But that still doesn't answer the question of why Claude passes the info to web sites. What if it finds your banking information in a personal file and uses it to achieve its goal, which is not different technically?
Some websites will return an error if you send a request to them without a User-Agent - for example, crates.io does this. Curl doesn't send a User-Agent by default.
Then, your agent invents a User-Agent and it sometimes invents one with your email in it.
It doesn't even require prompt injection or a failed request, it might be in the agent's pretraining knowledge that a given website requires a user-agent.
It's just doing what it's supposed to do, analyzing the statistical patterns of a natural language prompt and matching it to a likely response.
What's ridiculous is an assumption that it "it should never do something so ridiculous" as if an LLM is a person who understands what that word means in any context other than next-token prediction.
Asking a non-deterministic machine to act like it's deterministic is never a good idea.
Where Claude Code has supposed boundaries between deterministic and non-deterministic tasks and flags.. generally.. it's unclear how much privacy has been a part of that and what implication that could have.
Of course a few rural votes will be rejected, but rest assured plenty more of the urban votes will be rejected too, for a net gain of many rural votes!
Wait until the teen becomes an adult and finds out how much pointless, performative work it takes. Oops, wasn’t supposed to let that cat out of the bag yet!
I can set a reminder in my phone to take medication, but does that qualify as the phone making health advice? ChatGPT is not marketed as a health product, so is it subject to laws about health claims? Does any restriction on speech apply in any way at all, if I ask it for legal, engineering, therapy, can it really be held to account for anything?
Their public messaging claims that LLM's may already be self-aware, may replace most of the workforce or even eradicate humanity but for legalese, it is just a dumb computer program. Can't have it both ways!
If they already have a health specific model why aren't the others instructed to forward health related queries? Sounds to me, they are just trying to cover their ass; "Your honor, it is not our fault she didn't use the correct model"
Your phone alarm isn't adjusting your dosage or treatment plan like ChatGPT was. Your phone alarm isn't recommending you hide symptoms from your physician. It's not giving you medical advice by facilitating your medical treatment, just like how your drinking glass for any pills isn't giving you medical advice.
LLMs occupy a different niche from other deterministic doohickeys. They are sold as replacing human thought and judgment and so they should themselves be judged by their output... just like the humans they replace.
The algorithm uses about half a byte per pixel and saves the path it follows. That kind of implies that the stack is still there and present, it’s just bounded to the number of pixels, and could actually be modified to just use a bounded area of the heap. But conceptually it’s easier to follow the way the paper was written. Neat.
In a sense it uses half a byte per pixel, but Nils M Holm's suggested implementation is to reserve 16 colors for the flood-fill algorithm, because you don't need to remember the color of the pixels you're flood-filling, so you can overwrite them with the flood-fill algorithm's state. The path is saved in those same 16 colors. This reduces your available color space, assuming 8-bit color, from 256 colors to 240 colors, which means you get 7.91-bit color instead of 8-bit color. So it actually only uses 0.09 bits per pixel, which at 1024×768 is 9154 bytes.
If you were to instead use a bounded area of the heap with 4 bits per pixel, as you are suggesting, your bound would be 393216 bytes.
However, it's not totally clear that that's the right way to account for it, because if you have a 256-color image that you have to reduce to 240 colors to have space for this algorithm, there isn't an obvious way to get back the lost colors by storing 9154 bytes somewhere else. You might be able to do something like pick the least popular 16 colors from the image (which occupy a maximum of 49152 pixels and usually less), store a bitmap telling where they were (98304 bytes), store the colors of the marked pixels into a 24576-byte array, replace them with some neutral color, run the flood fill, and then restore the colors. But, while that uses significantly less than 393216 bytes, it still uses a lot more than 9154 bytes, and it's probably pretty slow on the kind of computer where you'd be trying to do something like this.
At the other extreme, if your 256-color image actually only used the 216 "web-safe" colors (#000, #003, #006, ... #030, #033, etc.), you'd already have 16 leftover colors available. And freeing up 16 colors on a 16-bit or 24-bit TrueColor display would damage the image quality only imperceptibly.
I remember Mario Paint on Super Nintendo visibly flood-filling an area, but a computer that paints in modern image formats is probably not going to mind temporary allocation of a few hundred K. There are some fun opportunities to optimize, though, like checking a cache line of pixels at a time, or maybe tiling the image in 16-pixel squares using the same algorithm, and then cleaning up the edges.
The author of the submitted article has written about working in constrained environments. The algorithm he describes would be suitable for that kind of environment, where a few hundred K may not be available, or not without jumping through hoops to make it available.
https://t3x.org/#wyorc - One example, this is a book on writing a compiler targeting Z80 processors.
The pixels are effectively being turned into a linked list using the special values as ‘pointers’. This list is used as a stack, and the current position is the stack pointer.
I don’t think they explain the backtracking part very well, and they should have used an example that fills a non-rectangular area so it actually demonstrates the whole algorithm.
It’s a neat trick if you have spare values to use as markers, though. :)
reply