This is The Diff’s weekly roundup of longreads, books, and open questions. Feel free to drop in links and comments of interest to fellow readers.

You’re on the free list for the newsletter. For full access—and to read the back catalog—subscribe today!

The Diff is a reader-supported publication. To receive new posts and support my work, consider becoming a paid subscriber.

Longreads + Open Thread

AI Novelists, Language, Stripe, Science, Currencies, Oil

Byrne Hobart

Jul 23

Longreads

Josh Dzieza of The Verge has a wonderful look at AI-assisted novels. This is a move towards the end state I described here, where it's possible that text-, image- and video-based AI will get good enough and cheap enough that most content will be produced for a single user. One way to get closer to that is for micro-genre authors to be able to increase their production of stories in particular niches. This work still requires a human in the loop, since text models can't keep track of details in a long story.¹ So for now, it's making people more productive rather than making them obsolete. That complementary relationship is a function of specific limitations in some text models, and if there's demand it's possible someone will create a layer around GPT-3 that keeps it on track as it generates chunks of the story, to the point that the author's contribution is minimal.
Benjamin Bratton in Noema has a thoughtful piece on what it means that we ask if AIs are sentient. One fun speculation: consciousness may have evolved from our efforts to model other humans; if we're trying to predict what other people do, we might get around to trying to predict what we ourselves will do, and if we do that in real time it will feel a lot like making conscious decisions. (Daniel Dennett has a fun short story on why these sensations might feel almost identical.) The piece has more meditations on the limitations of language models, and what these say about the limits of language itself.
Brie Wolfson in Every on what she misses about working at Stripe. The most notable detail: when she was asked what her favorite time at Stripe was, she mentioned a period when the company was struggling with API issues. As it turns out, this was a common answer to that question! It’s hard to optimize for, but people look back fondly on times when they suffered with people they like, while working on something they cared about (especially if the equity comp worked out in the end).
Charles Piller in Science on signs of fabricated data in Alzheimer’s research. As science advances and discoveries get more specialized, it gets more important to have confidence in science as an institution, because most of us won't be in a position to double-check the results. At the same time, the fact that results are hard to check means that we should have lower confidence, because the risk/reward increasingly favors cheating. It's fortunate that there's a community dedicated to spotting malfeasance (interestingly, this story includes both people motivated by the fun of sleuthing for scientific sleight of hand and short sellers trying to validate a thesis).
Dylan Patel of SemiAnalysis has a great writeup on DISCO Corporation, a Japanese firm that specializes in equipment for grinding semiconductor wafers. (There's a paywalled portion with financial analysis, but the general writeup is free to read.) A highlight: the company has an internal currency which workers earn by completing tasks, and which units in the company can spend to compensate one another: "At DISCO, everything has a price, from conference rooms, office desks, and PCs to a spot for your wet umbrella." This could be an exhausting system, but it's a striking example of how diverse companies' internal structures can be.

Books

Oil Capital: For as long as there have been independent energy companies, there have been complaints from independent energy companies that it's hard to raise money for new projects. This book is a detailed look at how oil and gas exploration businesses have historically capitalized themselves, how their banks have operated, and how the legal system has evolved. (As it turns out, this legal evolution is a big deal; the US is almost unique in letting individuals own mineral rights, which has led to a more decentralized system for extracting resources and for funding that extraction.) The book also has some anecdotes that are less edifying and more fun: Sinclair Oil may have gotten its seed capital from an insurance payout when its founder perhaps-not-accidentally shot off his toe.

Open Thread

Drop in any links or thoughts of interest to Diff readers.
It's fairly easy to keep tabs on advances in text- and image-generating AI, because these produce cool demos. And the companies that make them have realized that a demo with a waitlist leads to lots of Twitter FOMO. What's happening in the parts of AI that lead to more subtle changes, like better content recommendations and spam filtering? And what's the best way to track that?

Reader Feedback

Last week, a subscribers-only piece ($) talked about BMW's decision to charge a monthly fee for heated seats. This comment from Levi Ramsey was a spin on it that I hadn't thought of:

For BMW, heated seats and the like are something that has to be standard in markets where cool mornings are normal, but in other markets almost nobody is willing to pay for them. On the other hand, making this an option or having a cold-morning version of each model with heated seats and a version without heated seats dramatically complicates the manufacturing process, so it's plausible that having heated seats be a global default saves enough in supply chain and manufacturing complexity that it comes close to covering the cost of the hardware to heat the seats.
But in markets where they've previously not shipped cars with heated seats, suddenly making them standard leads to buyers assuming that the feature they never use is built into the price and trying to negotiate down (ask a Hampton Inn front-desk agent how often they get a guest trying to knock $5 a night off the rate with a promise that they won't have the free breakfast in the morning). So putting a subscription on the feature sends a message that the MSRP hasn't been increased to reflect even partially the added cost of the hardware (it probably has been, though by far less than the option typically retails for).

And last week’s open thread has some good thoughts on which jobs will be automated and which ones won’t.

A Word From Our Sponsors

Import millions of rows of messy customer data in minutes

“Hey, can you send me a spreadsheet with the following columns at minimum?”

Sounds simple enough, but ask 100 customers to do it, and chances are you’ll get 100 different spreadsheets, documents or formats sent back to you. Now try collating all that data into one place, without losing your mind. Or worse, asking your customer to do it.

If this sounds anything like your team’s reality then you need to get some help from Flatfile — who have built the ultimate data importing cheat code.

Their innovative platform lets your customers drag and drop their data into your system without formatting beforehand. Then the magic starts. Smart matching, file hooks and algorithmic data detection means that a messy file that used to take hours can now be imported, cleaned and validated in literally minutes. The best part? You can embed Flatfile straight into your product, or in a prebuilt hosted page.

Try Flatfile

1

It's kind of funny that one of the earliest functions of computers was keeping track of lots of details that humans would be prone to forgetting: "Remind me of what every single employee's salary and tax withholding is" was one of the earliest businesses asked computers to do. Computers excel in structured memory tasks but have a ways to go for unstructured memory tasks—not just learning things, but keeping track of what’s important.