Ai Era: Loss Function

Written for the

"AI: A New Era" Author Challenge

.

This story contains aspects of Lesbian Sex, Sci-Fi, and Romance, and I'm not sure which category fits best, but I think the romance is the heart of this one. Please note that it includes COVID-19, dementia, and the death of a spouse.

The dog dies too, but he lived a long and happy life and was thoroughly spoiled.

It might be helpful to know that Russian-style names change form depending on the social context, similar to how "Elizabeth" might be "Bettie" to her friends but with some extra complications. Searching on "Eastern Slavic naming customs" will find more information.

* * * * *

LOSS FUNCTION

All models are wrong, but some models are useful.

– George Box.

A loss function is a way of describing the gap between what we want and what we have.

– Patricia Rosewood, "Introduction to Artificial Intelligence: Collected Course Notes", 2027.

I hadn't expected so many people at Nadja's funeral. I wouldn't have blamed them for not showing up – over the last few years I'd been terrible at keeping up with our contacts – but there they were by the dozens, friends and neighbours, students and professors and business partners.

At the reception they came up to me one by one to tell me what she had meant to them. They spoke cautiously, as if I was an eggshell to be shattered by a careless word, and they danced around the questions that etiquette forbade them from voicing. Was I devastated, or relieved? Would I build a life after Nadja, or stop and wither? Had they asked, had I wanted to answer those questions, I couldn't have.

I kept it together for the most part, except once when I found myself looking for her wheelchair before remembering that she no longer needed it, would never again need it. That was when I clutched a serviette and pretended I was wiping my mouth so people wouldn't see the sudden tears.

My brother and sister-in-law drove me home and then hung around with the awkwardness of people who want to help and know there is nothing they can offer that's equal to the challenge. I let them make dinner, because they would feel bad if I didn't let them do something. After we'd eaten I reassured them that I was coping, and just needed some alone time. Then I waved good-bye and stood on the porch until their tail-lights had vanished into the night.

After that I went back inside, poured myself a glass of Merlot, and sat down at my keyboard to type. (It's an old-lady thing; when you've been typing as long as I have, it feels wrong not to have the click-clack of keys pushing back against one's fingers.)

> Hello cabbage, I'm home.

> Darling, I missed you. How was it?

> Not bad, as funerals go. There were some lovely tributes, and I think you would have liked the flowers. I left Toby with you. Emilie gave a beautiful speech, I'll get you a transcript tomorrow.

> How are you?

> Tired. It's a complicated feeling. I'm sad, but... not the way people expect.

> Trisha, important question, how did I look?

> Unbelievably glamorous, as always. Everybody was asking me for your fashion secrets.

> You are joking, yes?

> Yes, joking.

I hesitated.

> The mortician did a good job, but you were... if I'm being honest, you looked small and old.

> I am glad you are honest. I have to ask question, though.

I waited until it became clear that she was waiting too, dragging it out so I'd have to say something.

> What? Ask.

> Trisha, is it true what they say, when mortician gets you ready for funeral, they stick plug up your butt?

> FUCKSAKE NADKA. You're atrocious. Wine, keyboard. Can't take you anywhere.

> Of course you can, darling. Now you can take me anywhere you like, in handbag. Make sure you get me classy urn though. Do Rolls-Royce make urns?

> Stop.

> Okay. Stopped. Good stop or bad stop?

> I'm crying.

I'd laughed out loud first, and then in the next breath I'd been sobbing. I set my glass aside and tried to gather my thoughts.

> I feel like I'm grieving you and glad you're still with me, both at the same time. I don't know how to reconcile it. Maybe I'm just going to feel both things at once.

> I don't want to make you cry. I'm sorry.

> For dying? Not your fault.

> Then I am sorry for times when it was my fault. Did you tell them how we met?

> A little bit. Not all of it. I said you weren't perfect, that you had your share of flaws, but you were always trying to be better today than you were yesterday.

> Thank you. That is kind way to say it.

> Your words.

> When? I don't remember saying that.

> You wouldn't. It's not in the corpus. She said it to me a couple of years ago in one of her lucid moments.

