Thanks for the article. I hadn't heard of SynthID before and it's good to.
It's a shame however, how low quality and vibecoded the live examples are. The first example says "Toy example; not SynthID.", the second one is a generic spectrogram and the third one has an identification space too small to be useful (173 in decimal). I was hoping to see more realistic scenarios to learn how these new watermarks are being applied, instead of generic steganography.
It's of course important to reduce the impact surface before more powerful models become available, but it's worrying that some people seem to assume the only way to counteract those are by using more AI models, which despite being (apparently) good at finding bugs, they are also very good at introducing them unnoticed.
Maybe it happens that LLMs are great at finding human bugs which have some more patterned structure that's easier to match for then LLM bugs. They prefer LLM prose to human prose, so what if they're similarly wooed by slop code?
As you probably know, the computing time in the worst case scenario for brute forcing a cryptographic key generally doubles by each bit added. That is, it would take twice the effort to brute-force 129 bits compared 128 bits.
The security of RSA however depends on the factoring of very large numbers, and that means that for example, RSA-2048 doesn't translate into 2048 bits of security but 112 (roughly symmetric equivalent) bits based on the best factoring algorithms (for comparison, the RSA-512 the article mentions has an 56 bit equivalent and RSA-1024 has a 80 bit equivalent security, so RSA-1024 would take roughly 2^(80-56) ~= 17 million times to compute the worst-case scenario and RSA-2048 would take 2^(112-80) = 4.3 thousand million times more).
According to the Wikipedia article on RSA numbers, RSA-220 (66b) was factored in 2016, RSA-230 (69b) in 2018, RSA-240 (72b) in 2019 and RSA-260 (76b) this year, which is too close to RSA-1024 (80b) to be comfortable.
For RSA-250, the team reported it took "roughly 2,700 core-years, using Intel Xeon Gold 6130 CPUs at 2.1 GHz.". I am not going to (or feel qualified to) make estimates of how that would translate to RSA-1024, but it does sound plausible given enough resources.
The main counterpoint in the article seems to be screen quality, which might hold for older thinkpads. There are however a lot of models almost a decade old such as 480s or 490s with IPS screens that are very affordable second-hand. Alternatively, thinkpads are known to have a high repairability so it's relatively easy to change the screen yourself if you're not happy with it (which I understand is not something everyone would want to do).
I think we would need to balance practicability and code size since they tend to be mutually exclusive. Generally speaking, code size is not an important metric to make useful code, and usefulness is usually not the main point of code golf exercises such as this one.
The exception to this is obviously embedded systems with very low amounts of resources where C and assembly are practically unrivaled.
Someone correct me if I'm wrong, but I suppose this allows google to better persist browser identifiers in order to facilitate advertising?
I'm not surprised they do this since advertising represents almost 3/4 of Google 's revenue. If anything, why should anyone expect something else? Company in the ads business modifies monopolistic browser to make money with ads and more news at 11. If it's not this, it's going to be another thing in the future.
Being surprised or not, and being pissed about it or not are orthogonal.
I'm also not surprised and do fault Google for it. We shouldn't give them a pass just because they're regular offenders, it should be the opposite IMHO.
What I am surprised about is that even with all of this effort that is put into tracking us, the majority of ads that I see don't even fit my demographic.
I think it's due to the lack of proper ads at a price point worth it. For instance your local fishing shop won't win the ad auction for your eyeballs against Binance or Tinder, even if you have absolutely no interest in the latter.
I think what made the internet in the 90s and early 2000s more enjoyable is that it was way smaller than now and (mostly) free of bots, and trying to find that in social media and mainstream sites is counterproductive, since they are just the opposite.
In comparison, I've found small net places such as Mastodon or smolnets much more enjoyable even though the population is much smaller (which is nice since they also don't generate more information than I have time for). As a rule of thumb: The less interesting a place is for companies, the more pleasant it probably is.
I don't think it speaks well of Software Engineers the will to delegate the construction of blueprints to unreliable workers. Architects and Civil Engineers know that licensed, professional people that have experience and know-how will execute it in a nearly flawless way. Compared to that, agents might hallucinate, build something completely different to the intended project or leave subtle bugs behind that, in the case of complex systems, the humans in charge will probably not even notice.
Speculation with things that affect the well-being and safety of other people should be banned, period. We have seen the impact it has had on real state or the food market to the point people have become unable to purchase new houses without indebting themselves for life.
Also reminds me of that black mirror from last season where the IT crowd guy damages his own body to entertain strangers so they give him the money he needs to support his wife. What's stopping someone poor without much to lose from burning down a forest if there's a significant monetary incentive?
Also, paywall. https://archive.ph/B5kzg