A Bull Market in Facts; A Bear Market in Narratives
A Bull Market in Facts; A Bear Market in Narratives
Last week, a distributed machine learning project delivered its first big success. Some models come up with predictions of consumer behavior, or stock prices, or the weather. This one produced the word "πορφυρας," a term for a purple dye— it’s the first word to be been identified by a new project involving the Herculaneum papyri, which were last read almost two thousand years ago.
It's an exciting time for ancient history nerds (this piece has a preview of what we might find). And we do know generally that most of the ancient documents we know of, we don't have copies of, and that in some cases copies weren't made with complete fidelity, especially if a document flattering to one political faction was reproduced under the auspices of another.
This raises all sorts of interesting questions. For example: suppose we find a lost dialogue of Plato. It's interesting, but is it part of the Western canon? In one sense, surely it has to be: Plato had interesting thoughts that were immensely influential. On the other hand, part of the way the canon works is that the reader is in dialogue with whoever they're reading at that moment, but is also accepting an invitation to a party attended by everyone else who has read them and had comments. You can read Machiavelli in part to understand Machiavelli and in part to understand or argue with Leo Strauss and his disciples. It will be interesting to look at works that are just as old as the ones we have, but with two millennia less commentary about them.[1]
Another question it raises: emerging technologies often go through a period where they're obscure and there's lots of room to quickly make progress, but quickly reach some kind of talent/capital bottleneck that makes the companies involved more institutional. It's harder for someone to come out of nowhere and immediately make big contributions to machine learning, because there's so much context for them to catch up on and the cost of acquiring the relevant data or compute has gone up. Which makes this papyri discovery a good indicator of a useful mental model going forward: there's probably more upside and less competition in fields downstream of AI. If you want to be a 99th percentile in a field that's changing fast, maybe instead of learning Pytorch you should learn ancient Greek.[2]
But perhaps the most important takeaway from this is that facts, and narratives that tie them together, are getting continuously cheaper to manufacture. This talk is a timely look at one way to respond to this: make it easier to have chains of attestation behind any given image or video (i.e. "X was there and saw it happen; X knows Y, Y knows Z, Z knows you, so unless someone in the chain is lying, it's probably legitimate.") This solves one particular kind of disinformation problem, that of repurposing old photos and videos and claiming they're current. But it's also hard to get people to adopt this solution, since they already have two mental models:
- Trusting major media outlets, and assuming that they wouldn't outright lie though they might shade the truth. (A notorious way to do this is a close-up shot of a protest or celebration, which can make it look a lot bigger than it is. The toppling of a statue of Saddam Hussein in 2003 is probably the example, but just about any political rally organizer will make sure there's a shot that makes it look much more crowded and enthusiastic than it really is.
- Trusting confirmation bias.
These are both pretty unavoidable. Most news organizations will not knowingly run doctored images, but journalists are human, too, and if there's a picture that captures how they felt, but doesn't give a completely unbiased look at what's going on, that's the photo they'll run with. It's akin to any other judgment call; if the subject of a given article gives terse answers and doesn't offer the writer much time, whether that's described as evidence for focus or rudeness comes down to how the writer feels.[3]
And sometimes key facts emerge only well after they're useful. The US overestimated the size of the Soviet economy for decades, for example, but didn't get accurate estimates until well after war was a likely prospect (i.e. well after the US was making any big decisions that hinged on Russian manufacturing capacity and the like). The US had an incentive to be accurate, but Russia had an incentive to exaggerate; the size of the US consumer economy and the US's freer media environment meant that American manufacturing capacity was an open secret. (And the Second World War demonstrated that the US's manufacturing capacity for civilian goods could be repurposed for military ones—especially since plenty of those factories had been acquired from the government after the war.[4])
History is like many other domains where there’s a tradeoff between accumulating facts and assembling them into coherent narratives.[5] Mere recitation of facts is helpful to specialists, but only because they come to them with context and a narrative in mind. Much more common, especially for popular history, is to make some fairly extreme but appealing claim, generally about how some concept or group made the modern world. That's helpful, of course, but it gives writers lattitude. More translations, more ways to search them, and more ways to cross-reference different documents will make it much easier to find data to justify a thesis whether or not it's true. And since facts in question can be true even if the message is not, just investigating individual claims doesn't work; refuting a book like that means coming up with an entirely different thesis, and finding supporting evidence—which is exactly the kind of agenda-based research that caused the problem in the first place!
So, on one hand, we're in an increasingly hostile information environment specifically because it's easier to assemble enough verifiable, factual claims to support any arbitrary narrative behind them. That should be worrisome: no level of fact-checking can protect against it, and "narrative-checking" is either a redundant description of what people do all the time, or a product for which there's no demand. But that also means we live in a relatively better world for people who care about dead issues rather than live ones. Very few people particularly care if the Battle of Kadesh in 1,274 BC was a victory for the Egyptians or the Hittites (both sides claimed victory, and there's a snippy letter from the Pharaoh responding to these Hittite claims), but the physical evidence isn't getting any sparser while the ability to analyze it is growing. And, who knows, maybe Linear A is next.
