Profile picture
Kyle Machulis @qDot
, 54 tweets, 12 min read Read on Twitter
My AI blowjob thread was mildly popular, and I've been thinking about things I missed in the first round.

Time for a @SarahJamieLewis style "I'm still thinking about this" thread continuation.



First, a few prerequisites. The AI Blowjob paper I'll be referring to is here:
This thread will also reference @SarahJamieLewis's work on "Blowjob Fingerprinting", which starts here:
I was curious other responses to the @motherboard article on the AI Blowjob paper, tweet searched the URL, found this tweet and corresponding thread addressing the problems of mass data usage as automated objectification. It's a good point I didn't hit.
In my thread yesterday, I skipped past the human identity/gender context 'cause fuck if I have anywhere near the language or reference base to be pontificating about that topic on the internet.

Not that it stops anyone else, but believe it or not, I get self-conscious sometimes.
I will let the many wonderful feminist porn academics address "female action as abstraction". It's a complicated topic that my "I just wanna do the math" brain can't quite grasp.

But in the thread on using abstract data of women to make the ideal blowjob, I realized something.
Nowhere in the paper does it say they used 109 hours of purely heterosexual porn.

It just says 109 hours of "AAA" porn.

So uh.

Also, we should revisit the beautiful UI the company used to encode the data.

I fully expect to see this interface in the next iOS/macOS release next year.
To ground this in reality before we fly off into the wild outer space of Saturday Night Twitter Blowjob Algorithm Philosophy Wackiness, I'll make 2 points.
First, Very Intelligent E-Commerce, purveyors of the Autoblow, 3FAP, etc, are well known for misogynistic advertising. We can probably assume it was all het porn, because in their advertising universe it is illegal for male mouths to contact dick.
This isn't to say that gay men don't also think with their cock when in consumer situations involving sex (as anyone who's spent any amount of time at, say, @mrsleather knows), but it's never been part of V.I.E-C's repertoire to use that fact for advertising, as far as I'm aware.
Secondly, the head on the UI has a shutterstock watermark. This was obviously lifted off a webpage. There's a good chance this whole paper is a sham. Putting any trust in this project is probably just a bad idea, even if it is giving me material for 2 threads of Good Content.
Ok. So. Probably het. Male head on the UI shouldn't make a difference.

But that's not what you're reading this for. Oh no.

So now it's time for everyone's favorite critical conjecture...


*gameshow music*
Let's ignore the 80 tweets I made last night about how dumb this projects is, and just consider that a lot of people are going to consider this a good idea and maybe try to do it themselves because we live in a neverending episode of Recursive Jackass.
How does the data that makes up the seed vectors for procedural blowjob generation matter, in terms of context, consent, and culture?


Starting with context. So we're makin' a blowjob algorithm, using porn. How do we state the orientation of the algorithm? Is it a heterosexual algorithm if it's only heterosexual porn? Is it gay if it's only gay porn?
(BTW, quick note: I probably don't have shit for answers for a lot of the questions I'll be asking here, so if you need a thesis topic, just grab a question and go. And let me know when you publish 'cause I'd love some answers too.)
I could absolutely see this becoming a product differentiator in the future. Same hardware, differently oriented datasets for training.
But this problem in general goes back to my questions about data biasing in the first thread.

To make this more understandable, let's go back to something most everyone has seen: A stop sign.
Whenever you have to fill in one of those annoying captchas that asks you to find all the stop signs, you're usually doing it to fill in training sets for, say, mapping utilities, or self driving cars.
Those of us in the US know that stop signs are octagonal, red signs with white lettering, because of course they are.

(This is the point where I realize I should've researched stop sign internationalization before making this point so bear with me while I try to salvage this)
(phew ok found some weirdly shaped stop signs thank fuck)

Let's say we're the company collecting this data, and now we take it and put it in a car but that car goes to Japan or Paupa New Guinea and OOPS FUCK WE JUST KILLED SOMEONE.
This is a rather gory concretization of our problem.

The context of what the car was trained for matters in the terms it's being used in. If it's taken out of that context, things will go wrong, and in the case of the car, kill a person.
But, we're talking about Blowjob AI, not car AI.

We hope that if a Blowjob AI is trained incorrectly, it doesn't kill anyone. (Ok we hope that in reality but I would absolutely watch a horror porn parody about this. cc @wood_rocket)

But what DOES go wrong?
Our blowjob AI is going to have contexts, both inherent and marketed, based on the perceived sexual orientation of the material used to trained it.
In terms of the Autoblow AI, this is marketed as "blowjob scenes from AAA porn". So the only context we have is "blowjobs from porn". Not what kind of porn, or scenes. Might be real penises. Might be fake penises. Might be straight. Might be gay. Might be zoo. Who knows.
This put us, as consumers, in a sympathetic position to the self driving car.

We're given some data that says "This is a good blowjob", much like the car was given data that says "this is a stop sign".

We're put in an environment of using that data and hoping for the best.
Could we reverse engineer blowjob algos to figure out what kind of sources they came from?

