Chapter 44 — The Brain Specialist
Rufus considers how AI will impact society and jobs.
Followed by Chapter 44 —— The Brain Specialist, in which Saskia talks to a brain specialist.
Listen to full episode :
Hello Friends,
Today’s chapter picks up in the hospital where we left Saskia last week. It also picks up on some of the machine learning that she touched on earlier.
It’s funny to me, when you start looking for it, the myriad contexts in which machine learning crops up. Patient care is a well documented one, what with AI X-ray readers gradually becoming the gold standard first pass. Moreover, the potential to give those who currently survive without any access to doctors some form of interactive medical care stands to revolutionize healthcare for great swaths of the planet.
But it goes well beyond medical care. Indeed, listening to the Possible podcast the other day I learnt about JusticeText, which aims to benefit another poorly served segment of our society. As Devshi Mehrotra described, her AI software startup, has developed resources to help public defenders sift through the mountains of case materials that they are often presented with; materials they rarely have time to review with the care they would like.
For impacts that affect those of us at the more privileged end of the spectrum, art is the canonical example of a field about to be upended. With this in mind, it was fascinating to visit the Warhol museum in Pittsburgh a couple of weekends ago with some artist friends of mine. Particularly thought-provoking was to engage with one artist who felt they were collectively on the front line of job-loss, and ironic too, given we where walking through rooms celebrating another artist who had no qualms about appropriating other people’s work. I understand my friends’ fears, but as I noted to him, getting image generation tools to do what you want is far from simple; it requires skills, in the same way that photography——also once frowned upon——requires skills very different to those of painters. I certainly struggle aplenty when using Midjourney to help generate the episode artwork for this podcast, invariably at least half of my work is done later, in Photoshop. Wherever one comes down on the question of appropriation, none of us in Pittsburgh were able to leave the museum without feeling that Warhol himself, were he alive today, would certainly have been among those to embrace these new tools.
In the end, I remain of the opinion that generative AI is not so different from the excavator that makes the manual labor of trench digging less arduous. That, even if, art is “somehow different”, as was asserted to me. It should also be noted that human ditch diggers are not completely redundant either, their roles have just changed a little … excavators still require operators.
One other hypothetical that always gets me thinking is to consider what it would mean for LLMs to be able to write the next great novel. We can only absorb so many stories——there are, after all, competing interests for our attention——and some will always resonate more than others. Even without a magical story-generator, I regularly start novels that I fail to finish. So, where does that leave us? As an author, I suspect I’m committing blasphemy when I say this, but, as a reader, I’m kind of excited by the prospect that there might be more to choose from … and/or that maybe the AI might help to sort out which books are worth my giving a try to?
Anyway, one last note of commentary before turning to today’s chapter: the guts of Saskia’s explanation of LLMs today are a refining of an email I sent to an old grad school roommate who was curious about how the chatbots work. One piece in that email that didn’t make it into the chapter was a fascinating interview with Geoffrey Hinton (sometimes described at the godfather of AI) where he describes, among other things, the difference between how neurons in our brain work and those in neural networks do. Definitely worth checking out.
Oh, and for those of you listening along on Spotify, apparently you can now comment on episodes and podcasts, and doing so definitely helps raise visibility. So if Spotify is your listening app of choice, please chime in once in a while when you’re feeling generous, I would be incredibly grateful!
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
And now, without further ado, here’s chapter forty four in which Saskia talks to a brain specialist.
— 44 —
The Brain Specialist
Doctor Brian Siebling was a cheerful man from Boulder Colorado. His easy way with conversation had, at first, made Saskia wonder if it was a way of putting off the delivery of bad news, but it seemed that he was just genuinely personable. Now, after a sequence of narrative links that Saskia could no longer recall——all natural in the moment——he was launching into another brain-related anecdote: “Do you remember that white and gold dress that was actually black and blue?”
“It was always black and blue to me,” Mica avowed.
