pull down to refresh
111 sats \ 4 replies \ @Cje95 OP 1 Sep \ parent \ on: AI is Unmasking ICE Officers. Can Washington Do Anything About it? AI
I'm the first one to say a gun doesn't kill anyone a person does but AI’s are an entirely different monster we haven't seen before. In theory they could set up the AI to run off of X posts and completely remove the human from the loop as it would track hashtags and cross reference with scrapped data, make a guess, and post it. That system doesn't need a human and easily allows for an AI to act alone. Its a serious issue in specific cases like this where a whole process can be automated.
In theory they could set up [..]
Unless "they" is an AI that acts autonomously, I think my point stands: someone has to code it, deploy it, pay for the servers and/or subscriptions.
However, let's consider for a moment #1090262 and then we know that it's not theoretical. 1 It's just that the capability has been asymmetrically available to megacorps and govts exclusively due to the compute cost for inference in the past.
That is different now, and it's interesting that those that have been using this to exactly identify people - without their consent or even being informed - are a bit worried by the higher parity in capabilities. Surveillance itself hasn't been exclusively available to governments, but mass surveillance was, and this has now changed, because with the right code, anyone can process massive amounts of data. 2
To come back to this:
Its only going to further move the needle in a negative way for privacy and safety.
I have to agree.
It's an awful thought though that this could turn into an "AI arms race" leading to further increase in violence, while the underlying problem, a constant public exposure of private data for everyone - which is awful in the US, and only a little less awful in the UK/EU - actively impacts real safety of people, day-to-day.
However, the alternative requires taking the long view, and patience, which the 2yr (4yr if lucky) power cycle doesn't really incentivize. Escalation pairs very well with short term goals, so I'm grudgingly waiting for things like this to escalate.
What's the way out? Is it too late?
Footnotes
-
Hasn't been for years: the first time CBP didn't need my passport but just did everything based on my live mugshot from their lil camera was early 2022, when I flew in for entry at a smaller airport and not MIA (because MIA truly is the suckiest point of entry). ↩
-
Just last week I ran extensive comparisons and aggregation over a 20M records deep, 240 attributes wide parquet database file within a 500ms execution window on 6yo hardware. With 4 lines of python, 2 dependencies, and 70 lines of SQL. In the late 90s on a highly tuned Oracle db this would take hours on a HPUX or AIX cluster. In the 2000s this would take 10 mins up to half an hour on commodity Linux. In the 2010s this would take up to a minute. Now it's sub-second on aged commodity hardware. And no AI was used in any of this (transformers/diffusion/NLP implementations are slow af anyway and skilled data scientists rarely need it post-ingestion.) I don't run this on people data but on industrial process data, but the nature of the data doesn't really matter anymore... though a moral being would not want their code weaponized for total privacy invasion. ↩
reply
I think the question is what is the representation you need to be able to index and lookup facial data. I feel like last time I did a review of the SOTA i didn't find any easy open source implementations but this actually looks promising https://github.com/serengil/deepface
I have this weird idea where everyone seems ok with the government and corporations having these capabilities so we should implement stuff to make public indexes of faces so EVERYONE has the capability then people might care. Like if i setup my own flock cameras and started publishing licence plate data i feel like there may be some objections. We should just make an open source flock that people can put on their own land that makes all the data public (or however the person wants) https://github.com/openalpr/openalpr
reply