It doesn't take a lot of compute resources to have multiple profiles, and when evidence of a high assurance level (a referring URL that is known to designate a specific user of a major service) to link it with other profiles that also have that designation.

To me, that seems par for the course for any service that's generating profiles of browsing behavior and trying to make any sort of decisions based on it. It reduces cruft and duplicate profiles while also providing more accurate information. Why wouldn't it be done?

> the information-theoretic validity of your argument

The portion about canvas, WebGL and AudtioContext hashing is not theory at all, it's well known practice from years ago. Jest the other day here there was a story about some advertiser on Stack Overflow trying to use the audio hashing to tracking purposes.

Hell, if you get enough identifiable bits of entropy, you can probably assume weak to strong level matching using a bit-level Levenshtein distance that's low enough.

GitHub is always at your disposal. NV doesn’t sell the consumer cards to enterprises. So on AWS a multi-GPU box will cost you about 12 dollars an hour. If you can disambiguate, let’s just say 85% of profiles absent IP or cookies, well I think you just broke the academic SOTA and I’d love to make some calls.

Cheat sheet: you can’t.

I sorta understand the sentiment where there are tools like :

https://amiunique.org/ https://browserprint.info/

whose results would have you believe that one's footprint is very unique. I'd be interested in hearing more about why this is hard to implement into an efficient process.

> GitHub is always at your disposal. NV doesn’t sell the consumer cards to enterprises. So on AWS a multi-GPU box will cost you about 12 dollars an hour.

I don’t see how this is related to the claim, since it doesn’t solve the problem. But the advertising company that I let run code on my website will certainly do the job pretty well, I’d say.

I was pointing out that it’s a commercially applicable of a very strongly worded claim that I know would be expensive to test because I’m optimizing GPU intensive code at the moment. I don’t know where in this thread I generated so much ill will for trying to add knowledge to the conversation, but I’m not making shit up.

There are tools that will supposedly do this to a high degree of accuracy. Are you saying that they are fake/don't work as well as they'd want us to believe?

not much effort goes into de-duping customers? genuinely curious

I would've thought that would be a pretty useful exercise