“But you recall the confusion,” Dr. Siebling insisted. He explained that that was the point; different people saw different colors, even though the pixels in the image were what they were. “It’s because our senses don’t just perceive the world around us blindly——pardon the pun.” The doctor smiled at his little joke. “It’s the same reason photographers shoot a white board to color balance later.”
Saskia’s eyes shot side to side as the penny dropped for her. “It depended on whether you thought the photo was taken indoors or outdoors?”
“You may have sustained a bump to the head, but you’re sharp.” The doctor touched a finger to the outside of his nose as he turned to Mica and Miranda. “From the context it wasn’t clear whether the photo was taken indoors or outdoors. Nor, if the dress was in the shade or the light.”
Saskia couldn’t help herself: “If you thought it was white and gold, that was because you mistook where the photo had been taken. You assumed that it had been taken outside. And your brain color corrected the blue hew as white.”
“Our perceptions are only an interpretation of what’s in front of us,” Siebling enthused. “Our brain merely interprets the stimuli its given. And its whole life-goal is to create a model of the universe around it.”
Miranda followed the doctor’s eyes as they spiraled about the curtained bubble they were sitting inside of.
“Your brain is just a blob trapped inside your head,” Siebling concluded in triumph.
A light flickered in Mica’s eyes.
Siebling cocked his head. “Did I say something funny?”
“That’s exactly how Saskia described our brains the other day.”
The doctor turned to Saskia, “Do you remember that?”
“I was explaining how AI works.”
“How Chat GPT works?” the doctor asked, genuinely intrigued.
“Sure.” Saskia shrugged. “We weren’t talking about LLM’s but——” she broke off, turning to check with Mica.
“Coming out of a coma——it helps with recovery to do cognitive work,” Siebling interjected. He was like a child trying to sway a decision. “Think of it as exercising structured thought. Ideally, you combine it with some social; that tends to make people feel better.”
Mica shrugged. “I’m interested in LLMs too.”
Saskia was happy to oblige an explanation; after all everyone wanted to know how LLMs “did it”. She started by reminding them of what they “probably already knew”: that the LLMs are just trained——on the entire text corpus of the internet——to guess the next word. “So they answer based on your prompt and the words they have already written, adding another word at each step.”
“Right, but how?” Siebling pressed.
Saskia described her dials and knobs metaphor for neural nets. To Mica, she noted that it was pretty similar to the handwriting-reading neural nets except that rather than ten outputs for the ten possible digits, there was now an output cell for each word in the English language. “Actually, that’s not exactly how it works,” Saskia admitted. “It actually breaks the words up a bit, so parts of words like ‘un’ get their own output cell. But for the purpose of making conceptual sense——getting your head around everything——just think of the output in each cell as registering the likelihood that the word that cell represents ought to be the next word.”
“Great,” Siebling enthused. “But how do you know how to set the dials and knobs? That’s in the training, right?”
Saskia nodded at the doctor’s enthusiasm. “It’s called gradient descent.”
“Gradient descent?” Mica asked.
“Gradient is just fancy math-speak for slope,” Saskia explained. “It’s the sloping tangent plane approximation to the error function that I told you about.”
“Which means?” Siebling asked with one eyebrow raised.
While Saskia overloaded the doctor with an explanation of the notion of stepping from one setting of the dials to the next——by looking at how their configuration could be improved based on running one piece of data through the machine——Mica reflected on the similarities between brains as hidden blobs that can’t actually access the world around them, and the past as something we recall, something that informs the now, but that is just as inaccessible in the now. Except to Saskia, for whom the past was accessible. Mica gazed at the gorgeous brunette as Saskia made yet another inaccessible idea relatable.
“Can you make that more concrete?” Siebling asked.
