18. Talk to Her
“He’s coming around.”
Those were the first words I understood, though there’d been a burbling of speech before. The words were Clay’s.
A big hand pressed on my forehead, sausage fingers pulling at one eyelid. I blinked and shook my head. The hand removed itself. Dappled sunlight washed my face, dazzling.
Something else pinched my wrist in a firm, cool grip. The pressure asserted itself over the tingles in my skin.
The fugue had passed. The disorientation was short-lived, as usual. As the tingling faded, faces swam into focus: Agent Clay’s to my right, Agent Katz’s to my left. The car door on my side was open, and Clay was leaning in. Outside were trees and a grassy verge. Traffic sped by. The pressure on my wrist was from Katz’s hand, her thumb on top, two fingers pressed firmly up from below—feeling for a pulse.
And my cold coffee was still in its cup in the console holder, not all over my knees or in the footwell. Something to celebrate.
“Okay, I’m alive,” I growled, shaking off Katz’s hand. I reached into my jacket pocket for the notebook and pen. Flipped open the notebook and wrote.
I knew that my memory of the experience would fade quickly. For me, it was the sort of memory connected to drawing and writing that lasted the longest. Even when I could no longer recall anything to speak of, I could sometimes write fragments down.
My hand stopped. I looked to see what I’d written, because otherwise, I’d have no idea.
Talk to Krome.
Talk to her.
It wasn’t much. That’s usually how it was. Frustrating.
Katz and Clay watched in silence. Katz’s eyes were very round and her face pale. She looked frightened, though she didn’t strike me as someone who scared easily.
“Did I say anything?” I asked. “Anything at all?”
“No,” said Katz.
“So, who’s ‘her’?” I wondered aloud.
Katz gave a slight shake of her head.
Clay put a shiny black shoe on the doorstep and scratched his neck. “What the heck was that?”
“My demon,” I said. “An epileptic seizure. Where are we?”
“Alongside Van Cortlandt Park,” said Katz.
We’d gone farther than I recalled. Must have been a dilly of a seizure. As I’d grown older, they’d got milder. Usually I’d be out for less than a minute and not go limp at all. I’d just hang out the blank-eyed ’nobody home’ sign for a bit, then pick up more or less where I’d left off.
“Are you, I mean, d’you need …” began Katz, all brow-furrowed concern.
“Fine. No. Let’s get moving,” I snapped.
After a moment, Clay got back in, and Katz pulled out onto the Expressway. Her eyebrows were stuck high on her forehead for a while.
A couple of minutes of silence later, I regretted my snappishness. These were my mission partners, after all. The snappishness was an after-effect, like the tingling of my skin. A hangover from the crashing electrical storm in my brain. The sour taste of ozone in my head.
“It’s nothing to worry about,” I said. “Happens a few times a year. When I was young, I got some bruises, falling. But that doesn’t happen now. Sometimes, I don’t black out at all.”
Katz glanced at me. “What you wrote … I know people who try to write down their dreams …”
“That wasn’t sleep and I didn’t dream. It’s a different process. An epileptic seizure is a cascade of neural firing in a region of the brain, caused by a chemical imbalance. ‘Electrical storm’ isn’t a bad analogy. My condition inspired me to study neuropsychology, and it’s why I don’t drive.”
Katz nodded. “And writing things down?”
Another after-effect was that I wandered off track.
“In my case, the epilepsy involves a part of the right hemisphere of my brain that turns memories into insight. When I experience the fugue—the seizure—I sometimes make connections I wouldn’t have made otherwise. But the insights, if there are any, fade quickly. And because language skills are centered in the left hemisphere, but the storm is centered in the right, I find it hard to articulate whatever insights I gain. So I’ve learned to jot down or draw the first things that pop into my head after an episode.”
“Happens to me all the time,” said Clay from the back, in a jocular tone. “I solve my cases just before I wake up. By the time I hit the shower—gone.”
I rubbed my neck and rolled the stiffness from my shoulders. I didn’t feel like responding to Clay or saying anything more. We drove on in silence. Glancing back, I noticed Clay was slouched with his arms folded and eyes closed, like he was napping. I didn’t know him well enough to say for sure, but he seemed a lot less enthusiastic about our trip than Katz. Bored, almost. Of course, he was a lot more experienced.
My own peevish impatience faded, and I felt sorry that my little episode had damped Katz’s earlier ebullience. I knew such things were a downer for people who weren’t used to them.
“I’m sure you’ll be able to meet Krome,” she said quietly, glancing at me.
“Yes. Does Pythia … talk?” I wasn’t sure who I meant when I’d written ‘Talk to her.’ Katz, Alex, and Pythia were the only ‘hers’ that came to mind, and I’d been talking to Katz all along. That left Alex and Pythia.
