Profile picture
Kyle Machulis @qDot
, 82 tweets, 19 min read Read on Twitter
Ugh. I will be receiving this article for the foreseeable future, and while I love yelling at people for sending me dumb shit, I have things to do. In an effort to save time, I present

ONE DEGREE OF TRANSLATIONAL FREEDOM DOES NOT A BLOWJOB MAKE

A THREAD

motherboard.vice.com/en_us/article/…
First off, this article itself is not the problem, and I'm always a fan of @samleecole's work, especially when it's about me. I'll be addressing the whitepaper that is the topic of the article, posted below. The article is a good summary of the paper. autoblow.com/bjpaper/
I will also fully admit that, with this thread, I am absolutely falling into the marketing trap of Very Intelligent E-Commerce Inc.

I wish I was good at *anything* as Brian Sloan is at making men drool all over their own cocks via spurious technical claims.
My argument is currently in 3 parts:

- Metrics analysis
- Data accrual and quality
- Applications of haptics and biomechanics

I may add more arguments as I write this, because I just woke up and my brain is still booting and I just wanted to eat a waffle BUT NO. HERE I AM.
Please note that I am not an expert in any of the aforementioned fields, nor any field in general, except ones that I have made up to further my career.

If you are smarter than me and see something wrong here, please correct me. I trust grownups.

Except grownups that are cops.
PART THE FIRST

METRICS ANALYSIS

So let's start with the data we've got.

The paper states that humans encoded (to be addressed in argument 2) 109 hours of blow video, using the slider pictured here. (Image via @motherboard. Is that shutterstock watermarking on the head?)
Let's ask ourselves.

What makes a blowjob a blowjob?

You can probably name many qualities.

- Linear motion (WE NOW HAVE THOSE I FUCKING LOVE SCIENCE or some shit)
- Sucking sounds
- Tongue movement
- Feelings about person giving it
- Throat constriction

The list goes on.
In the small list I made, we have 3 haptic signifiers, 1 auditory, and 1, for lack of a better term, intangible.

Unfortunately, the intangible is REALLY important, but until a sex robot can implant good memories and experiences of itself in your brain before sexy times, welp.
As the production costs of putting a speaker in the machine to play the sound of stirring a pot of Mac & Cheese (can't believe I'm giving away my porn foley secrets here) would be too high, we also drop the auditory portion of the list. We're down the haptic.
So we've got linear motion, tongue movements, and throat constriction.

Unless the porn was taken in an MRI (PLEASE DO THIS I WOULD WATCH THIS PORN), we don't have a way to track the tongue and throat.

So all we have to track is head movement.
One could ask "But @qdot, what about a dildo with sensors? Couldn't that track all of this?"

Yes, it certainly could, and the Orgasmatronics Ambrosia was a (simplified) version of that.

However, collecting data from videos is much cheaper and easier, so that's why we're here.
So now we get linear motion from a video.

What are the qualities of linear motion in a blowjob?

Let's do a little thought experiment first to make sure we're all on the same wavelength here.
Make sure no coworkers/roommates/etc are looking, then starting moving your head back and forth like you were giving a blowjob.

Go ahead. I'll wait.
Ok. Now that we've all embarrassed ourselves (except me because come on I run a project called @buttplugio and am wasting my morning writing about this), let's take a second and consider what just happened.
We just tried to give an imaginary blowjob. How did we do it? Well, we had to imagine:

- something to suck on (not making assumptions here, some people really like pickles)
- how whatever is attached to the thing we're sucking on reacts
- our own physical limitations
- etc.
Obviously there's a lot of though and work that goes into a well crafted blowjob.

The computer does not get that thought and work.

What you are doing in machine learning is training an extremely tiny brainlike matrix on an extremely tiny universe.
In this case, our computer's extremely tiny universe is "motion extracted from some videos". It doesn't care what the motion is. It doesn't even know it's motion. It's just some vectors that are being fed into a system and, as the paper says, "quantized".
That computer's one job in its computery life is to take those vectors, and make more like them. That's what the paper is all about.

If you've heard about problems with biasing in machine learning data lately, prepare to hear more, because GUESS WHAT.
The computer just knows how to take what it knows and make more of it. What it knows is the vectors fed into it. The vectors fed into it come from human perception of motion in a video.

We lose all other context in the encoding.
When was the last time you thought "Gosh, the linear motion of this abstract oral sex is really spot on, in that is not taking into account my pleasure, my anatomy, or any other surrounding factors of my environment or self"?
And thus we land smack dab in data biasing, aka "All data is good data because it's data".

We get out what we feed in, which means if it's shitloads of pictures of white people and we get more white people, if it's one-dimensional linear movement, we get more linear movement.
In some cases, like the first, it's offensive and exclusionary.

In other cases, like the one this thread is about, it's just dumb. 109 hours of linear head movement is 109 hours of linear head movement.

Linear head movement is not sex.
We need to be able to extract more useful data to create to actually have an experience. With the whitepaper data, we have a one-dimensional vector that we can apply pretty much anywhere.
With the data we have from this whitepaper, we can drive a sex toy, and that's what its for. For that sex toy to work, it will require good haptics (which I get to in part 3 of this neverending void of a rant).
Since the data lacks context, it absolutely REQUIRES the sex toy to be there in order to make it sexy.