Probably not, the data is low fidelity and looking at the insides of the NN's that generated it will just show you Deepdream Blowjobs.

No one wants to see Deepdream Blowjobs.
This is where I'm directly referencing @SarahJamieLewis's fingerprinting work, so, once again, see this for more info by someone who actually did work on this problem.
So if we can't get the training sets (because 109 hours of video data is a lot anyways and is a relatively small set for NN training), and we can't get the context...

What does this mean for algorithmic consent?
This is where the human side of algorithmically generated sex gets super messy, so to speak.

If I'm staunchly of one orientation, and I unknowingly use a dataset trained on another orientation, what does that mean for me in terms of consent?
Movie synchronization with sex toys goes back at least 20 years now. With movie synchronization, a toy will move along with a movie.

The movie, having auditory and visual aspects, provides the context for user consent. Don't want het stuff? Don't use het movies. The end.
In the brave new idiotic world of Autoblow AI, we no longer have that sensory vetting. We just have some movement vectors, generated from data trained from "porn".

If we don't consent to the type of porn the movement was made from, how do we deal with that?
There's also the consent of the models involved, which is a larger issue with ML dataset training. 109 hours video of models giving blowjobs to build a product that a company will sell as the perfect blowjob giver. Those models will go unpaid and uncredited.

So, with neither context nor consent being made any clearer, let's head to culture.

We as humans haven't really done much to deal with how ML trained in mass by other humans is acted on us. So we're gonna have some fuckin' tabloid freakouts about algosex pretty soon, I reckon.
(And if any of the ethicists/philosophers working on the topic of ML and humans want to interject with some references to people doing work in this field or papers/articles, by all means please do and I'll retweet)
The cultural questions overlap with the consent questions quite a bit, but it basically sums up to the 37 Blowjobs Argument from Clerks.
There's already a *ton* of writing on spectatorship and porn, because for the past [however long humans have been drawing/painting/photoing/filming/recoding porn] years, it's been a spectator sport.
Even with movie synchronization, which puts the spectator in a vaguely-sympathetic haptic replication of the scene they are watching/hearing, it's still a recorded instance of a past act.
With algorithmically generated sex, we now have a training set made of hours and hours of human actions on other humans, that are now generating NEW actions on other humans.

There are probably more than 37 blowjobs in 109 hours of blowjob videos.
While many of us give approx zero fucks about ye ol' cishet monogamous lifestyle writing, algosex takes those "is using sex robots cheating?" questions and turns them into "is using sex robots cheating with everyone involved in the training set?".

How will societies deal with the fact that algosex data is human trained?

There's already all kinds of sex worker social stratification to begin with, and most of the data will come from them.
But, we'll also have shitloads of amateur porn from pornhub and who knows what other sources. Porn is cheap if not just plain free, and plentiful.
So to the sex worker issues, we add the lack of origin around the training data, and BASICALLY EVERY PROBLEM HUMANS HAVE WITH ONE ANOTHER pops up at one point or another, 'cause it's usually gonna be humans having sex with humans, or, uh, other things.
In summary, algorithmically generated sex from human based training sets is going to be #problematic.
We're still figuring out how to deal with masses of humans being driven in massive directions at massive speeds. This problem isn't new, we're dealing with it in social media, politics, etc. I'm just throwing another log on the fire.
Unfortunately, since I'm not providing much in the way of answers here, it's kind of "throwing a log on the fire and running away".

I work in sex tech, so as I ended the last thread with, this is as much me thinking out loud as it is doing some "starting a dialog" bullshit.
These are the kind of questions I try to process when working on new applications for Buttplug. Since I'm providing haptic actuation to digital experiences or communication, it's on me to at least consider the impact.
Lots of unanswered questions to think about, lots of work to do, but can't really have one without the other. By throwing my thinking on here, knowing I've got followers who are actively working on these topics in academic contexts, I hope for guidance. And do pay for it too. :)
Anyways, that's pretty much it for this thread. Don't have a neat bow to tie it up in, nor do I feel any more enlightened about whether an algorithm can have a sexual orientation.

But at least I tweeted a bunch?
We now return you to your regularly scheduled buttplugs.

At least, those of you that haven't muted my chatty ass over the last 2 days.
Oh yeah and as always I like money.
Missing some Tweet in this thread?
You can try to force a refresh.

Like this thread? Get email updates or save it to PDF!

Subscribe to Kyle Machulis
Profile picture

Get real-time email alerts when new unrolls are available from this author!

This content may be removed anytime!

Twitter may remove this content at anytime, convert it as a PDF, save and print for later use!

Try unrolling a thread yourself!

how to unroll video

1) Follow Thread Reader App on Twitter so you can easily mention us!

2) Go to a Twitter thread (series of Tweets by the same owner) and mention us with a keyword "unroll" @threadreaderapp unroll

You can practice here first or read more on our help page!

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just three indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member and get exclusive features!

Premium member ($30.00/year)

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!