Katz grinned lopsidedly. “You mean, with a sexy voice like Samantha’s in Her, or like the robot’s in Ex Machina? Or maybe a sinister one like HAL’s in Space Odyssey?” She lowered the pitch of her own voice, and softly intoned, “I’m afraid I can’t let you do that … Dave.”
“I don’t know. How does Pythia communicate?”
“It opens an online case file and posts the alert in an internal chat group, so we see it immediately. But seriously, Pythia has the personality of a toaster.” Katz glanced at me with a little smirk. “That’s why I only call it ‘her’ to annoy Daffy.”
“What does she—it—typically say?”
This time, Katz made her voice toneless and robotic. “New case opened at ten oh five Monday July 5^th^ based on social media traffic analysis. Subject James Arklow of blah blah, New York. Majority prediction of antagonistic action against subject with probability of twelve percent, margin of error of—”
“Okay, I get the idea. Why ‘majority’?”
“Pythia’s not just a single neural network. She’s a committee of twelve Pythias, twelve neural networks that have learned from different data sets. They’ve had unique life experiences. Computer scientists call the committee members ’experts.’” Katz glanced slant-eyed at me. “Krome says the temple of Apollo at Delphi had a dozen trained priestesses, and it wasn’t always the same one who served as the oracle. The rest mingled with pilgrims, learning the gossip and politics of the day. Whoever acted as the oracle would speak for all.”
My eyebrows went up. I definitely wanted to meet this guy.
Katz continued. “All the Pythias—the trained experts—monitor the same social feeds and each one comes to its own conclusions. At the end, they report the consensus opinion. But individual Pythias might disagree.”
“Like in the movie Minority Report,” I said, channeling Katz.
She flashed a bigger grin. I was glad to see her cheerfulness return.
“Yeah, kind of, except if one of the Pythias gets pulled out of the committee, that doesn’t stop it, the way it did in the movie.”
“Why would one be pulled out?”
“That’s how Alex’s programmers upgrade the system. They replace an older, underperforming Pythia with a newly trained one, maybe with a different neural architecture. There are usually twelve Pythias online, but they all … speak with one voice.”
“Twelve. Like a jury.”
“Yup.”
“And these twelve Pythias learn their jobs just by reading social media?”
“Oh, no. An algorithm that learns only from social media data can discover interesting patterns, but that’s all. Krome wanted more. He wanted a system that predicts crime. So each expert—each Pythia—is trained on a huge corpus of social media data, time-aligned with the FBI’s crimes database. That way, the Pythia can learn what social media traffic predicts actual crime.”
“Supervised learning, right?”
Katz gave me a sparkly grin like I was a kid who’d got the right answer in class. “That’s right. But until Lohman, it didn’t work. The problem is that we can attribute very few crimes to social media chatter. The chatter might foreshadow or even influence, but it doesn’t cause. Remember the Slender Man stabbing?”
I nodded.
“Well, the scientists Krome hired tried to train a Pythia that would have predicted the stabbing based on the Internet traffic about Slender Man that came before. They never succeeded.”
This time I grinned. “Yeah, I can see how that would be wasted effort.”
“But maybe the Frost Angel traffic is different, because it’s targeted at individuals.”
“I thought you said Pythia learns by competing in simulated bad-actor scenarios. The GANs thing.”
Katz glanced my way. “Right. But like I said, there aren’t enough actual crimes.”
I made a skeptical grunt.
“Sounds weird, huh?” continued Katz. “But there’re hundreds of billions of social media messages, and only a few hundred crimes with documented links to social media, and many of the links are tenuous. Really, when you pare it down, there are only a few dozen examples like Pizzagate and El Paso. That’s not enough for Pythia’s neural networks to generalize from. The solution was for Pythia to fantasize a lot of crimes. But to keep the made-up crimes plausible, they’re rooted in actual crimes from our files. Like deepfakes of crime, see? That’s one place where GAN’s come into it.”
My head was hurting, and not only from the fugue. “So these Pythias are also master criminals?”
“Nope. The twelve regular Pythias only try to predict crimes. They’re the ones that get to listen to social media traffic in real time. There’s a thirteenth neural network—a thirteenth Pythia—that’s responsible for inventing the bad-actor scenarios. It can’t connect to the Internet, only to historical archives. While they’re being trained, the twelve Pythias compete against the thirteenth. That’s the ‘Adversarial’ in ‘GAN’.”
She gave me the sideways-smirk again. “Number thirteen is like the system’s dark imagination. We sometimes call it Darth Pythia.”