I think the above is a good idea of what's wrong with python (and Go), because in your example the list comp is evaluated in what seems to be this order:
FOURTH-> print([THIRD-> v*v for v in FIRST-> reversed(a) if SECOND-> v*v % 2 == 0])
Which is all over the place. I'd rather see:
a = [1, 2, 3, 4]
a = reversed(a)
a = [v*v for v in a]
a = [w for w in a if a % 2 == 0]
print(a)
I often use generator expressions for the intermediate values (so I don't allocate a new list for each step), but I find this to be much more readable.
Having read the paper, I agree that this is an enormous effort, but I didn't see anything that was particularly surprising from a technical point of view - and nothing of Singularity-level significance. The use of AI to train AI - as a source of synthetic data, or as an evaluation tool - is absolutely widespread. You will find similar examples in almost any AI paper dealing with a system of comparable scale.
Yeah I know, but you sometimes see posts on HN that talk as if AI isn't already being used for self-improvement. I guess the subtlety is that people tend to imagine some sort of generic recursive self-improvement, and overlook the more tightly focused ways it's being used.
Stenography is writing in shorthand. What you mean is steganography.
You can also watermark plain text by generating "invisible" patterns.
Of course, in all these cases, the watermarks are trivial to remove: just re-encode the output with an open model. Which is why I hope there will be no federal law that tries to enforce something that is categorically unenforceable.
If La Jetee was just some photos stitched together plus meaningful narration, then of course, you could use AI-generated photos.
But would AI be able to quote Vertigo, like La Jetee does? Doesn't art, at least to some degree, require intent (including all intentional subversions of that intent dogma, of course)?
> In the short term, it's a smart strategy for Israel, but they've likely opened Pandora's box in the process.
Absolutely. From today on, this type of attack will have to be considered part of the arsenal.
While building thousands of explosive communication devices and swapping a large shipment requires substantial resources and intelligence, the actual "sophistication" of today's attack seems to lie in the fact that the perpetrator managed to specifically target a clandestine adversarial organization.
If you don't care about that last part, I don't think it's completely out of reach for a hypothetical "state sponsored terrorist organization" to have a thousand smart phone explosives shipped into a target market, say the European Union or the United States. Such an attack, if successful, would be devastating.
> Absolutely. From today on, this type of attack will have to be considered part of the arsenal.
I doubt that. It’s way more likely that from today on, every batch of devices used by VIPs will be tested for explosives, thus rendering such attacks moot. Finding such devices would also enable figuring out who your enemy’s operatives are.
Nothing, they aren’t looking for 2”x1” sheets of copper within electronic devices, and presumably the thin layer of explosives would be sealed and washed.
I meant for inline strings (i.e., typed by the programmer). This is an inoffensive change. The only possible ``accident'' is when the programmer wants to write "{x}" instead of the value of x. This is such and exceptional case that it may be best treated by forcing to escape the curly brackets.
If anything, user-input strings must be treated as tainted whatever the case.
Then you would run into the same problem Kotlin has with its raw strings – the brackets need to be escaped, but raw strings disable escaping, so you'd have to write r"{}" as r"{'\{\}'}" which is quite ugly and not very raw at all.
That would break any existing strings that use curly braces, including most legacy use of str.format and docstrings that include code examples with dictionaries. It would be a big backward compatibility issue.
Also: if you don’t have high precision timers, you probably could also go and increase the number of guesses to cancel out noise. After all, this is a statistical side channel attack. This will be fun for quite some time.
Doing that is harder than usual, because the time before rollback is very limited, reducing the amount of cache lines you can pull into the cache and you can measure a given cache line only once.
The attack is already pretty damn slow (1kB/s) in C, if you have to collect a statistically significant sample in JS it might well slow down enough to not really work properly.
> which just shows how little moat these companies have
Flux was developed by the same people that made Stable Diffusion.