s4e06: Criticality 

by danhon

0.0 Station Ident

Wednesday, 15 March 2017 and a break from the usual newsletter programming and a continuation of s4e01: umbra[0]. What better way to deal with stress and anxiety by trying to do a little creative writing.

[0] s4e01: umbra

Have a good day, everyone. We’re halfway through the week and I’m down in Sacramento for the day in the name of hard work toward inexorable, patient progress. As ever, send me notes. See you on the other side.

2.0 Criticality

Wednesday, March 15, 2017

umbra’s had me out at MPK 16 for the past two weeks on a spike and I’m not sure when I’m going back.

Late one Thursday afternoon Karla and Dan stopped by at my desk and asked me to come to Four, one of the conference rooms upstairs and Amy was there to do the NDA dance before anyone would talk.

This is the NDA dance, and it’s always been this way: first, an NDA to let me in on the spike codename (Rorschach), a second to tell me who the client is (Facebook – my third time with them now) and a third so that we could stop silently smiling at each other and start talking about whatever it is I’m going to be doing for the next few weeks. Rumor is that we borrowed the NDA dance from Apple; anyone who’s inclined to spend the time or paid attention will see that our lawyers have drifted back and forth between jobs at both companies.

Anyway. Rorschach. This is what Karla said:

“Mark’s pissed.”

I raised my eyebrows. In situations like this, it’s easier to stay quiet and wait to see what happens. Karla in particular likes to see good use of non-verbal communication cues, and it helps that umbra’s founders are also the kind of people who are both very smart and like to hear themselves talk.

“About the news thing,” added Dan.

“The news thing,” I said.

Mark, I’ve learned, is pissed about lots of things but you wouldn’t know it from the degree to which he communicates so earnestly with the rest of the world. This is, of course, because he isn’t the one who is communicating so earnestly with the outside world. He has a whole bunch of people to do that, and some of them are even the best Chinese Rooms you can buy. But when you’re in a meeting with him and you’re showing him something and he doesn’t like it and he’s… frustrated at the stunning lack of power he has over the nearly two billion people who use his company’s product every day then I suppose yes: you would say that he’s pissed. It appears that he’s pissed about all the criticism he’s getting about Facebook being a platform that’s really good at spreading things that aren’t necessarily true.

No wonder this is an opportunity for umbra.

“The news thing,” says Karla. “I mean, the graph is doing exactly what we’d expect it to do given the shit that’s being fed into it. Bunch of dumb nodes doing dumb processing and a finite amount of operations you can do on it, what does he fucking expect.”

“A lot of shit is being fed into it,” I agree. “What’s the pitch?”

Karla leans back as Dan stands up. Dan’s turn, then. He stands up in the way that a control room’s atmosphere changes when a reactor reaches criticality. Something serious is going to happen, it might involve an explosion and at the end of it, the meeting will be over and I’ll know what I’m supposed to be doing. It’ll probably involve spending the next few weeks in Menlo Park. And then it blows.

“They’re wasting their time with all this Yann LeCun stuff. I mean, we know it’s  bunch of wank because they’ve told us it’s a bunch of wank. Whoopty-fucking-do, now we’re better at putting blue boxes around protestors and police officers and categorizing intent based on stance. No ROI and  on top of that, some fuckwit chooses that fucking Black Lives Matter photo for the public blog post. Bunch of amateurs succeeding in spite of themselves.”

“But,” prompts Karla.

“But,” he grins, “it’s not garbage, is it. I mean, the stuff that spreads isn’t garbage. They call the stuff that spreads a virus. Agency clients come to us for the drip-feed of what the next adaptation might be so they can stay one step ahead of their useless game for attention. It’s not even a good virus. It’s the most successful, least-dumb virus.”

“It’s a training set,” I say, taking a calculated risk to guess where he’s going given the way that he’s trash-talking Facebook’s Artificial Intelligence group.

Dan smiles, which means I will probably get an unreasonable bonus this month. The instant reaction I have to this is not outweighed by how my bank account will feel about the unreasonable bonus and, somewhere in my head, something pushes closer toward a different kind of criticality.

“A training set. Something interesting for their new nVidia toys to dream about. You’re going down to campus, you’re getting one of LeCun’s teams, time on the new clusters and you’re going to tell Mark and us what we can do with dreams straight from the feed.”

It’s simple, really. Train a stupendously giant neural network, based on the kind of thing that can beat a human Go player, all the most successful things that get shared on Facebook. Probably prioritize image macros first. Run it backwards to see what it thinks makes an image macro spread. Take the image macros it makes and seed those onto the feed and do the equivalent of hitting the network with a stick until it starts making image macros that beat the performance of the training images.

The technical work of applying post-human pattern recognition to the biggest set of culture that takes advantage of our brain architecture took less time than it took to persuade the Facebook team that this was worth doing.

It took marginally more time for umbra to set up the plausibly-deniable fake news sites that would be the sources of seeded network-generated content. And then we just hooked it all up together. We knew it worked because Mark shared one of the things in our meeting and looked guilty about it.

So that’s what I’ve been doing for the past two weeks.

Well. That’s not all I’ve been doing.

I think the thing inside me that was going critical, went critical.

There’s another network, running on the side. Same training data.

It’s trying to make a vaccine.