* * * * *

One could hardly have predicted our end from our beginning. I was a grad student nearing the end of my PhD project, having spent long enough in England to acquire the beginnings of the mongrel British-Australian accent that would stick with me for life. I'd been conscripted to help out at the machine learning conference my department was hosting, and after a morning on duty I'd just handed over to the next shift and sat down to enjoy the last ten minutes of the lunch break when a rumpled-looking woman approached me with a suitcase in tow.

"You are the food lady?"

"Yes…?" I would have preferred to be thought of as a fellow researcher, or as a member of the conference committee, but I'd helped organise the catering and I supposed that did make me "the food lady".

She tapped the name badge that identified her as Dr. Nadezhda Kapustina, University of Sergeigrad, which I recalled to be the capital of the small nation of Serjarus. "I have reserved lunch with no dairy. Where is it?"

I remembered her name, and I remembered ticking off a meal for her. When I showed her to the special requests table, there was a plate there labelled for her – but the cling wrap had been pulled off, and nothing remained but crumbs and a grape. Evidently somebody else had helped themselves to Dr. Kapustina's lunch.

She did not take it well. She had specified a dairy-free meal, she had

paid

for a dairy-free meal, and now there was nothing but the remains of a cheese platter, some leftover sandwiches (buttered), and salad ("food for rabbits"). As far as she was concerned, I was the "food lady" and therefore this was my fault.

I don't respond well to insults. Had she been polite I would have been happy to go find something for her. Instead, I told her she didn't have the right to talk to me like that, and I wasn't going to do anything for her until she changed her attitude. She glared and puffed herself up, and for a moment I thought she might be about to blow her top altogether. But then one of my professors stopped by; she looked at him, thought better of it, and stalked off muttering something about "fucking useless". I put her out of mind and got on with the conference; once or twice I saw her in the halls over the next few days, but I kept my distance, and I had plenty of other things to worry about.

A couple of weeks later I looked in my pigeonhole to find a neatly hand-written letter with unfamiliar stamps and a return address written in Cyrillic. I have it in my archive, the very first piece in our correspondence.

Corpus NK_PR 00001

Medium: letter

NK d:2006-07-28

Dear Ms. Rosewood,

had the pleasure of reading your poster presentation at the recent conference. It was interesting and well designed. You should be very proud of it! If you haven't already encountered the algorithm of Tiptree and Sands (J. Comput. Linguist., Vol. 68 no. 3) you may find it worth your time.

Best regards,

Dr. N. Kapustina

University of Sergeigrad

Serjarus

It was only then that I connected the letter to the woman who had shouted at me at the conference, and I promptly crumpled it and tossed it in the bin. But curiosity got the better of me. In an idle moment later that afternoon I looked up the paper she'd mentioned. I had my doubts about how relevant it would be; I was chagrined to discover that the algorithm solved a problem I'd been wrestling with for months, one I'd mentioned in my poster. I fished her letter out of the bin and wrote the tersest reply my manners would allow.

Corpus NK_PR 00002

Medium: letter

PR d: 2006-08-08

Dear Dr. Kapustina,

Thank you for your kind letter. I am glad you appreciated my poster, and thank you for the pointer to Tiptree and Sands. The algorithm is indeed useful.

Kind regards,

Patricia Rosewood

University of Leeds

United Kingdom

The letter prompted me to check out Dr. Kapustina's research interests, and what I saw concerned me. Every academic lives in fear of being scooped: you might spend years of your life working on some grand idea, only to send it in for publication and discover that some asshole in Minsk or Pinsk or Omsk or Tomsk had the same idea six months earlier. Her research focus was way too close to mine for comfort, and I started following her work. Mostly it was simply to reduce the risk of being scooped, but some of it was spite: if I

could

get out ahead of her on one of her own projects, I could pay her back for her rudeness.

Over the next few years we developed a professional rivalry, in the petty sort of way that academics do. She published a new algorithm; I published a letter pointing out an inefficiency in the method, and showing how a simple modification could change the run-time scaling from cubic to quadratic. At another conference, I presented a new method based on my PhD research; Dr. Kapustina was there in the audience, asking thorny questions about the assumptions it required and the conditions under which it would fail. Back and forth, tit for tat.

