Chapter 52 — Teaching Wassily
The Curve of Time, Chapter 52 —— Teaching Wassily, in which Saskia takes on a student.
Followed by Rufus’ thoughts on machines learning.
Listen to full episode :
— 52 —
Teaching Wassily
Wassily looked at Saskia. “So, where have you gone?”
Saskia shrugged sheepishly. “I won the Lottery.”
“I mean, sure,” Wassily brushed her confession aside. “But what was it like? Have you ever felt yourself splitting apart? Branching into different timelines?”
Saskia said she had no idea if she’d split at any point. No more than she had any idea if the world was or wasn’t merely four-space.
“Wait,” Mica stopped her. “Earth is just a sphere.”
“Right, but the earth-sphere sits in three-space——three dimensions——and it’s possible that adding time just means you can fast-forward or rewind through another dimension, making our space-time universe four-space.”
“OK.” Mica nodded her head, comprehending. “So how could it not be four space?”
“Well, if you wind far enough forward through time, maybe you get back to the big bang.”
Wassily jumped in: “You might have to go forward a billion years. But just because we can’t test it, doesn’t mean that doesn’t take you back to the start of the universe. It’s like an ant going west and hoping to come back to where it started; it’s unimaginable that the earth is a sphere from the ant’s perspective——”
“You and your ant analogies!” Mica interrupted him.
Wassily grinned. “Unless, of course, the ant crawls onto an airplane, or a rocket.”
“But we can’t see the time dimension,” Mica objected. “I can look up in the sky. See the clouds.”
“We can see a little way,” Saskia corrected her.
“No, you can!”
Saskia shook her head. “It’s called memory.”
“Actually, everything you ‘see’ is the past,” Wassily piped in. “It takes light time to reach your eyes.”
Mica ignored Wassily’s annoying pedantry, preferring to object to Saskia that memory wasn’t the same as looking out and seeing.
Rather than seem chastened, Saskia reminded Mica that our brains were just blobs in boxes, and that accessing memories from storage looked pretty similar to accessing vision from your visual cortex; in neither case was it a direct experience.
Wassily nodded. And, apparently appreciating Saskia’s argument, he added his own note: “We don’t exist at a constant moment, anyway. Even if we don’t have control over our movement in the time dimension, the way we do in the other dimensions.”
“Even for me,” Saskia picked up his thread. “I can’t move around that much. It’s like you guys are sitting on a ‘time train’ whereas I can get up and move about the cabin. As soon as I stop consciously slipping in time, though, I just move like you again. Maybe I’m ahead or behind you on the tracks, but not by much.”
Wassily focussed on Saskia again. He asked her what it felt like to slip through time faster than the ordinary flow.
Saskia said it felt like a blur. “And then, when you stop——you know how sometimes when you’re driving and you zone out, and then you arrive where you’re going? And it’s hard to remember getting there. Once you’ve stopped moving through time at high speed, it kind of feels like that.”
Wassily posited that perhaps the mind had trouble processing the world around us at anything other than the normal flow of time, which, in turn, reminded Mica of the time dilation experiment Eagleman had done. But before she could add her thread to the conversation, Saskia had taken it in another direction, noting that her phone battery didn’t feel the dilation of time: it ran down at the speed of the world around her.
“Well, it doesn’t matter if you’re carrying your phone east to west, or west to east,” Wassily rejoined, “your phone’s battery still runs down. So, if time really is just another dimension, why should that be any different? Backwards or forwards in time, the battery is always dying. Fits with bastard theory.”
Saskia cocked her head and squinted at Wassily.
“Bastard theory,” he clarified, “the extension of Murphy’s law, that says, not only will something go wrong if it can, it’ll do so in the worst possible way.”
Saskia nodded her head with a comprehending smile. “There’s a logic to that. Going at twice the speed would obviously kill the battery twice as fast.” Then a pall crossed her face. “Do you think my own life has been slipping by faster too?”
“How long does it take to feel hungry?” Wassily asked.
Saskia’s eyes popped wider as she considered. “I have no trouble skipping meals when I’m slipping through the world at speed.”
“So, it’s a relative effect,” Wassily shrugged.
Once again, Mica felt excluded from the conversation. It clearly wasn’t malicious intention, but she didn’t like it anyway. She turned to Wassily. “Why are you here? What was it you came to tell me?”
Wassily looked back at Mica, as if surprised by her intrusion. “Well, the way you described time travel to me——it seems like it would also imply an ability to teleport.”
