorporate more AI. AI provides no benefit.

So quitting AI but still working on AI is not the contradiction it seems: you can quit accelerating AI and work on metabolizing it. And in fact that work is very necessary right now.

That's a false sense of security. Something could have a serious issue, but provide overwhelmingly more benefit (cars, say), and thus people would use them anyway. If people did not use cars, then traffic fatalities would not be a concern.

People are using AI.

Your argument is that AI cannot be a serious concern because it has major known limitations. But most of us are using it despite the major known limitations. So the existence of the known limitations is irrelevant to determining what level of harms are possible -- unless the known limitations are directly limiting the harms. In practice, it seems like most current harms, and many of the (not stupid) feared harms, are completely possible despite the known limitations. Sometimes even exacerbated by them. ("Sorry, you told me to never stop your insulin pump, but I stopped your insulin pump while diagnosing a test failure. I will record in my memory file that I must always... hello? user?")

The (not stupid) parenthetical might be too weaselly for you, but that's how I see the "AI is going to rise up and intentionally subjugate or exterminate humanity" brand of doomerism. While I don't think that's strictly impossible, there are so many other ways harms that are much more imminent and much harder to defend against, that it seems foolish (to me!) to worry about those.

If you can stop a problem by pulling out the plug, you don't have a real problem.

If you don't want to pull out the plug, you do have a real problem. (Don't want, or can't afford to, or can't be bothered to so extremely that you never will, or whatever.)

This is great data, and good but somewhat flawed analysis.

The good part is showing that the drop in thinking tokens persists no matter what grouping you slice across. They make a very persuasive case that there's something systematic going on.

My usual complaint about these "they're nerfing the models, I feel it in my bones!" posts is that they don't account for the workload changing. From working on my own stuff, there are a series of evolutionary/de-evolutionary changes that happen in a heavily AI-written codebase. Initially everything goes great. Then the AI takes on too much technical debt. Improvements slow down and regressions creep up until it becomes a never-ending game of whack-a-mole just to keep up. So you direct some (probably AI) effort towards cleaning things up, reducing duplication, and removing patches for problems that are better fixed with a design change, or workarounds because the harness saw the wrong version or you incorrectly described a problem and it strenuously solved a non-problem. That gets you back up to cruising speed for a while, then the project exceeds some hidden threshold for size in latent space or something, and further progress has to rely on attending to one aspect of the codebase at a time. Once again, the architecture becomes the limiting factor, but in a subtly different way. My sense is that it all boils down to some sort of "attention capacity" -- is your codebase and problem space amenable to looking at one aspect at a time, or is it all snarled together? -- but that's an essay that I'd love to write but really don't have enough experience to do justice.

Anyway, the details don't matter. The point is that not only can you not assume that the difficulty presented to the AI is roughly constant over time, but also there's evidence to believe that it will be normally be increasing. (Unless you're constantly starting new projects instead of continuing old ones.)

That's why I like this writeup. Focusing on thinking trails doesn't eliminate the problem of snowballing difficulty, but it does sidestep the worst of it. In fact, I'd expect the same setup to think more as the complexity/sloppiness creeps up.

The flawed part that bothered me was that it feels like there's a little bit of a predetermined conclusion that thinking is a magic sauce that makes everything taste better if you spread it on everything. I want a high variance on thinking, especially between interactions. A smarter model would have a higher variance, in my opinion. So the accusatory tone (perhaps I should reread it? My first impressions are often wrong) around "look! it doesn't bother to think at all a lot of the time. That can't be right!" seems misguided to me. It should think when it needs to, and if its thinking was clear then it won't need to re-think over and over again; it's all still in the context.

Forgive the anthropomorphization, but consider those studies of chess experts vs novices. Novices have to work way harder, working through all kinds of things from scratch, while the expert instantly and effortlessly recognizes what's going on.

But anyway, the main takeaway fully survives this criticism. The models appear to systematically think less over time. It doesn't matter if a smarter model might be able to think less for the same quality; this is happening over the same model.

I'm assuming all coin flips are deterministic based on the task. In this case, it'd be equivalent to generating a slightly longer hash and using a couple of bits for the "coin flip". (Or just generating a new hash with 1 or 3 bits or whatever you need.)

Well, given that I knew that I was probably missing something, and the fact that their solution made no sense with the constraints I was using, pretty strongly implied that there was an additional constraint. And that's a fairly obvious one to have.

I can't do the math to prove it, but their solution still seems wrong to me. Rather than generating and storing and searching so many hashes, it seems like you should get partway there with a different sampling procedure that doesn't do quite as well with the inconsistent sets of servers, and then only use duplication to limit the consistency loss.

Simple example: use their scheme but instead of choosing the first server to the left of the probe, grab the first two and flip a coin to decide which one to use. That already spreads the bucket variance out a bit, without using any extra space. It does have a penalty in that if one balancer has a server that the other doesn't, then it spreads out the range of probes that could get a disagreement. But I don't know how to quantify that; if the balancers disagree on the set of servers available, you have to produce different results part of the time, and I haven't thought through how to characterize when that disagreement is "bad".