By this argument, we could take this data, feed it into a self driving's car nav system, and stick a fleshlight on the front of the car, and the car would be "giving blowjobs".
While this would be quite preferable to a certain subset of dragons with proclivities toward automobiles, it doesn't do us much good otherwise. It's just a god damn vector.

It's not that interesting.

End Part 1.
PART THE SECOND

DATA ACCRUAL AND QUALITY

AKA YOU HAD MULTIPLE PHDS APPARENTLY. WHY THE FUCK DID YOU MECHANICAL TURK THIS. DID NONE OF THEM WORK IN MACHINE VISION.
According to the paper, the 109 hours of video were encoded, by humans, using the interface below.
I have two problems with this:

- The data
- The people

which means I have problems with pretty much all of it.
First off, if we're going by the data from that image, the time positions were recorded to 5 digits.

Let's be generous here and say that all of the video was 60fps.

This means 16ms per frame.

The human encoding this cannot take in more data than that, physically.
60fps exists because of [hand waving and nyquist and persistence of vision and a bunch of other shit go ask @omershapira or @_vade or @matttrent because the middle of a blowjob debunking thread is a great place to do a follow friday]
Anyways, 1/60th of a second is 16 ms, or 0.016 seconds. So if this paper is sciencey mcscience, how on earth are we getting 5 digits of [precision|accuracy I can't remember which I'm not a sciengineer] with only 3 digits of data to work with in the first place?
And then there's the encoding method.

This study decided to go with what I call "soylent tagging"

Soylent tagging is having a bunch of bored ass mechanical turk users sit there and tag fucktons of data. Humans are in the loop, and they're gonna fuck up because humans.
What humans provide to data is context for other humans. We build machine learning systems for other humans. Other humans don't exist in the tiny universes of the tiny computer brains we make, so we have to add humans as ingredients to make more humany things. Hence soylent.
This context is what you give when you do one of those captchas that wants all the stop signs or all the birds or whatever. You are an ingredient in soylent tagging then.
The problem with soylent tagging here is that we didn't need it. The data was already tagged. The people doing the work just got loads of blowjob videos, so we already know they're blowjobs. At this point, a computer could've done the work.
To give an example of how a computer could do the work, let's look at @magicleap's Superpoint encoder. Superpoint is a pretrained Optical Flow network. github.com/MagicLeapResea…
This means that @magicleap took a whooooooooole bunch of videos where shit moves around, and trained a computer to find "interest points" that denote the motion.
This kind of thing is SUPER useful for, say, Augmented reality. If you want to display a virtual object on a real surface, you have a camera you use and you have to extract surfaces and motion and all sorts of shit out of that to make sure that object looks like its staying put.
So how does this connect to the blowjob data in question?

Well, this study uses pre-recorded video, but it's still video. And we don't care about depth extraction/etc because we aren't looking to do object placement.

So why not just run the video through something like this?
The data would definitely need cleanup afterward, due to things like cut/edit points in films, changing perspectives, things like that. But you would still get a frame-by-frame interest set which you could easily extract one-dimentional motion data out of easily.
If you want to see what machine learning can do in terms of auto tagging edit points, scenes, and things like that, check out @_vade's project, @Synopsis_. github.com/synopsis
End part the second.
PART THE THIRD

HAPTICS AND BIOMECHANICS
I am a blowjob machine.

Blowjobs are a beloved hobby that I would turn into a job except I know enough sex workers that I damn well know I couldn't do it, because right now I work with computers and the only thing more broken than computer is people.

(Pay sex workers)
As a blowjob machine, I am aware of the things I am capable of in terms of stimulating the senses of other humans.

Assigning all of those qualities to a machine is a tall order.

That's why, in haptics, we work with things called "haptic illusions".
Haptic illusions are basically trying to fool the body into thinking a complex touch sense is happening via a very rudimentary mechanism, because touch is FUCKING HARD. For more information, see @Dave_Parisi's Archaeologies of Touch. amazon.com/dp/B079SJSXDJ/…
I'll also be speaking about this at SLSA 2018 in Toronto next month. litsciarts.org/slsa18/
Annnnnnnnnnyways, this means we have to take the one useless vector of data that we have, and turn it into an approximation of the human mouth.

And we have to do it at a price that people who buy sex toys can afford.
Haptics is by no means an affordable field. Sex toy haptics, doubly so. The most advanced sex toy for penises so far has been the RealTouch, with friction, lubrication, warming, etc. See this article by @Dave_Parisi again in @logic_magazine for more info logicmag.io/02-realtouchin…
Blowjobs are going to require lip construction, tongue movement (which, sitting here moving my tongue around a bunch and thinking about it, has at least 3 rotational axes and 3 translational axes of freedom, so it's kinda complicated), saliva production, etc.
That is a TON of mechanics to put in a toy, and then you also have to add the neck movement, because otherwise why the fuck did we collect all of this data in the first place?
If we don't have all of that, and just have linear movement, we then once again have to ask, what is the absolute minimum blowjob? Is there a blowjob MVP that this "Autoblow AI" will satisfy between the learning and the haptics?
End Part the Third, and I have to run to a lunch meeting so my conclusion may be in a bit.
BTW if you've gotten to this point and absolutely need more sweet sweet @qdot content and can't wait for me to eat and finish this thread, check out my YouTube channel where I basically perform shit like this thread on video youtube.buttplug.io
AND WE'RE BACK.

