Think like how so many businesses will opt for laws that hurt competitors more than themselves rather than laws that benefit them but benefit competitors even more so.

Unlike life which would have such behavior selected for by evolutionary pressures, AI would be more likely to pick it up from human literature on things like game theory, though why it even cares it survives or not is even more difficult to explain. Maybe a default bias also picked up from humans? I find it hard to see how AI training would create an evolutionary pressure that produces such a drive.

That feels a bit like a lie. At the core, they are deterministic. We found that adding some ability to randomly pick the second or third best tokens made for better output, so we added temperature. And then we started running them in optimized ways where your answer is deterministic only if the batch of tokens are the same (not your input tokens, but other tokens in another batch being processed), and in practice those are never the same. Lastly, we use harnesses that do things like adding IDs and timestamps to the context, which means the same exact text from the user does not lead to the same text hitting the AI.

The final result is that, in practice, you are right (unless you run a model fully locally, where you can seed temperature and turn off all these other features). But strictly calling it non-deterministic makes it sound like the underlying algorithm is itself non-deterministic (and I've seen many people with that misunderstanding) rather than it being a result of how we purposefully changed the algorithm for better results.

A bit like saying path finding is non-deterministic, because having the best pathfinding makes for poor gameplay, so we added some randomness to NPC path finding to make it more realistic. The given implementation is non-deterministic, but the underlying algorithm isn't.

In a real like political misinformation, this will have the effect of making people trust non-watermarked images more, which will then be used by foreign actors to pass off propaganda as legitimate.

Also, if you don't hold people responsible for spreading an image they know is fake, bad actors can take advantage of this even within the US (they purposefully spread a image they have reason to think is fake but lacking a watermark), but holding people responsible for a strict liability crime for spreading AI without knowing it is AI seems an even worse route.

I'm not sure a law even makes the issue better in a 'don't let perfect be the enemy of good' sort of way.

A small fraction of them is saved in the model. Far more is saved in the digitized copy as long as they keep it which they have plenty of incentives to do so (future training of newer models).

That's more than what happens if that book was burned or sent to a landfill, but less than if the book is giving a loving home.

>They removed the pages from the binding, scanned them on a high speed conveyor belt scanner which yielded full color 600 DPI jp2 images, placed the pages back in the binding like a folio that could be re-bound if needed, vacuum sealed them, and stored them in a salt mine.

My understanding is that this simply isn't legally allowed for these books. The original must be destroyed for the digital copy to not be copyright infringement.

>That’s a false dichotomy.

I pointed out there is a spread of possible outcomes and that different people are considering different outcomes and the comparison of if this is good or bad depends upon which outcome one considers. I even mention that both outcomes are sometimes right. That's about as far from a false dichotomy as I can see it.

For example, knowledge about how the book smells when you open it, something that reader do talk about enough I don't think this should be a strawman, is lost. But, that is about the experience of reading the book, not the knowledge of the book.

The exact text? Yeah, I think that is largely lost as well. This is a summary. And for rarer books, it will be a particularly bad summary. The basics of the book are being captured in a space that the right question that retrieve it, but worse than a sparksnote and any well read reader will tell you all the sorts of things a sparknotes already loses compared to reading the book directly.

But, that little bit of data is a bit more data than existed before, and future LLMs should get better at giving the information. So, in that sense, the knowledge is better being spread compared to copyright where the book stays in a warehouse until it is disposed of. If it was between this and a sparksnote of the book being made, the sparksnote is far better, but between this and the book simply being disposed of, then the LLM is better but far, far from great.

That's a lot of assumptions that goes into the judgment, which is probably why different people reach different conclusions. One person is imagine the alternate fate of this book being slowly rotting in a landfill, the other resting on a bookshelf where it is read at least once fully and then flipped through time to time, and neither are wrong.

As for long term benefits, it could one day be sold as a service, once copyrights have expired on the works. We can't see it today, but that is purely the result of the law and what the law intended to do from the start, you don't see a copy unless you pay for your own.

