"AI art generators enable the creation of ignorant and lazy illustrations by outsourcing understanding to an idiot robot."
"Yes, but is it not the intent of the artist to be ignorant and lazy?"
It is possible to repeatedly iterate AI art gen and get what you want, but that's not what happened here. And even so, it's not at all the same thing as drawing a picture: "iterating on what you want" is equivalent to curating art, not creating it. In the US you can copyright curation and that extends to curation of AI art - the US Copyright Office correctly said that tweaking prompts is the same thing as tweaking a Google Images search string for online image curation. But you can't copyright the actual AI-gen pictures, they are automatically public domain (unless they infringe someone else's copyright).
I am specifically talking about DALL-E or Stable Diffusion, your link describes something very different. The point was the "Google Images" analogy, which applies to 99.999% of AI art but this is an exception.
> I am specifically talking about DALL-E or Stable Diffusion, your link describes something very different.
No, it doesn't. It describes artwork done on Invokeai, one of the popular hosted web frontends for Stable Diffusion (and some similar models), with a process very much like what many AI art hobbyists use (whether with hosted frontends or locally-hostable ones like A1111, Forge, ComfyUI, or Fooocus.)
I don't understand your ridiculous pedantry! I am talking about DALL-E and Stable Diffusion. I am not talking about other front ends to these services, nor did I dispute that your example deserved copyright protection. Invoke is very very different from plain text-to-image generation, WHICH IS WHAT I WAS TALKING ABOUT.
I think it's best if I log off and ignore your replies.
> I am talking about DALL-E and Stable Diffusion. I am not talking about other front ends to these services
Stable Diffusion is a series of models, not a service. There are various services (including the first-party Stable Assistant service from StabilityAI) and self-hosted frontends that use the models, most of which (including Stable Assistant) support considerably more interaction than simple text-to-image.
See the other reply for a half-counterexample, but the major difference is the specific software is more like generative PhotoShop, and the final image involved a lot of manual human work. Simply tweaking a prompt is not enough - again you can get copyright for curation, just not the images."
Of course AI can't be credited with copyright - neither can a random-character generator, even if it monkeys its way into a masterpiece. You need legal standing to sue or be sued in order to hold copyright.
I am unable to think of a single animal behavior popularly ascribed to intelligence but in fact explainable by rote instinct. Do you have an example?
OTOH there are plenty of animal behaviors that can only be explained by intelligence: seeing-eye dogs perform a task which is economically useful for humans yet far beyond the ability of AI (even if the robotics mech eng issues are resolved). But it also doesn't really make sense that one of my cats is "instinctually" able to understanding my words while the other cat is "instinctually" able to outsmart me when we play with toys. The more sensible explanation is cats are intelligent and intellectually diverse.
The problem is that it also included "doxxing" DOGE employees, even though they are significant public figures and their anonymity is almost certainly illegal. Protecting the identity of government employees who are more powerful than cabinet secretaries is plain abuse of ToS. (I will add that I don't believe most of the alleged death threats were real. I think Musk's concern was the "doxxing" and the death threats were a pretext.)
Specifically, better scores on the measures of impulsivity, emotional awareness and flexibility, and also the personality trait of agreeableness were all linked to higher patience scores.... But these results do suggest that patience is not so much virtue as a method to help us to deal with frustrations — and that some of us are better equipped to employ this coping mechanism than others.
The flaw in this argument is that non-impulsiveness, emotional awareness, and agreeableness are all considered virtues! And I really don’t like the implicit suggestion that impatient people are just born that way and patient people won the luck of the draw: I was far more impatient before I started getting mental health treatment.
I struggled reading the papers - Anthropic’s white papers reminds me of Stephen Wolfram, where it’s a huge pile of suggestive empirical evidence, but the claims are extremely vague - no definitions, just vibes - the empirical evidence seems selectively curated, and there’s not much effort spent building a coherent general theory.
Worse is the impression that they are begging the question. The rhyming example was especially unconvincing since they didn’t rule out the possibility that Claude activated “rabbit” simply because it wrote a line that said “carrot”; later Anthropic claimed Claude was able to “plan” when the concept “rabbit” was replaced by “green,” but the poem fails to rhyme because Claude arbitrarily threw in the word “green”! What exactly was the plan? It looks like Claude just hastily autocompleted. And Anthropic made zero effort to reproduce this experiment, so how do we know it’s a general phenomenon?
I don’t think either of these papers would be published in a reputable journal. If these papers are honest, they are incomplete: they need more experiments and more rigorous methodology. Poking at a few ANN layers and making sweeping claims about the output is not honest science. But I don’t think Anthropic is being especially honest: these are pseudoacademic infomercials.
>The rhyming example was especially unconvincing since they didn’t rule out the possibility that Claude activated “rabbit” simply because it wrote a line that said “carrot”
I'm honestly confused at what you're getting at here. It doesn't matter why Claude chose rabbit to plan around and in fact likely did do so because of carrot, the point is that it thought about it beforehand. The rabbit concept is present as the model is about to write the first word of the second line even though the word rabbit won't come into play till the end of the line.
