You Look Like a Thing and I Love You

by Janelle Shane

You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place was my early February 2020 stall-out book. 

Not for reading: it is way too fun not to cruise through. Check out this little taste:

This failure to plan ahead shows up fairly often. In level 2 of Super Mario Bros., there is an infamous ledge, the bane of all game-playing algorithms. This ledge has lots of shiny coins on it! By the time they get to level 2, AIs usually know coins are good. The AIs also usually know that they have to keep moving to the right so they can reach the end of the level before time runs out. But if the AI jumps onto the ledge, it then has to go backwards to get down off the ledge. The AIs have never had to go backwards before. They can’t figure it out, and they get stuck on the ledge until time runs out. “I literally spent about six weekends and thousands of hours of CPU on the problem,” said Tom Murphy, who eventually got past the ledge with some improvements to his AI’s skills at long-term planning.

But for writing reviews, it broke me. See, after reading You Look Like, I had this plan to implement a GPT2-based machine learning “large-scale unsupervised language model which generates coherent paragraphs of text” concept where I would dump in a hundred or so of my book reviews as a base library, feed in the You Look Like quotes that I excerpted, and compare the two for hilarious (!?) results.

And now we’re nearly three months from when I read this book. The weight of this Machine Learning (ML) project broke my motivation and, much like not responding to a text for way past the point of politeness, it was getting harder every day to sit down and just do it. But this was really my only reasonable opportunity to fiddle with GPT-2 or any generative text software, so I finally went ahead and did it. I smashed my face against GPT-2 for about a day, fed it a bunch of my old reviews—thirty, to be mostly accurate, rather than the lofty goal of one hundred—and then picked through the gibberish. Even my tiny experience confirmed everything You Look Like taught me, not that I expected any less:

I’ve found that even the most straightforward task can cause an AI to fail, as if you’ve played a practical joke on it. But it turns out that pranking an AI—giving it a task and watching it flail—is a great way to learn about it.

This was a great way to learn about machine learning based on text! Yes, the broken bits were mostly user error: attempting to import .docx, then .rtx, before realizing when the library is built off “text files” literally nothing but .txt files will work. Terms of art, I would say—because in most cases, to most people, a Word File is a text file, and vice versa. Not so here. But that technical error was the number one reason why my first attempt at generative text returned nothing but sequential numbers (and why I ended up with thirty random reviews rather than one hundred of the more tech-focused ones). Once I did finally get some real text, I was hoping to see a broad sense of my own writing style divorced from a personal sense of creation, but all I found was lofty prose and convoluted thoughts that lacked any sense of clarity, cohesion, or approachability. Heavy levels of self-deprecation, too.

Two days removed and I would be starting from scratch if I tried to redo this. [runtime > restart runtime] but for my brain.

So what I’m going to do is break out a bunch of relatively coherent sections from my GPT-2 dump—format them in bold so you know they come from the DinaburgWrites.com oeuvre—let you look at them, maybe toss in some editorial comments, and then we’ll reconvene to discuss You Look Like as a book for a little bit.

These are questions that cannot be answered or illuminated by facts or reasoning. They are rhetorical devices that vaporize any thought or acknowledgement that they are not representations of the author or the reader. The pepperoni pizza I ate was so good I almost died. It was too late. The story has followed me into the future, but it has not yet ended.

This is the ending I was promised. This is the beginning.

I hope you are too. But never, ever, to read this book. Ever.   I wanted to put this book down before I wore my tights, but I didn’t, so I’ll stick to my burgers.

This paragraph took about...six hours of tinkering to create. I will say that, at the point this was being spit out, I did not think my grand experiment was worth it. I mean, I was thrilled i got anything to work (see above, wherein I produced nothing but sequential numbers), but this sample of my trained model was pretty raw. Perhaps you could separate out some of the parts about “pizza,” “burgers,” and “my tights” to tell a story. But it wouldn’t be a very good one. I want you to see, warts and all, the nonsense I had to pick through to dredge up a few good lines. Context, thy name is not Machine Learning.

OveralI, if you wanted to use two or more contiguous sentences, things did not get much more coherent at any point.

AI is all about context.  The modern decontextualized object is a tweet; a post; a photo; a short story; a poem; a preamble; a cover art—all the time and at once. Think about what you just read and say it aloud; say it once or twice. Partly because it is a tweet, part because it is a post. Part because I, like most things, am extremely self-aware.