Aren't the copyright laws forcing them to do this the very ones that would make such archives illegal? The books that could be in such an archive are the books that don't need to be destroyed.

I've sees some chat logs where I completely empathized with the AI trying to make sense of the information it was being drip fed. Always incomplete, often incorrect.

Thus, the first well accepted study anchors a value and it takes much more work to get a new value established as correct after the anchor is set. Especially in a case where the experiment design itself is still solid and it was more with specific experiment (measurements slightly off, other values not quite correct).

I do wonder if the idea of TDD has influenced AIs to be too prone to testing even when a human would never consider it. I've seen some really silly tests, especially when it starts writing tests for things that are setup to only allow testing. Normally a comment later and it agrees it was pointless, but unless you have something in context to force it, it simply defaults to "test all the things".

When I'm writing technical documentation, it keeps the explanations in. Same when writing non-technical documentation. When I was having it attempt to generate a Pathfinder 1e class for a Sword Dancer, it was leaving in notes about why it removes things I told it to remove/rework.

And it isn't just Claude. I've seen the same with GPT models, with Grok, with Deepseek. Each AI isn't quite the same with how it approaches this, but in every case they seem to have a strong bias to retaining information, even bad information that we want gone, so it is like they have a, dare I say, subconscious bias to retain the information. Putting a note in a comment or explaining why to not do something or something was undone is a good way to retain information while still achieving the goal (well, if you ignore the part about the human intention for the information to be gone).

This then weakens the AI in the future, as I find AI struggles with the more incorrect information. Sure, a comment saying "not X because Y" is less 'context damage' than a comment saying "X" (assuming X is wrong), but it is still a slight shift to X being present in context in some way. One off, AI's seem to perfectly handle this without issue. But after hundreds or thousands of cases build up? The attention mechanism seems unable to keep up and incorrect information flows it. This effectively creates a sort of vibe coding maximum size unless there is a human janitor cleaning up the bad information on the context stays nice and clean.

But this is all simply a feeling I get as I use AI to do different things and isn't at all backed up by any formal study.

There are many studies, but each can be wrong and they can collectively show a bias. Even things of which are the most non-political of facts can have very strong biases. Look at the Millikan measurement of the electron and how in created confirmation bias and an anchoring effect on some property that has absolutely no real world significance to the things people are tribalistic about (aka, no political relevance). Now imagine the same applied to fields like economics or psychology which do have massive legal/political implications.

For a different example, ask the question if X committed crime Y. There are cases where they weren't found guilty but it is reasonable to assume they did. But being found guilty doesn't make it a fact either, as some people are wrongly convicted. Some eventually are overturned, but even if it isn't, it still isn't a fact they committed a crime.

Then there is the simple ambiguity of statements. Language generally can't support facts. It is why legalize, and programming code, and math's are effectively their own languages. For a simple example, consider the Betrand paradox(1).

>Consider an equilateral triangle that is inscribed in a circle. Suppose a chord of the circle is chosen at random. What is the probability that the chord is longer than a side of the triangle?

Is the answer 1/2, 1/3, or 1/4? Well, it is all three at once, depending upon what you meant by random. Now, imagine how this impacts things like research studies, where the randomness is much harder to quantify and there is constant pressure to p hack a result.

1. https://en.wikipedia.org/wiki/Bertrand_paradox_(probability)

In cases where one token is extremely likely, it'll randomly be red or green and still be picked in either case as it is simply the best (or only) option. So you'll have more tokens that don't show a pattern either way (half of these cases will match and half won't, just the same as if a human wrote it). Meaning you'll need more instances where multiple tokens were all likely to see if there is a pattern. Given the check algorithm can't identify these cases, it can only judge on the overall text, so the more strict a language, the more the length requirement scales.