>later Anthropic claimed Claude was able to “plan” when the concept “rabbit” was replaced by “green,” but the poem fails to rhyme because Claude arbitrarily threw in the word “green”!
It's not supposed to rhyme. That's the point. They forced Claude to plan around a line ender that doesn't rhyme and it did. Claude didn't choose the word green, anthropic replaced the concept it was thinking ahead about with green and saw that the line changed accordingly.
> Here, we modified the part of Claude’s internal state that represented the "rabbit" concept. When we subtract out the "rabbit" part, and have Claude continue the line, it writes a new one ending in "habit", another sensible completion. We can also inject the concept of "green" at that point, causing Claude to write a sensible (but no-longer rhyming) line which ends in "green". This demonstrates both planning ability and adaptive flexibility—Claude can modify its approach when the intended outcome changes.
This all seems explainable via shallow next-token prediction. Why is it that subtracting the concept means the system can adapt and create a new rhyme instead of forgetting about the -bit rhyme, but overriding it with green means the system cannot adapt? Why didn't it say "green habit" or something? It seems like Anthropic is having it both ways: Claude continued to rhyme after deleting the concept, which demonstrates planning, but also Claude coherently filled in the "green" line despite it not rhyming, which...also demonstrates planning? Either that concept is "last word" or it's not! There is a tension that does not seem coherent to me, but maybe if they had n=2 instead of n=1 examples I would have a clearer idea of what they mean. As it stands it feels arbitrary and post hoc. More generally, they failed to rule out (or even consider!) that well-tuned-but-dumb next-token prediction explains this behavior.
>Why is it that subtracting the concept means the system can adapt and create a new rhyme instead of forgetting about the -bit rhyme,
Again, the model has the first line in context and is then asked to write the second line. It is at the start of the second line that the concept they are talking about is 'born'. The point is to demonstrate that Claude thinks about what word the 2nd line should end with and starts predicting the line based on that.
It doesn't forget about the -bit rhyme because that doesn't make any sense, the first line ends with it and you just asked it to write the 2nd line. At this point the model is still choosing what word to end the second line in (even though rabbit has been suppressed) so of course it still thinks about a word that rhymes with the end of the first line.
The 'green' but is different because this time, Anthropic isn't just suppressing one option and letting the model choose from anything else, it's directly hijacking the first choice and forcing that to be something else. Claude didn't choose green, Anthropic did. That it still predicted a sensible line is to demonstrate that this concept they just hijacked is indeed responsible for determining how that line plays out.
>More generally, they failed to rule out (or even consider!) that well-tuned-but-dumb next-token prediction explains this behavior.
They didn't rule out anything. You just didn't understand what they were saying.
>They forced Claude to plan around a line ender that doesn't rhyme and it did. Claude didn't choose the word green, anthropic replaced the concept it was thinking ahead about with green and saw that the line changed accordingly.
I think the confusion here is from the extremely loaded word "concept" which doesn't really make sense here. At best, you can say that Claude planned that the next line would end with the word rabbit and that by replacing the internal representation of that word with another word lead the model to change.
I wonder how many more years will pass, and how many more papers will Anthropic have to release, before people realize that yes, LLMs model concepts directly, separately from words used to name those concepts. This has been apparent for years now.
And at least in the case discussed here, this is even shown in the diagrams in the submission.
We'll all be living in a Dyson swarm around the sun as the AI eats the solar system around us and people will still be confident that it doesn't really think at all.
Agreed. They’ve discovered something, that’s for sure, but calling it “the language of thought” without concrete evidence is definitely begging the question.
Came here to say this. Their paper reels of wishful thinking and labeling things in terms they prefer it would be. They even note in one place their replacement model has a 50% accuracy which is simply a fancy way to say the model's result is completely by chance, and it could be interpreted one way or another. Like throwing a coin.
In reality all that's happening is drawing samples on the joint probability of the tokens in the context window. That's what the model is designed to do, trained to do - and that's exactly what it does. More precisely that is what the algorithm does, using the model weights, the input ("prompt", tokenized) and the previously generated output, one token at a time. Unless the algorithm is started (by a human, ultimately), nothing happens. Note how entirely different that is to any living being that actually thinks.
All interpretation above and beyond that is speculative and all intelligence found is entirely human.
tangent: this is the second time today I've seen an HN commenter use "begging the question" with its original meaning. I'm sorry to distract with a non-helpful reply, it's just I can't remember the last time I've seen that phrase in the wild to refer to a logical fallacy — even begsthequestion.info [0] has given up the fight.