Ah jeez, that’s starting to sounds a little bit like a threat. Going a little SkyNet on me, there, jefe. Regardless, that’s the last of the larger text blocks, so if you’re just kind of skimming, your long nightmare has come to an end.

Next up, I’ve got a list of generated alternate titles for You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place:

• You Look Like a Thing and I Love You, But You’re Not a Thing: The Neuroscience of the Looking At the Self in the Twenty-First Century’s Most Trash Common 

• You Look Like a Thing and I Love You To The End Is the Key [Repeat]

• You Look Like a Thing and I Love You, and I’d sell you my college coat and T-shirt, too. 

• You Look Like a Thing and I Love You, You’re Just a Bomb and it’s Not Even Happening 

• Machine Learning for the 21st Century: From Factory to Cloud to Factory or Less in 5 min. 

Ok, so, we’re finally getting to the part where I can see the appeal of ML. I could have refreshed that particular set of parameters for a long time and spit out ten or twenty titles that made me smile as much as You’re Just a Bomb and it’s Not Even Happening, but that’s kind of like cooking the dice, so we get what we get. Once the setup was done, refreshing takes like, a minute, tops. You could spend an hour best-of-times-blurst-of-times-ing till you got Pulitzer material, but let’s move on.

Well, before we go too far; there was definitely a way for me to embed an API so you could roll up your own random titles, but c’mon, we don’t want this review delayed to August, and I would clearly need three months of procrastination-based anxiety before I even got started on the two hours worth of reading it would take to do that. So enjoy this video of The Simpsons reference I just made:

Here are some of what I’ll call the COLLEGE FRESHMAN STONER VIBES: sentences that seem cool or deep as long as you don’t think about them for too long:

• I don’t know if makeup will help but at least I can’t see myself in the mirror. 

• I’d create stories of things that made me what I am now, and of things that made leaving home more simple than ever.

• The hand that punches the paper points toward the far right, while the other one looks toward the left. The color of the ink on the page will vary depending on the orientation of the crystal, but will generally be in the opposite direction to the sky. 

• What happened when I didn’t even bother? What happened when I did something I didn’t.

• Lists are lists—they are you, everywhere’s existence, at least according to one side of the page, so long as you don’t tell the whole story 

• I don’t know if there is a more resonant form of intellectual life than juvenile history, and in that spirit I’d create a collection of childhood dreams that I I believe you will recognize as reality. 

• It was hard to know when you’re going to take advantage of our collective unconscious. I thought it would only be accessible to you now. I’d built a story of what the story would be, listing it off as time went by.

• It is important to remember that the twenty-first century is a flood, and a storm is coming. More than anything, it is a warning: keep your mind in the water, and be prepared for the next wave to come 

• No human can ever become an animal or a plant or a animal or a bird or a dragon or a bird or a tree or a mountain or a 

Yeah, that last one just kind of ended. I admit, I’m curious as to what else humans can never become, but I feel like it might have just continued to repeat “animal.” Earth to ML:AI, humans are animals. Now whose mind is blown.

Ah, but there’s so much more just from my small samples, but I think you get the point.

You Look Like is like this, except better: funnier; smarter; more giraffes. 

...[E]ach image was seen an average of three hundred times, which is why their dataset contains lots of similarly worded answers:

no, just the 2 giraffes

    no i just see 2 giraffes

    nope just the 2 giraffes

As you can tell from the answers below, some respondents were more committed to the seriousness of the project than others:

    bird is staring at giraffe asking about leaf thievery

The other effect of the setup was that each person had to ask ten questions about each image, and people eventually run out of things to ask about giraffes, so the questions got a bit whimsical at times. Some of the questions humans posed included:

    what is giraffe song

    does this look like elite horse

    how many inches long are bears, estimated

Humans do weird things to datasets.

That’s the real conclusion. Humans do weird things to data sets, which are weird enough already. Also, elite horse. That’s just quality content. Ah, so, if you’ve read this far into a review for this book, you should just go to the author’s website.

You’ll like this book. And if you don’t believe me, here are some closing remarks from my dataset:

• This Book was a Likeness. It was also a Moment. And a Time, too 

• This Book is Wonderful, Everlasting, and it contains the best advice I have ever seen. 

• This Book is extremely wise, extremely smart. I love that this book is being written; it reminds me of why I came into being as a reader

• If you couldn’t decide between them, it was a fantastic book and an utter disaster.

Don’t quit your day job, ML algorithm.