This has been known from the beginning of the field (the 1960s). The problem is that hype gets more funding, and at the beginning of each AI cycle no one knows how it turns out, so all incentives are aligned for more hype and grandiose claims.
Conversely, those who are cautious about claims are just drowned out by the hype chorus and get fewer resources. (This is not unlike a lot of the other scientific fields, but AI cycles happen very fast compared to natural and social sciences).
Latest news is that meta pirated a crapload of books to train on, and they’re not even good seeders!
“Leachers” is an adequate way to describe LLMs on the whole.
To be fair this is more of a problem with book copyright law than meta in general 
Maybe this is too much to hope for but after 10-15 years(Maybe less), books should be open digitally for everyone who wants a copy and cannot afford it. Especially non fiction.
Internet Archive’s “one viewer at once” method is good, but not enough, and they got in trouble when they allowed more than one person… quite sad.
That said, seed your books(and everything else) everyone 
Highly disagree, if you look at the reporting around it, its headlines like “meta pirates books, makes copyright cases murky”… The law is the law; doesn’t matter if the law is good or bad, meta obviously violated it, it isn’t murky at all. Copyright is problematic in general… But " should we treat meta differently because AI?" No.
I started asking a recent version of llama some programming questions at work, and man, it really just launders answers from stack overflow, including the wrong parts. Hilarious (or not).
Don’t get me wrong, I don’t think Meta should be given a pass. If the common citizen, and morally good(imo) agencies like the Internet Archive, aren’t given a pass, Meta shouldn’t either.
Meta should be prosecuted even more so than the Internet Archive and similar institutions that are nonprofit.
I guess there’s a risk that barriers, in general, might cement large incumbents with deep pockets. For instance, I heard an argument that removing platforms’ immunity from libel would help rather than hurt the dominant players by building a moat. Haven’t thought through whether that’s useful comparison for copyright law, though. Seems like big corps are already adept at extending their copyright protection beyond the statutory periods. Which isn’t to say we just have a free for all for tech companies to arbitrage away another set of regulations.
I came across this image this morning.
What do you see in the image? Did you need prompting to see it?
Without prompting, would an AI see the object in the image that you do?
This is an empirical question without a general a priori answer. You would have to check specific ML engines to see what happens.
It’s a Dalmatian, or dog at least, sniffing the ground. No prompt necessary
I would argue that it shows the difference betwen syntactic and semantic information.
Tbh, I didn’t read through the discussion above I don’t quite understand how grammar in the linguistic sense applies to photography.
I uploaded the image to Gemini. I renamed it to puzzle.jpg (I did not check if there are keywords in the metadata - update: according to the browser’s EXIF plugin, there is no metadata attached). The prompt was simply: Describe this image.
The response:
This image is a black and white, high-contrast photograph that cleverly uses negative space to reveal a Dalmatian dog.
Here’s a breakdown:
- Contrast: The image is starkly black and white, creating a strong contrast between the dark and light areas.
- Negative Space: The dog isn’t defined by black lines, but rather by the white space surrounding it. The black shapes come together to form the dog’s silhouette. This use of negative space is the key to the image’s effectiveness.
- Subject: The subject is a Dalmatian dog, clearly recognizable by its characteristic spots and slender build. It’s facing to the left and appears to be walking or standing still.
- Composition: The dog is positioned slightly off-center, with the black shapes forming a sort of abstract background or foreground. The composition emphasizes the dog’s form through the surrounding negative space.
- Style: The image has a graphic quality due to the high contrast and sharp edges. It’s reminiscent of stencil art or a Rorschach inkblot test in its ambiguity and reliance on perception.
Overall, it’s a visually striking and creative image that effectively uses the interplay of black and white space to depict a recognizable subject. The use of negative space makes it intriguing and encourages the viewer to actively engage with the image to “see” the dog.
ChatGPT:
This is an optical illusion or high-contrast image that appears abstract at first glance. It consists of black and white patches, creating a visually confusing effect. However, upon closer inspection, a Dalmatian dog can be seen sniffing the ground. The dog’s head is lowered, and its body is camouflaged by the surrounding textures. The image relies on the brain’s ability to recognize patterns and fill in missing details.
Claude (failed):
Claude: This appears to be a high-contrast black and white image with a very grainy or noisy quality. The image shows what looks like a figure or object in the center surrounded by a speckled, textured background. The stark black and white contrast and grainy quality make it somewhat difficult to make out precise details, giving it an abstract or artistic quality. The image has a rough, somewhat scattered appearance, almost like a photocopy or heavily processed photograph where much of the original detail has been reduced to basic black and white shapes.
me: Can you make out the “figure or object”?
Claude: When I look more carefully at the central part of the image, it appears to be an astronaut in a spacesuit. Given the extremely high contrast and grainy quality of the image, it looks like it might be a historical photograph, possibly from one of the early space missions, though I can’t be certain of the specific mission or context without more information.
I am beginning to warm to ML because I imagine it will save endless hours of high school students writing pointless essays about literary works. I hated these with all my heart back in the day.
The funny thing is that 90% the time I enjoyed reading the actual book that was assigned, just hated writing a something with the sole point of proving that I read the book.
The point was that you read it, understood it, and were able to apply critical thought to it.
But you are right, the point is to make us dumber and dependant on them
That image seems to be from 2011, what are the odds that someone already uploaded it and went through this exercise with the models and they learned from it, and can now discern it?
Would be fun to also try producing a similar image with gmic or similar, and then testing that one with ChatGPT and Gemini.
I tried the dog image with LeChat (What a name 
) and it failed as well.
Given the context, I find this ironic, almost cheeky:
The image relies on the brain’s ability to recognize patterns and fill in missing details.
and yet, most AI links have the caveat that they’re not always right …
I should say the odds are quite high as this image is a classic in the “Vision” chapters of cognitive psychology textbooks …
We’ll have to see. It seems models, and probably the hardware, too, are getting more efficient:
The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use. You can see this in the token cost from GPT-4 in early 2023 to GPT-4o in mid-2024, where the price per token dropped about 150x in that time period. Moore’s law changed the world at 2x every 18 months; this is unbelievably stronger.
(Three Observations - Sam Altman)
Of course he has a stake in this, so no wonder he says things like:
The socioeconomic value of linearly increasing intelligence is super-exponential in nature. A consequence of this is that we see no reason for exponentially increasing investment to stop in the near future.