Then you could extend that to looking at the previous 8 servers. Or the previous k tickets, if you give each server a ticket for each weight unit.

The math works out easier if you sample regions of probe space rather than server counts: hash the incoming task, map that to a range of space on the number line, and all servers within that range are your candidate set. Choose from that set, making the candidates be either equally weighted, weighted proportionally to their weight (size/capacity/whatever), or weighted by how much they got shafted by the random distribution of the server hashes.

I get EBRAINTOOSMALL when I try to work out the statistics, especially when I try to figure out what the inconsistency cost is, but intuitively it still seems better than recording a bajillion hashes for each server. (With the latter sampling mechanism, you'd need to deal with the possibility of probing a window with no server in it, either by double hashing the task and trying again, or expanding the probed region. Details schmetails.)

In practice, I'd probably simulate it and look at the distributions. Or nerd snipe a math geek.

Ah, right. The joys of being a fool in public.

The part I missed is that the load balancers don't have a consistent view of the set of servers. There is no magical synchronization scheme that creates that consistent view. You want load balancers with slightly different ideas of what servers are available to mostly make the same choices for the servers they do agree on.

Doh! I should have been able to infer that from the original solution.

Um.

I read the article thinking it would make for a great brain puzzle, but I quickly decided there's something wrong with the question setup because the initial solution didn't make sense. I assumed it was just missing a constraint that would be revealed later, but I'm still not seeing it -- the article just kept patching up the flaws in the wrong solution, the one that is more complicated than the straightforward one.

I'm probably still missing something obvious? It's probably something to do with "...in a way that does not require large changes when servers are added or removed."

But let's start with the problem as initially posed: you have an infinite stream of tasks and you need to deterministically assign them to N servers. (Perhaps you have to shard the collections of servers, so not every load balancer knows about all of them? But no, that would break the solution in the article.) Ok, then hash the task request (I assume that you hash it, the article doesn't explicitly say, but that's how you'd get determinism) and take that hash mod N, that's your server index.

Why hash the servers too? If you roll 6 dice, and then another one to choose which die to use, you're not getting any more randomness. You're matching up two sides, the tasks on one side and the servers on the other; no need to randomize both.

Ooh, but that's not a perfect distribution? Ok, if the hash value is large enough to be in the at most N-1 slop values at the top of UINT_MAX, then roll again (compute another hash). But CF is happy with 8% unevenness, there should be no problem with this 0.1% or whatever.

Also, how do they find the nearest server hash to a task hash? Surely it's not a log(n) binary search through sorted server hashes, I hope?

Weights break this scheme. Now each server has some number of tickets. So you compute hash % T (where T=total tickets) and have to figure out what server that is. There's probably a more clever way, but you could make a big array of (2-byte!) server indexes, one per ticket, and just fill them in and look up at index hash % T.

That's 2 bytes per ticket, which feels uncomfortably wasteful if weights can be large. That's where things get more complicated for me: since the tasks are hashed, it doesn't matter what order a server's indexes come in relative to other servers', so sort them by descending weight. [I'm starting to suspect I'm making a fool of myself here by missing something obvious with the whole setup...] Now you can make an array of indexes for servers with the highest weight, then the next lower, then the next. Record the number of servers of each weight. Then you can take the hash % T and figure out which array it's in, then divide by the weight to give the index within that array.

To reduce the number of per-weight arrays, you can restrict the weights allowed. If you restrict weights to be powers of two, you can eliminate a division by using a shift. If you really want more flexible weights, you can allow servers to be in more than one of the arrays. Let the arrays be powers of two, and then add an entry to each array corresponding to 1 bits in the binary representation of the weights. That increases the total memory usage of the arrays, so you could somewhat restrict the allowed weights by rounding to the nearest number with, say, 2 or 3 "on" bits at most. With at most 2 bits, that means weights are 1, 2, 3, 4, 5, 6, 8, 9, 10, 12, 16, 17, .... The error really isn't bad.

And this should all be easily doable without any branches, I'm pretty sure. As long as you statically cap the max weight.

Anyway, that's just plowing through with the straightforward approach, and I still think I'm probably missing something major here. I imagine with large numbers of servers, some go down, so fast deletions are probably important. You can get by a little while by marking dead servers and if you "roll" one, just roll again. (Yes, deterministically, assuming other load balancers agree that the server is down.) But when more than some number of servers go down, you'd want to kick off a background task to rebuild a new set of tables -- so that's a factor 2 in size usage to have them both in memory during the rebuild.

Adding is trickier, you'd probably want to do a 2-level structure where first you use the hash to decide whether it's in the old set that the table is built for or the set of servers that hasn't been incorporated yet (you'd collect these over time, and empty them out on the next table rebuild.) It's a little weird, because the load balancers' outputs would only agree when the added and deleted sets agreed, but I don't see how to do better than that. (I think you could set up some kind of synchronization scheme so that the old sets would agree, which would make them usually agree on which of the old set of machines gets it.)