Saskia laughed, but Mica turned to her. “You didn’t tell me you can teleport.”
Wassily turned back to Saskia: “You’ve tried it?”
“Teleport? Sure. And fly.”
“No,” Wassily responded. “Teleporting is just moving about with time stopped. You can be everywhere, all at once.”
∞
Saskia felt the obviousness of Wassily’s idea like a slap in the face. How had this not occurred to her? But then, despite determined efforts, she couldn’t hold time still. She got close. Invariably though, she’d inadvertently switch the march of time backwards, slipping one side, or the other, of static time. There was something unstable about her execution. It was like a marble rolling down a banister; there was no way for her to hold that path. And at the point she fell, she’d disappear from Wassily and Mica’s view.
The trepidation of jolting herself, á la how she’d done so back in Mica’s bed, was a prospect she feared instinctually. It bestowed on her a presence of mind to move from where she’d been standing when she turned time about. When she returned to Wassily and Mica, she invariably arrived just before she’d left. It gave the appearance of a sort of haphazard teleporting, though Saskia knew it was only the appearance.
More significantly, whatever Wassily thought about teleporting, she was obviously blowing his mind with her actions. Saskia recognized it in much the same way she’d recognized it on Mica’s face when she had tried taking the ant back in time with her in Mica’s kitchen. Wassily was immediately convinced that there was a there there.
“Why don’t you teach Wassily?” Mica proposed, pushing for her own agenda.
Saskia couldn’t tell if Mica was annoyed at her, or if she really did want to take the burden off Saskia. She’d been put on the spot, which she didn’t like, but maybe it had been meant kindly.
Mica and Wassily both watched her expectantly. Still, Saskia took her time, wrestling with what to do. Eventually, she reasoned that if she were going to try teaching anyone——Wassily was as good a candidate as any. There was no one she trusted much more. Might as well take a flier on him.
It was also as an olive branch to Mica, who she recognized that she’d upset.
Wassily was visibly giddy at the prospect, and professed as much.
“Well, sure, you’d like to slip in time,” Saskia needled him gently, as much in an effort to patch things up with Mica.
“Well you got what you wanted. A bottomless bank account.”
“Winning the lottery?” Mica asked.
“That was an experiment,” Saskia objected.
“So? I want to do my own ‘experiment’”——Wassily air-quoted the word for effect——“I just want another shot at love.”
“It’s hard to go back far,” Saskia noted, significantly. “Even a week took it out of me.”
“Give him a break. He could be talking about last night.”
But Saskia knew what he was talking about. She could protest all she wanted that she’d given her money away, but it wasn’t relevant. It also raised the question of how love worked with time travel, though Saskia didn’t hold herself out as an expert on that topic, even in the present.
Wassily shrugged. “Maybe different people have different parameters.”
“I’m not so sure,” Saskia pushed back. “If other people can slip further back in time, and if it’s possible that anyone else can learn, then why haven’t we seen others? For that matter, if it’s possible for others to slip in time at all, why haven’t we seen another time traveler anywhere? Even if people can’t go back far.”
“Maybe the time is now,” Wassily suggested. “Maybe now is the time for time travel to bloom. Just because we can’t see what made it suddenly possible, doesn’t mean something hasn’t. Nobody realized the world was on the brink of understanding calculus in the mid seventeenth century, but the preconditions were in the air.”
“So, Saskia is Newton?” Mica raised one eyebrow.
“Or Leibniz, or whoever else would have discovered calculus had they not existed.” And in response to Mica’s confused look, Wassily explained: “By the late seventeenth century the ideas underpinning calculus were in the air. It was just a matter of time.”
Saskia, however, was mulling Wassily’s dashed hopes of rekindling what he once had with her. It recalled the do-over that she had permitted herself, when she first met Mica. But had anything really changed? Nothing that she knew of had changed when she went back to confront Zeno in Texas, either. Suddenly, a thought pressed insistently and she blurted out: “There is only one world.”
No sooner had the words left her lips, however, than she recalled her other self. The one she’d seen outside the bathroom at Cleo’s. If there was only one world then where had that other version of herself come from?
Whatever the case was, both Wassily and Mica were keen for Saskia to try teaching him what she could, and, so, Saskia obliged them. What none of them realized, was that Wassily was too concentrated to take on board how one slipped in time. Saskia did her best to describe how it felt. What to look for. But nothing changed. In the end there was only the struggle.
