It’s not looking for stuff like you’re describing, skeletons, graves, birds, whatever spoopy “dark” stuff you put on there. It’s looking for nude people, then actioning/alerting if it thinks minors are in the images. You definitely don’t want to honeypot this ai bro.
That’s the thing: they’re looking for something, and in order to look for something, they must look, and lots of computation (AI inference) must happen, especially when it involves nonindexed files (my own private art, from digital drawings and handmade sketches, going through photos of my own ritualistic installations, all the way to Blender 3D renders and animations, all of which is yet to be crawled). And if this involves gibibytes of data, on a smartphone whose hardware can barely run 1B parameter clankers locally without hiccuping, I can see my private files going to Google’s servers for “efficient inference”, files of which will likely end up seen by (and sowing some interesting inspirations for nightmares in the subconsciousnesses of) Google’s employees during server maintenance/audits or, mainly, end up training their clanker.
As for the latter, hopefully Google Gemini would end up “seeing” enough of my digitized psyche to randomly start silly infodumpings about, for example, different owl taxonomies and their diets down to the very gory microdetails about how owls make a tasty dinner out of mice, and/or the different interpretations by historians and archaeologists regarding the figure (Ereshkigal or Lilith) depicted in the Burney Relief, a Mesopotamic plaque of which, by the way, is part of the phone’s gallery too (including my own artistic works inspired on it) and seemingly an efficient trigger for nudity detection algorithms, which will likely put the rest of the gallery under automated scrutiny (because there was a positive, albeit a false one) and, thus, serving as the semantic Trojan Horse for my dark artistic private content to make its way to their servers.
I mean, how else Google Gemini would’ve taken the “inspiration” to suddenly start telling people to “eat rocks” and “add glue to pizza”? Its training data is crowdsourced so there’s nothing much Google can do to avoid sources poisoning the well (and I must nod at the Nephentes algorithm which I praise as way more efficient than Anubis to hurt the western AI corps) other than:
Adding said poisoned sources to their .gitignore equivalent if they know they’ll likely find poison they don’t want messing up with their expensive AI training.
Doing manual curation of people’s personal files, which is time-consuming and potentially unfeasible when dealing with billions of phones.
Trying to improve the detection algorithms to avoid poisoned sources (which requires training data, which requires them to rely on potentially poisoned sources, which defeats the whole purpose of improving the algorithms in the first place).
Stopping to put everything else into scrutiny when the detection algorithms stumble upon (potentially false) positives, risking false negatives, which will likely attract some heat from regulatory agencies worldwide and fines for failing to comply with things such as age checking laws (and there it goes the GOOG stocks, together with the entirety of NASDAQ, like a Titanic as investors realize they’re paying for fines not theirs; be it GOOG or other stock, the popping of the ongoing AI bubble will inevitably be something in this regard, Western corporations having their clankers failing on them).
At best, this will consume some portion of Google’s resources (including but not limited to human resources), albeit as infinitesimal as my silly existence can get to a worldwide corporation, but it’s still something I guess…
Image classification is old tech that existed long before ChatGPT and similar huge models became the new norm. These models are already embedded in many data flows and run within milliseconds on non-specialized hardware.
You were talking about 1B models. More modern image classification are in the mere millions. Yolo v10 has like 7M, MobileNet, which Google often uses, has around 4M parameters depending on the version. Image classification is not why there are data centers being built everywhere.
Camera apps, image and video filters and content moderation (the app in question) all use them for live processing. I can guarantee you this runs flawlessly on your phone unless it’s older than 10 years.
You’re employing AI guerilla warfare against pre-LLM technology. It’s not effective.
It’s not looking for stuff like you’re describing, skeletons, graves, birds, whatever spoopy “dark” stuff you put on there. It’s looking for nude people, then actioning/alerting if it thinks minors are in the images. You definitely don’t want to honeypot this ai bro.
That’s the thing: they’re looking for something, and in order to look for something, they must look, and lots of computation (AI inference) must happen, especially when it involves nonindexed files (my own private art, from digital drawings and handmade sketches, going through photos of my own ritualistic installations, all the way to Blender 3D renders and animations, all of which is yet to be crawled). And if this involves gibibytes of data, on a smartphone whose hardware can barely run 1B parameter clankers locally without hiccuping, I can see my private files going to Google’s servers for “efficient inference”, files of which will likely end up seen by (and sowing some interesting inspirations for nightmares in the subconsciousnesses of) Google’s employees during server maintenance/audits or, mainly, end up training their clanker.
As for the latter, hopefully Google Gemini would end up “seeing” enough of my digitized psyche to randomly start silly infodumpings about, for example, different owl taxonomies and their diets down to the very gory microdetails about how owls make a tasty dinner out of mice, and/or the different interpretations by historians and archaeologists regarding the figure (Ereshkigal or Lilith) depicted in the Burney Relief, a Mesopotamic plaque of which, by the way, is part of the phone’s gallery too (including my own artistic works inspired on it) and seemingly an efficient trigger for nudity detection algorithms, which will likely put the rest of the gallery under automated scrutiny (because there was a positive, albeit a false one) and, thus, serving as the semantic Trojan Horse for my dark artistic private content to make its way to their servers.
I mean, how else Google Gemini would’ve taken the “inspiration” to suddenly start telling people to “eat rocks” and “add glue to pizza”? Its training data is crowdsourced so there’s nothing much Google can do to avoid sources poisoning the well (and I must nod at the Nephentes algorithm which I praise as way more efficient than Anubis to hurt the western AI corps) other than:
.gitignoreequivalent if they know they’ll likely find poison they don’t want messing up with their expensive AI training.At best, this will consume some portion of Google’s resources (including but not limited to human resources), albeit as infinitesimal as my silly existence can get to a worldwide corporation, but it’s still something I guess…
!technology@lemmy.world
Image classification is old tech that existed long before ChatGPT and similar huge models became the new norm. These models are already embedded in many data flows and run within milliseconds on non-specialized hardware.
You were talking about 1B models. More modern image classification are in the mere millions. Yolo v10 has like 7M, MobileNet, which Google often uses, has around 4M parameters depending on the version. Image classification is not why there are data centers being built everywhere.
Camera apps, image and video filters and content moderation (the app in question) all use them for live processing. I can guarantee you this runs flawlessly on your phone unless it’s older than 10 years.
You’re employing AI guerilla warfare against pre-LLM technology. It’s not effective.