(I don't mind language evolving over time, but I also think we need to save the precious few phrases we have for describing logical fallacies)
Even in self-driving, Telsa's behavior proves there is a market for cars that are programmed to speed and roll through stop signs. Waymos are safer than the average human, but the average human also intentionally chooses a strategy that trades risk for speed. Indeed, Waymo trips on average take about 2x as long as Ubers: https://futurism.com/the-byte/waymo-expensive-slower-taxis
What happens if an upstart self-driving competitor promises human-level ETAs? Is a speeding Waymo safer than a speeding human?
I can imagine that wide adoption of AVs could increase the speed limits (at least for them if not all vehicles). Also I believe that high saturation of AVs will finally shorten ETAs naturally as stable, predictable driving without cutting in, forcing others to stop suddenly etc. reduces traffic jams. (Can't find the study that opened my eyes on that right now).
> a country which is unilingual English by policy?
To be clear the US only has a unilingual policy because Trump signed an executive order this year (and I believe even this SCOTUS would strike the order down as unconstitutional if anyone had standing to sue over it).
The US has always been de facto unilingual, but de jure we don't have an official language since Trump has no legal authority to establish that. The "policy" is political and legally empty.
The former de facto status of the US unilingual nature, even with its soft acceptance of Spanish here and there, is so far from Canada's official bilingualism it's not even in the realm of comparison.
And not anywhere close to what would be tolerable for Canada's francophone population, and especially Quebec which would simply immediately begin separation.
I am not disputing any of that, nor am I trying to put a positive spin on anything coming out of Trump. My only point was that it's misinformation to say that the US is unilingual by policy - that only became partially true in March 2025 via a toothless (and blatantly unconstitutional) executive order.
And it's not "soft acceptance of Spanish here and there," all levels of government are legally required to print official documents in whatever languages their community speaks; cities usually have ballots in dozens of languages. This is a constitutional requirement, bolstered by the Voting Rights Act, and Trump has not yet done enough damage to make those legal requirements go away via diktat.
You genuinely do not understand the level of personal freedom we have restricted in Québec to protect our language. As a francophone, I am forbidden by law to send my children to English school. Any public signage MUST have French predominantly displayed. Any company of a certain size MUST allow its workers to work in French. Any immigrant can receive government services for six months in a language other than French.
We are not kidding when we say that the American way of life and the Quebec one are incompatible. But this is all moot, the Canadian constitution would require a referendum to decapitate Canada, and that would never get the 50% vote per province that is required. Right now they'd get at most 10%.
All federally managed things & services, including travel on airlines, airporst etc, fully bilingual. All packaging of goods, across the whole country. Access to legal services, no matter where, bilingual. Highway signs, etc. The list goes on.
Here in Ontario, at least, right to full public school education in French, in the French system, if you come from a Francophone family.
The idea being that even in areas of the country that are not Quebecois or predominantly Francophone there are rights granted because of the French being part of the founding of the nation.
This would be a valid POV if there was any solid evidence that LLMs truly increased worker productivity or reliability - at best it is a mixed bag. To stretch the food analogy, it seems like LLMs could be pure corn syrup, without any disease-resistant fruits and unnaturally plump chickens that actually make modern agriculture worthwhile.
Or, since LLMs seem to be addictive, it's like getting rid of the spinach farms and replacing them with opium poppies. (I really hate this tech.)
That's 100% false, dogs and pigeons can obviously think, and it is childish to suppose that their thoughts are a sequence of woofs or coos. Trying to make an AI that thinks like a human without being able to think like a chimpanzee gives you reasoning LLMs that can spit out proofs in algebraic topology, yet still struggle with out-of-distribution counting problems which frogs and fish can solve.
Maybe a roundabout answer to your question, but Peano's axioms are equiconsistent with many finite set theories (even ZFC without axiom of infinity), and I do think philosophically it makes more sense to say weak axiomatic set theory + predicate calculus forms building blocks of arithmetic[1]. The idea of "number" as conceived by Frege is an equivalence class on finite sets: A ~ B <-> there is a bijection, which is in fact a good way of explaining "counting with fingers" as an especially primitive building block of arithmetic:
{index, middle, ring} ~
{apple, other apple, other other apple} ~
{1, 2, 3}
as representatives of the class "3" etc etc, predicates would be "don't include overripe apples when you count" etc. Then additions are unions and so on, and the Peano axioms are a consequence.
[1] In my view Peano axioms are the Platonic ideal of arithmetic, after the cruft of bijections and whatnot are tossed away. I agree this is splitting hairs.
"Yes, but is it not the intent of the artist to be ignorant and lazy?"
It is possible to repeatedly iterate AI art gen and get what you want, but that's not what happened here. And even so, it's not at all the same thing as drawing a picture: "iterating on what you want" is equivalent to curating art, not creating it. In the US you can copyright curation and that extends to curation of AI art - the US Copyright Office correctly said that tweaking prompts is the same thing as tweaking a Google Images search string for online image curation. But you can't copyright the actual AI-gen pictures, they are automatically public domain (unless they infringe someone else's copyright).