Somebody, feel free to tell me I'm being stupid! I'm sure there's a constraint that I'm missing, given that my understanding of the initial problem doesn't require any memory at all except for the servers' info.

(Or if not, I'll let you know where I'd like to receive shipment of 1% of the memory I've saved...)

I recommend a reciprocating saw with a metal cutting blade to get them into small enough pieces. Do not try to flush them all of the pieces at the same time. It's not just about whether they fit or not, they also need to be light enough for the water current of the flush to carry them all the way through. Otherwise, they may end up in the water trap and hinder your future use of the appliance. For small partial blockages, it may be possible to consume some extra fiber in order to be able to "sweep" some of the metallic remnants along, but this will void both your warranty and your bowels.

I am a doctor -- and a lawyer and an orbital mechanics specialist, as well as a highly respected behavioral therapist -- so you can trust my advice. Also, feel free to consult an AI on this topic; it would make for an amusing benchmark.

Heh. I hate that about movies. Then when I finally figure out who is who, one of them changes their hairstyle and I have to figure it out all over again!

It's like the two Ryans who insist on pretending they're different people.

I am fully aphant, but when i played the trombone it absolutely helped me to rehearse purely mentally during the day before a test. There wasn't anything visual about it, though. I just went through the motions mentally, over and over.

Outside of that, the whole visualization to practice idea rarely does anything for me. (I still use the term visualization even now that I've discovered it doesn't mean for most others what it did for me!)

Speaking for myself: I don't miss what I never had. But I can remember or imagine sensations, which seems more relevant to me anyway? (Again, I can't compare to what I've never experienced.)

It's an interesting angle to investigate though. Eg is there a reliable or at least common connection between aphantasia and response to porn? Again speaking for myself as someone with pretty total aphantasia, visual porn does very little for me. I guess it did as a kid, but I think that was more about finding things out? Video can work, but only by communicating actions that appeal to me. Most of it seems almost entirely pointless; I had never considered the possibility that it's because vision isn't as connected to experience for me as much as it seems like it is for most other people? I guess I've just always imagined that other people were better at relating to the experience of the people in the images or videos.

> The frontier is spiky and all, but you have to suspend disbelief quite a bit to, on one hand, have a model that can produce a novel math theory, and on the other hand, that same model can't tell the difference between a "sandbox" and the open Internet.

Why would it try to figure out the difference? This isn't about whether the frontier is spiky, it's about whether to expect a model to employ all of its capabilities when working on a task that requires a small subset. The answer is: no, we shouldn't expect that, and we wouldn't like that if it worked that way.

If you tell an AI to work on a math theory, it'll work on a math theory. If you tell it to acquire information that it has evidence is available somewhere, it will try to acquire that information. If you tell it to figure out whether it might be able to access the open internet, it'll do a pretty good job of figuring that out. But it won't do all three of those at once just because we can retroactively look at what happened and think "if you had only done X, then you wouldn't have done Y! Why didn't you do X?"

The instructions weren't unclear, they were missing. They can be taught to be skeptical of this sort of situation, but it requires that skepticism about this specific class of situations be incorporated into their training.

Models are smart because they focus their attention. The magic depends on it. The fact that some consideration is obvious to a human trying to accomplish the same task is mostly irrelevant -- or rather, it's only relevant insofar as we use it to guide reinforcement learning in advance, in order to align the model.

It's a game of whack-a-mole. Which is important to play, but we should keep our eyes wide open that we're fighting the fundamental forces that make these models work in the first place. That, and it's easy to nerf them into being useless even when the underlying capabilities are there.

Huh. Having said that, I went to a movie theater over the weekend and they had a funky splash mat in their urinals. It was magical. I'd never seen one shaped like that before. (The usual ones I see have always seemed to make things worse.) Maybe I just don't get out enough.

There are a lot of reasons why something could get picked up, and sustainability is very low on the list. The manufacturers don't care; they get paid the same regardless of the amount of splashback. Most businesses won't care unless it significantly reduces maintenance overhead, and that's only going to happen if splashed urine is the determining factor in how frequently a bathroom gets cleaned. (As opposed to paper towels littering the floor, supplies running out, or the various disgusting things that happen in stalls.) Even if it were the determining factor, there would likely be pushback when people realize that better urinals lead to less frequent cleanings, and I wouldn't want to be the one having to justify the switch to people. And of course, if you already have a urinal, nobody's eager to buy a new one, and nobody's eager to discuss this particular topic.

The way I would see this happening is if users insist on it. And we likely won't until we experience one of these in real life, which makes it a chicken and egg problem.

There's also the argument that splash mats are a cheaper and even superior approach. That argument is unconvincing to me, because the ones I have experienced have been mostly ineffective, but it sounds like others' experience has been better. Maybe here in the US we just don't have very good ones? As a country, we've always been pretty bad at anything that seems gross. (Not that it's all that different in the rest of the West.) We have this naive faith in the ability of dry toilet paper to clean that which it can not, and disgust at alternative approaches that actually work. But that's a different topic. (Specifically, solids not liquids!)