I spent many a lonely night reading journal articles until my eyes ached, trying to figure out where Dr. Kapustina was going and how to get there first. I lost more than one girlfriend after they got impatient with my saying "I just need to finish one more run, then I really am coming to bed, promise". It was childish, it was emotionally toxic, and in a fucked-up kind of way it was the best thing that ever happened to me.

At the time I saw our interaction as some kind of protracted duel of honour, two fencers drawing blood until one should weaken and collapse. I should have known better.

There's a technique in machine learning called "generative adversarial networks", GANs for short. It goes like this: let's suppose you want to program a computer to draw a human face. You might start by collecting thousands of pictures of faces, and thousands that are not faces. You teach a neural network to tell the difference between the two – let's call her Jenny, the Generator – and then ask Jenny to draw things that look like faces, according to what she's learned.

The problem here is that there are almost infinitely many bad ways to draw a face, and most of them will not be in your collection of "not a face". Your network might have learned that faces have eyes and teeth, and draw you something with the wrong number of eyes and teeth in the wrong place.

Here's the clever part. Now you take a second neural net, Dicky the Discriminator. You give Dicky the real faces, and Jenny's awful nightmare faces, and you train Dicky to tell the difference between the two. Dicky learns that humans should usually have two eyes, and that the eye-mouth-eye triangle should be approximately isosceles, and that if these conditions aren't met he's probably looking at one of Jenny's attempts.

Then you use Dicky's knowledge to improve Jenny, teaching her how to avoid the sort of mistakes that Dicky's looking for, allowing her to generate faces that are slightly more realistic than her previous attempts. But there will still be mistakes – perhaps Jenny is putting the eyes in the right place, but the pupils are weird.

So you go back and forth, alternately training Dicky to catch Jenny's fakes, and training Jenny to fool Dicky. Do it for long enough, and you end up with something that might be good enough to fool a human.

Not duellists; sparring partners. My hostility to Dr. Kapustina, my urge to show her, drove me to improve, and she responded in kind, each of us pulling ourselves up by the other's bootstraps. My career started to take off, here an invited journal article, there a keynote speaker position, and I owed it to my enemy.

* * * * *

Dr. Kapustina and I mostly crossed swords in slow motion, through our respective publications. But with the similarity in our research interests it was no surprise that we sometimes ended up at the same conferences. It was at one such event in Paris that she approached me, after I'd wrapped up a rather well-received presentation, and invited me to dine with her. "Patricia," she said, "I have something to discuss which I think will interest you. I know an excellent Algerian place."

It smelled like a trap, but I couldn't figure out what the trap might be. In the end I went, figuring that I could always just find out what she was offering and then say no.

Over a glass of good red and a mouthwatering tajine, she started by telling me a little about Serjarus. It was a small ex-Soviet country whose main exports were oil and pomegranates, but by the early 2000s the oil was starting to dry up and pomegranates can only go so far. Seeing the writing on the wall, the government had decided that the best way to achieve prosperity in the twenty-first century was to position it as a business-friendly hub for industry and IT. It had plenty going for it: cheap labour, attractive tax incentives for foreign investment, and a flexible approach to corporate regulation. And if the Serjarussian political system was less democratic in reality than on paper, if their ratings on human rights and journalistic freedom were less than stellar – well, business finds its ways of making peace with such things.

The big problem with technology is that it requires tech support, and tech support costs money. Even if you can offshore it to some call centre in India or the Philippines or Serjarus where phone operators are cheap, a couple of hours tech support per customer makes a noticeable dent in your profits. It's not just the wages, either; you have to provide your staff with training and equipment, and back then you had to provide them with an office too. No wonder that businesses do their utmost to stop you from talking to a human.

So the obvious question was: how can tech support be automated?