Unfortunately, slipping in time was like putting in golf, or playing a drop shot in tennis. There were multiple elements at play, and the worst thing you could do was to concentrate too hard. An experienced painter can, with a fluid flick of the hand, strike a line that defines the edge of a face in a portrait, but the beginner, trying too hard, tends to overshoot or clam up altogether, delivering a shaky edge, a consequence of slow and too deliberate.
An hour after they started, Saskia was exhausted. Defeated. Moreover, though her primary reason for trying had been to appease Mica, Saskia couldn’t help feeling that Mica held the failures against her. As if Mica felt she’d been deliberately floundering in her attempts.
When Mica’s cat, Fish, dragged his tail past her leg as she sat on the couch, Saskia had had enough. She looked up at Mica. “I miss Tomato. I think I’m going to go home.” She turned to Wassily. “It was nice to see you again.”
That was chapter 52, Friends, I hope you enjoyed it!
For those binging the story, you can now skip forward to chapter 53, or you could join the rest of us and listen to me muse a minute longer.
Teaching has been an important part of my life, it is the flip-side of learning. It can happen one-on-one. Or, in a first year university calculus class, it can happen five hundred-to-one. On an even grander scale, with the advent of books, and now Youtube channels and the like, it can happen millions-to-one.
At the other end of the scale, at the risk of confusing teaching with research, it can happen zero-to-one. That is, sometimes it is not so much another person who teaches us, but nature and the world around us. The story might be apocryphal, but it could be argued that it was the apple tree that taught Newton about gravity. Or even a single solitary apple.
Since our story has touched multiple times on machine learning, and LLMs in particular, and since the way they are trained (aka the way they learn) is oft in the news, I’d like to take a moment to toss my two pennies into the ring on how the next few years are likely to play out. Bear in mind that I am certainly not an expert on this subject, but in my defense, much of the dialogue surrounding how this will play out is based on both intuition and first principles thinking.
There is a school of thought that contends that what makes LLMs powerful is that they are trained on what effectively amounts to the corpus of human knowledge and thought, collected via words on the internet. That it is our human interpretations of the world around us——synthesized into words which make it ingestible by the machines——that gives LLMs the illusion of thought. Adherents to this view of the world suggest that since we have, more or less, run out of new “training data” we are likely near the apex of model power. True, a few in this tribe, believe that synthetic data——text generated by the LLMs themselves——might offer some small extension of teaching capacity for our models, but most don’t believe that that extension will take the models much further. To believe that it would, kind of amounts to believing in a perpetual motion knowledge machine.
There is, however, another camp. One into which I find myself more readily leaning. That camp says that the path to the next generation of models will not so much be through the billions-to-one learning model we’ve talked about thus far, but rather turning the paradigm upside down. This camp believes that the next leap in artificial intelligence will come through models learning directly from the world around them.
Critics of the second camp contend that it is the human understanding of the world around us that is important, and the true value of what the LLMs have gleaned from all the text they have ingested. In some sense, what they are suggesting is that it is our human interpretations of events that are paramount. In the language of model architecture, the structure of the text we feed the models makes it relatively easy to design an objective function from which the model learns (recall the objective function for an LLM was the computer checking its guess for the next word against the actual next word of the text it has been fed).
To me, though, this misses two points. Firstly, the newer multimodal models are now in a place to learn directly themselves, substituting the prediction of the next word of text with the next frame of video image, or the next sound, or maybe even the next smell, of the world they are observing around them. Indeed, the sensors on Teslas have already amassed a trove of such data very specifically of use in predicting how to drive. Of course, sometimes we still have to intervene and help the autonomous driving programs with edge case, but that just brings me to the second point.
Secondly, though human knowledge has been hard won through millennia of collective human experience, machines can aggregate far more data than any human ever could; more data than any collective of humans ever could. In fact we’ve seen trivial examples of the power of this sort of massive data absorption already, with AI assisted models catching tumors on x-rays that humans never would.
Anyway, my two cents says that running out of text might actually be a really good thing for model training: it’ll remove the training wheels, as it were, and let the models start inferring things that we have struggled to see ourselves. Of course, this does make so-called alignment of models even more important, but that’s a topic for another week.
Until next week, be kind to someone and keep an eye out for the ripples of joy you’ve seeded.
Cheerio
Rufus
PS. If you think of someone who might enjoy joining us on this experiment, please forward them this email. And if you are one of those someone’s and you’d like to read more