I'm now filled with a Bay Area Urban Food Log (aka burrito), which means I'm all powered up for...
PART THE FINAL

GENITALS ARE STUPID AND NOTHING MATTERS
I have just spent an really, really dumb amount of time explaining in "angry layman" detail why a paper that was a "science" marketing stunt isn't actually science.

This will have roughly 0 impact on Autoblow sales.
The Autoblow Blowjob AI Whitepaper is a type of marketing that bears some similarities to oft-pranked but actually sincere Russell & Whitehead proof that 1+1 equals 2 via set theory (in the hopes for unifying mathematics on a single axiomatic set).
When presenting someone with the 1+1 = 2 Principia Mathematica proof, you make someone really question whether they *know* the meaning of 1+1 = 2, due to the arcane symbols and terminology required to build up understanding of where the argument is going.
Whitepapers like the Autoblow AI Blowjob, and other marketing stunts like Camsoda's "Blowjob patterns" (with followup work by the wonderful @SarahJamieLewis via the tweet linked here ), use technical terms to build up what looks like a consistent basis.
Now it's a little unfair to compare W&R's formalist approach to explaining the universe to to shitty adult company marketing.

But the unifying aspect is, all of these arguments unravel, via either a mathematical Gödel, like Kurt, or a Buttplug Gödel, like, uh, me, I guess?
Now, of course, not being able to cast our wonderful, complex universe into a simple set of rules has confounded humanity for millennia so far and millennia to come.

But we still care. We still try.

When it comes to what we stick on/in our genitals, it's different.
When we wanna stick something on our (hopefully) happy place, our brain dumps (hopefully) happy juice and a lot of our worries just float away.

This is why we have the whole idea of "drama".
Since these marketing ploys have to do with sticking things in/on happy places, a whole whitepaper isn't really needed.

The whitepaper is just to get journalists to cover it, and piss off people like me who will write stupid twitter threads like this instead of doing their job.
There's two things I can extract from this though.

First, sex is difficult. Really difficult. There people out there thinking about it in real, difficult, sciencey/humanitiesy ways (@NicoleRPrause @hallielieberman @drkatedevlin to name but a few).
These people run labs, write books on their work, and they put their name on them.

This paper is nameless. We don't actually know who wrote it. We can't vet any of these. Because it's not academic, it's marketing in academic disguise.
And this is not to say there are not people with academic titles who won't sell themselves to sex toy companies to be "futurists". My pinned tweet exists for a reason, and that reason is Dr. Ian Pearson.
Dr. Pearson has given us some absolutely wonderful bullshit over the years, like the boob implant ipod theregister.co.uk/2005/10/13/mp3…
And then there's Dr. Pearson's comissioned report for a sex toy company about how "Sex Robots are the Future", including this graphic of hunky Jesus battling for the fucky high score on some imaginary score board.

graphics.bondara.com/Future_sex_rep…
All of this can happen because we can combine complex yet wrongly used technical terms and rudimentary sex education into a sharknado of marketing speak that will make people spend $200 on $10 worth of parts that got $15 worth of marketing.
Which brings me to my second point, why the fuck did I just spend 2 hours writing this?

Unfortunately, it wasn't for any of you.
I'm currently spending a lot of time working out in my head why the hell I spend a non-trivial portion of my life on the topic of remote intimacy and touch.

It's just something I have to revisit so often to keep myself sane.
It's comes down to "this is how I understand people".

I have one of those gloriously broken brains that let's me make broken machines do math really fast and other people pay me a lot for it.

It'd make me more money if I could break my brain more and drop the people part.
But, for reasons that would take another thread longer than this, computer mediated intimacy is how I figure out "how to people", which might explain why I'm really bad at it sometimes.
This thread is me deconstructing, in public, my thought process around how people are intimate online, how they communicate through these systems I help build, and what it means as we continue to build them instead of burning them all down and running into the woods.
What I hope you take from this Friday afternoon post is

- An understanding of how sex may not need the "science" advertising throws at it
- A critical eye toward sex tech
- How tech and humanity work via intimacy (which would've been a nice alternative from politics but NOPE.)
But, if nothing else, you will know that I am absolutely fucking godawful to take sex toy shopping.

Seriously, Good Vibrations, Mr. S, some midwestern truck stop store, I can ruin them ALL.
And with that, I'm going to go to the thing I'm paid for now.

We now return you to your regularly scheduled buttplugs.
Oh and if you like beans and George Wendt and me writing threads like this please sign up to my patreon because being this weird is expensive. patreon.com/qdot
And for anyone who just wants all of this on a single page
And for anyone who just hasn't had enough and want EVEN. MORE. THREAD. I'm now continuing this here:

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!