“Sure,” Saskia paused, her eyes landing on Dr. Siebling’s name tag. “Say we input a Wikipedia article on Brian Siebling. Maybe the article starts: ‘Brian Siebling is a brain specialist.’ and let’s say our model has already had some training. This is the next text we’re going to train it with. First, it reads ‘Brian’ and then predicts the next possible word with some weights assigned to them, maybe ‘is’ has 25% chance of being the next word, ‘Eno’ has a 2% chance of being the next word, and maybe ‘Siebling’ has a 0.02% chance of being the next word.” She glanced up at Siebling, but the man seemed unfazed by the slight. “The program then looks at the next word in our training text, which is of course ‘Siebling’ and it says, ‘shit, got that badly wrong’——”
“With the error function?” Siebling checked.
“Right.” Saskia nodded. “The error function quantifies how wrong. Anyway, the thing is we have all these dials and knobs to potentially adjust and we can look at those dials and knobs and say to ourselves ‘was there a way to adjust those dials and knobs to make the error of the predicted outcome a little smaller’? And the answer, of course, is: ‘Yes!’. So now, we go back and change all those dials and knobs just a teeny tiny bit to improve the alignment——we bias towards the dials that make the biggest difference. Then we have the model guess what the next word should be and the model checks its guess against the next actual word——‘is’——and it goes through and readjusts the dials and knobs again. Rinse and repeat.”
“So the text on Wikipedia is like my handwritten digits?” Mica checked.
“Yep, it’s all training data. Different data for different models. The amazing thing is that with the LLMs——language being the way we describe the world——they learn how to describe the world, which is eerily similar to how our brains work.”
“Whoa!” Siebling put his hands on his head and let out a whistle. Before he could comment, though, the moment was ruptured by the ping of his phone, which he automatically turned to.
It was Miranda who filled the empty space: “What if the machine guesses right? If there isn’t a way to turn the dials and knobs to improve it?”
“There is never one possible word that follows any other,” Saskia said. “Say the Wiki text was actually ‘Brian is . . . ’, then ‘is’ was already the most likely next word after ‘Brian’, but the model just makes it slightly more likely now. Interestingly, it would also then implicitly learn that since it is now making ‘is’ more likely than ‘was’, ‘Brian’ is more likely to occur in present tense!”
Mica’s eyes widened. “That’s how it learns bias?”
Saskia smiled at her. “The miracle is that it learns at all. I mean, how can you be sure the model is actually going to get better. That it will spit out anything meaningful. What’s to say our training process is better than a random monkey twisting the knobs.”
“Definitely surprising to me,” Miranda agreed.
“Remember, though,” Saskia noted, “these models were built to mimic our brains’ neurons, so it’s at least plausible they might work. I mean, we were able to send rockets to the moon based—more-or-less—on physical models developed by Newton, hundreds of years ago. Computers help, but you saw Hidden Figures, that movie about the women who crunched the numbers.”
Siebling looked up from his phone. “So, it just ozmoses grammar and a world view?”
Saskia nodded. “Evidently that’s all part of understanding language.”
Siebling’s phone pinged again. “I’m sorry, I need to get back to this.”
Saskia felt momentarily self-conscious about having waffled on.
“Well, I’ve learnt something today.” Siebling pulled a tight smile. “You clearly haven’t lost much.” His eyes, however, averted to the floor.
“But . . . ?” Mica asked.
“The ‘but’ isn’t really my area of expertise. Your brain scan is normal, but——yes, there is a but——it appears that you have some scar tissue on your heart.” He glanced across to Miranda before returning his attention to Saskia. “Your mother said she was unaware you had an arrhythmia.”
Saskia met Mica’s eyes. She saw in them an echo of what she, herself, was thinking: the recollection of Saskia’s jerk upright in Mica’s bed. Both women turned back to the doctor.
Siebling, however, had his own train of thought to follow. “It’s not what we expected to find. Frankly, it’s not something you’d see after a blunt trauma to the head. But then we realized: we probably have the causality back the front.”
Saskia nodded slowly. Deliberately.
“We looked at the EKG from your last annual check up. There was no sign of arrhythmia. Can you think of anything that might have caused this?”
Unnerved, Saskia tried deflecting, “Isn’t that the sort of thing you’re supposed to be able to tell me?”
“Frankly, I’ve never seen anything like it.”
