Some interesting things to read the second weekend of September

Every now and then when I have writer’s block, I look up essays by writers I admire. The other day, I decided to read something by Katherine Boo, an alumnus of The Washington Monthly under Charlie Peters and a contributor for many years to The New Yorker. I found my way to a piece of Boo’s from 2003 called “The Marriage Cure,” about a creative program in Oklahoma for bringing women out of poverty by coaching them to find a husband. The story begins with a blazing opening sentence: “One July morning last year in Oklahoma City, in a public-housing project named Sooner Haven, twenty-two-year-old Kim Henderson pulled a pair of low-rider jeans over a high-rising gold lamé thong and declared herself ready for church.”

The whole piece is a work of art. My next favorite sentence might be this one: “Kim marvelled that Corean, who is forty-nine, seemed to know what to wear on such occasions. The older woman's lacquered fingernails were the same shade as her lipstick, pants suit, nylons, and pumps, which also happened to be the color of the red clay dust that settled on Sooner Haven every summer.” Boo’s genius is her ability to write about poverty and struggle in a way that is beautiful and evocative. It’s never condescending, and it also never seems fake. Whenever I find my way to Boo, her sentences snap me out of whatever funk I’m in and send me back to work. 

Here’s one more: “Kim keeps the things that matter to her next to her mattress, in a cardboard box stamped ‘Fragile—Eggs.’ In addition to a handmade card that her father sent from prison on her eighteenth birthday, and a tangle of blond hair extensions that her mother had mailed when Kim turned twenty-two, the box held several poems that Kim had written about the meagreness of what people around her termed love.”

I was also taken in this past month by an essay by Michael Sokolove about the conflict between Michael Oher, the former lineman for the Ravens, and the wealthy, white family given credit for saving him, as chronicled in Michael Lewis’s The Blind Side. As Sokolove writes, “It might even be that the positions taken by each side — one claiming to have been exploited, the other extorted — are both true. That would make this chapter of ‘The Blind Side,’ its epilogue, less a fairy tale of racial reconciliation and more a classic American story of money, misunderstanding and ruptured relationships.” 

Speaking of complicated relationships, I loved Elaina Plott Calabro’s story in The Atlantic about the very weird recent career of Kash Patel and Caitlin Dickerson’s incredible but devastating story about the families that try to cross the Darién Gap. And here’s a smart WSJ essay about the history of “settler colonialism.” I was captivated by this paper from the NBER that finds that bringing coffee shops to communities that don’t have them increases entrepreneurship overall. If people have third spaces where they can sit and talk to each other, they come up with good ideas and start new companies. The finding doesn’t hold for bars, or coffee shops that don’t have places to sit. If you happen to be reading this newsletter at a coffee shop, I suggest that maybe you strike up a conversation with one of your neighbors about the best idea you have for a new company in AI. 

This Ezra Klein podcast changed the way I think about the ways that young boys learn in school. I also really liked listening to Nilay Patel talk with R.J. Scaringe, the CEO of Rivian, and The Atlantic’s Jerusalem Demsas explaining gender dynamics in elections, among many other things. Meanwhile, self-driving cars may be arriving sooner than you think. And this is a truly amazing piece about how to make a button. 

I’m now deep into Yuval Noah Harari’s new book, Nexus. I’ve had a number of great conversations with him in the past. And I’ll be interviewing him Monday at the Lincoln Theater in Washington, DC. Come join! And then come say hello at the Aspen Cyber Summit (also in DC) two days later where I’ll be interviewing Lina Khan, Rohit Chopra, and Jessica Rosenworcel. And then of course The Atlantic Festival down on the Wharf on Thursday and Friday.

This section is sponsored by Elastic, the Search AI Company. Elastic enables everyone to find the answers they need in real-time, using all their data, at scale. Its solutions for search, observability, and security are built on the Elastic Search AI platform—the development platform used by thousands of companies, including more than 50% of the Fortune 500. Visit here to learn more.

During the summer, it seemed like every week brought an announcement of a new, more massive Large Language Model from the major AI companies. The most current version of ChatGPT has 170 trillion parameters that it uses to make predictions. Sounds like a lot. But some data scientists estimate that we will see a hundredfold increase in the size of these models in the next few years. What’s going on? What is the advantage of supersizing a Large Language Model?

The story in AI so far has been that bigger is better. The more data that an LLM ingests, the keener it seems to be at grasping context, performing reasoning, and doing structured tasks like coding. To put it another way, they can better understand all the wild, unpredictable things that humans ask it to do. My favorite recent example has been the Gen-Zers who upload photographs from social media and ask the AI to determine the true height of a potential date. The LLMs are able to take knowledge learned from one domain (the estimation of distance from the laws of perspective) and apply it to another (Instagram), and report that yes, he is 6’2”.

There’s also the issue of accuracy. Imagine a model that is 95% accurate versus one that is 99% accurate. It doesn’t sound like that much: it’ll still be good most of the time, but not always. But now imagine that you ask this model to do ten things in sequence. The first model will now be 60% accurate, or essentially useless. The second one will be 90% accurate. Every little bit of precision matters.

And then there’s the very interesting thesis of what AI researchers call “emergent capabilities,” the idea that sufficiently smart AI models are suddenly able to do things that previously seemed impossible. One day, a model can’t do complicated math and then the next day it can. It’s possible that this phenomenon is just a mirage. It’s also possible that there are real jumps in capability that happen that require more scale. And then there’s also the interesting possibility that the public will only want to use small models, but that these models will need to be trained on data created by the big ones. This is an approach OpenAI is reportedly trying, and one that Meta pursued too.

The rapid scaling of AI leads to talk that we will soon have models that are as smart as a PhD student. (I can hear the PhD students laughing.) But just like the mushy tomato that you left in the garden for too long, it turns out that growth isn’t all good. Gigantic LLMs require tons of power both to build and to operate. This is all very expensive and not very efficient. It’s also more difficult to understand the inner workings of large LLMs, and therefore they are harder to “align” with human values. 

What’s clear is that we all share a fascinating window seat onto a technological revolution. We may look back at the original LLMs in the same way that we look back at the early computers that filled three rooms and could be taken down by an actual bug. The notion of a “bug” in the software comes from a moth that lodged in Mark II Aiken Relay Calculator at Harvard University in 1947. That’s a fact that I looked up on my phone, a supercomputer that fits in my hand. 

Cheers * N

I hope you enjoy this newsletter, which, thanks to your support, now has more than 450,000 readers. Please continue to forward it to anyone else who might enjoy it. They can sign up here.