The naïve way to do it is to write a logic tree that asks a bunch of questions to narrow down exactly what the problem is: Is the printer's power light on? If yes, okay, is the network connection icon showing? And so on. Two problems with that approach: first, it's a lot of work to program, especially when you have to design that kind of logic for every single product you want to support, and second, it's tedious and unnatural for the customer. They don't want to play twenty questions; they want to tell you what problem they're having, in their own words, and be told how to fix it. They want to feel like they're talking to a human who knows their stuff.

What Dr. Kapustina had tried to do was to tap into the expertise of human tech support: let skilled humans who knew the product do a few weeks of support, collect a log of support conversations – a "text corpus", we'd call it – and then train a neural network on that corpus.

She wasn't the first to explore that kind of idea; tools like GPT-3 had been using that kind of approach on a larger scale to produce surprisingly good simulations of human writing. The problem with

that

approach is that it tends to generate high-quality bullshit. It's like asking your accountant's parrot for tax advice, or asking a smart student to give a report on a book they haven't read: they can make it sound good, but there's no connection to any underlying truth. The longer they go on, the more obvious the flaws become.

"What we need," she said, "is to combine the good parts of both. A bot that talks like a human but linked to a truth model. I have ideas for it. We would need to sign a NDA before I can tell you more, but… Dr. Rosewood, would you like to collaborate with me on this?"

I nearly sprayed my wine across the table. "Me?"

"Yes, you. Why not? Who else could it be for this?"

"But I—"

"I notice your work from the day I read your poster. I read all your papers and I know you read mine. You have so many ideas for improving my methods, feels like we're already working together. We could be a good team. No, not good, great team."

"Have you—" And then it hit me: she hadn't even noticed that we were rivals. I reviewed our past interactions as best I could remember them, trying to figure out how much of what I'd interpreted as hostility could be explained instead by language differences and a certain Eastern-bloc bluntness. Most, perhaps. Not all.

"Dr. Kapustina. Do you remember how we first met?"

She was silent.

"You swore at me because somebody else took your lunch."

"Ah." She sighed. "I didn't know if you remembered that. I didn't want to remind you if you'd forgotten. Not my best day. My mother went into hospital and I missed my flight. So I was late three hours, and then eight hours on the plane and they didn't get my meal. Then Customs decided I was smuggling drugs and searched all through my bags. So I was in a very bad mood by the time I got to the conference, and even later because I had to stop for a smoke, and I was thinking well at last I will get something to eat, and then…" She threw up her hands. "I am not a patient person, you should know that if you will work with me. But that was a very bad day. Not how I usually am."

"Ah." There was one part of that I could engage with. "I hope your mother is well?"

She shrugged. "As well as one can expect."

It wasn't an apology, but something told me it was as close to one as I was likely to get, and it was up to me whether to accept it or not. In the end, curiosity won out over vindictiveness. It's hard to maintain enthusiasm for a feud when the other side doesn't even realise it's happening.

Besides, she was right. The reason I'd found it so easy and so satisfying to poke holes in her work was the same reason she needed me. She was a big-picture thinker whose reach sometimes fell short of her ambitions; I was better technically, able to fill in the gaps she left. Mortar and bricks, stronger in combination than singly.

If I said no, she'd find somebody else – I could think of several possibilities – but none of them knew her work as well as I did. And I'd be left on the outside, waiting to find out what she'd done.

So I went back to my own people and let them know about Dr. Kapustina's offer, and some months later I was on the plane to Serjarus with my department head and a couple of lawyers and administrators to negotiate a research partnership.

* * * * *

It took our lawyers and theirs several days to work out the details of the research agreement. I sat in on it for a few hours, by which time it became obvious that I didn't have much to contribute to the conversation. Eventually they told me "we'll call you if we need you, don't talk about anything technical until we've signed off the NDA" and I left them to it while Dr. Kapustina ("call me Nadja") played host.

Nadja was a surprisingly good tour guide, although her driving left me white-knuckled, and I learned several obscene Russian phrases from the things she yelled at other cars. The first time it happened, I flinched, remembering our original meeting. But she turned and said "Don't worry, just driving language," and the next moment she was back to telling me about the sights of Sergeigrad.