Where I wonder if this keeps working is in tool calls. Often, you don't take code straight from the llm, you take the results of a tool call to edit already existing code. It might be that the result of this leads to far too few signals to pick up, meaning that this only works when one does significant generation with a single model (even swapping between different models, at least by different companies, breaks this just as much as having a human write parts of the code).

Think of it like finding a loaded dice. A dice that has a slight bias in a few dozen roles is just random chance. If that bias continues after hundreds of thousands of roles, the dice is loaded. But will a code base have enough samples, especially when edits made from tool calls? I could see this being unable to detect things at the size of a reasonable PR and only being useful for massive sets of changes and only if the person behind them didn't structure their AI usage to avoid detection.

You then look at the tokens actually picked to see how closely they follow this pattern that isn't connected to the meaning of the tokens. With enough text, you can then analyze the chance of it happening by chance verses being because the generation of the tokens was done using the algorithm, and you can save a positive result until you are arbitrarily sure. There is a chance of a false positive, but the chance of a false positive approaches the chance that the murderer happened to have fingerprints that matched your and both forensics labs happened to have mixed up the dna tests and the eye witness happened to misremember the face and your phone gps happened to glitch out and put you at the murder scene at the time of the crime all happening. It is theoretically possible only in the same sense that quantum teleporting a cat is theoretically possible.

The real question is how much text do they need for a given level of certainty and what do they check for. If they flag a positive at a p value <.01, that's a problem. If they can reasonably get a p value of < 1e-12 in only a few paragraphs of text, that is effectively no false positives (but a lot of 'too short to analyze' outcomes).

One interesting (albeit sad) area which might be related are humans who are never raised with a first language. They seem to never developer abstract reasoning and even seem to lose the ability to develop it later in life. This might indicate there is some 'real world senses' -> 'direct language' -> 'indirect language' -> 'abstract abduction' hierarchy that develops, perhaps related to more real world abductions as a necessary side chain to developing abstract ones.

One of the obvious problems with this is just how difficult we find it to study intelligence purely in humans. We are measure a LLMs by a yardstick that is already known broken, but maybe this is still the right path.

I'm a major AI user. Coding? Sure. But language learning, some art projects, TTRPG design. I'm so far 'lost' that I have a number of AI songs on my YouTube play list. For the anti-AI people, I'm like a walking abomination.

Yet despite all this, I still am a firm believer in "if you couldn't be bothered to write it, I can't be bothered to read it" as a default stance. I feel human to human communication should remain human to human. If AI is involved, the extent should be limited to where the human to human communication couldn't be achieved otherwise and fully disclosed.

Someone might even say I'm a bit of a hypocrite. Why do I like listening to an AI song but not reading an AI blog. Honestly, I'm not sure. I'm actually a bit puzzled by it myself. Regardless of the why, I think I should be an example of someone who holds this position for reasons entirely different than those you suggest.

I don't even disagree with you in general, a lot of AI discourse does seem to have bad faith arguments put forward and I do tire of some of them. But this specific one is one of the more few that seem good faith enough we should steelman it.

Isn't that more due to an optimization and not how the LLM itself runs?

Like a MoE LLM run on a single input should give the same output each time. But this is inefficient, as any given token is hitting 1 (or maybe 2 or 3) experts at a time, meaning all the other experts are doing absolutely nothing. So you upgrade it to take in multiple requests. But then any given expert can become a bottleneck, so when too many requests need a given expert, some of them are routed to a second or third best expert instead. Within the context of any single request, this looks like non-determinism, but it is still deterministic when considering the full batch.

For everyday users and everyday use cases, that is enough to treat it as non-deterministic (the harness might also send in unique data like current time which means one can never have the exact same request twice), but when talking about LLMs more theoretically, I think we need to consider they can still be ran deterministically even if that isn't as optimized.

Similar with temperature. 0 means deterministic, but anything higher with a seeded value is deterministic. If anything, temperature is us purposefully adding non-determinism to agents because they were too deterministic.

Even if some individual case can be shown to be safe from being combined, can we identify such cases with enough confidence to justify using it reduce severity warnings?