One way to view teaching the humanities is that it's an effort to make the books as life-changing today as they were when they were first written. But that also means readers can't evaluate the book independently; books get referenced in other works, or recommended by people, so you're almost always going into a book with the sense that there's a flesh-and-blood person who you'll understand better once you've read them. ↩︎
Another surprising AI impact is that it makes written tests harder to administer, and makes oral exams relatively easier. It's entirely possible that someone in the near future will get an educational credential by demonstrating their ability to master a dead language based on getting quizzed one-on-one by their teacher; a direct result of AI is that at least some kinds of education will look a lot more like they did in the 16th century. ↩︎
One subtle version of this, which shows up from time to time, is whether linguistic filler gets removed from quotes. Most people say "um," "like," "well," etc., while they're talking, or will start a sentence and then verbally backtrack. Keeping the former in the quote makes them sound dumber, and keeping the latter in makes them sound dishonest. Which may be true! A good writer isn't going to come out and tell the reader "I think my interviewee was lying to me," but will try to make it clear from the narrative. It's often a good thought experiment to think about what an article with the opposite mood affiliation would sound like if it were written from the same set of notes. ↩︎
It's frustratingly hard to find out how big this phenomenon was, but eyeballing table 3 here implies that the total was about 10% of US manufacturing capacity. ↩︎
You can even see this in some of the oldest documents we have that combine a story with a recitation of facts: the Iliad's Catalog of Ships, the recitations of family trees in the book of Genesis, the genealogical tables in the Records of the Grand Historian, etc. In a way these feel like a primitive example of pre-literate recordkeeping, but they're also quite modern: "What's up everyone? It's your boy Homer back with another episode of The Iliad, but before we get rolling just wanted to give a shoutout to my top-tier supporters, the Phocians and Bœotians. Love you guys." ↩︎
A Word From Our Sponsors
At martini.ai, we use AI to provide real time credit risk estimates for over a million private companies. We integrate all available market information into a massive knowledge graph to provide color on illiquid credit. With our optimized risk management solution, lenders can make informed decisions, safeguard portfolios, and adapt to the ever-changing business landscape. Say goodbye to manual processes and embrace the future of corporate risk management with martini.ai!
For July, martini.ai is offering a free one month subscription for The Diff readers and a free AI-powered portfolio risk evaluation for your corporate credit portfolio.
Elsewhere
The Decline of Meme Stocks
Even though the meme stock phenomenon is popping up in other markets, its home in the US has seen better days ($, WSJ): margin borrowing is down and retail-focused brokers are reporting lower trading volume. There are some good reasons for this: the risk-seeking investors had a good run for a while, but then blew up in Gamestop, AMC, or crypto; the megacap tech stocks driving the overall market are too big for memes to matter much; and self-fulfilling prophecies tend to work until they don't, with no way to get them working again.
But another reason is more prosaic: one driver of meme stock mania was the distinct sense of not having a plausible path to affording a home, or, more generally, to accumulating some money that could compound nicely over time. The increase in interest rates has suddenly made grade school illustrations of the power of interest meaningful again, because money in a high-yield savings account really does compound at an appreciable rate. When there's no chance of earning what looks like baseline middle-class wealth (i.e. enough money to make a downpayment on a house similar to the one you grew up in), people gamble. When cash yields more than 5%, they don't.
Volatility
One other effect of higher short-term rates: long-term bond ETFs have significantly higher volatility than the S&P 500. This happens from time to time—the article is accompanied by a chart showing through all of 2017 and much of 2013 and 2015. One way to think about this is that the stock market tends to adjust to higher inflation, in two ways: first, companies change their prices and their cost structure in response to inflationary pressure. It's usually costly at first, but can be profitable over time. The subtler reason is that the composition of the market also changes: when inflation is high, growth stocks will do worse (and be a smaller share of the market) and energy stocks will do better (and see their share grow). So shifts like this create a stock market index that's built to hedge against whatever the latest macro shock is, while the bond market has to keep making the same bet over and over again.
Lobbying as a Competitive Advantage
The FT has a profile of Microsoft's extensive, and expensive, lobbying operation ($, FT). Part of the bull case for Microsoft is that, as far as the DOJ and FTC are concerned, every other big tech company is Microsoft in 1999, and Microsoft itself is something closer to, say, Netscape. The company clearly does have market power, and exercises it, but not incredibly aggressively—and generally in categories where it's assumed that Microsoft's corporate customers can largely take care of themselves. One way to tell part of the Microsoft story is that the company was putting too little money and effort into keeping the government happy in the 80s and 90s, and then staffed up heavily to compensate. Money generally works in lobbying, but impact is closer to a function of money multiplied by time. Once the required expense was below Microsoft's threshold, continuing to pitch their story to regulators and politicians has given Microsoft flexibility that its peers lack.
Disclosure: Long MSFT.
Goldman
A few years ago, Goldman Sachs made the strategic decision to shift away from lumpy and unreliable revenue from trading and transaction fees, and towards reliable funding and steady income from consumer banking. They are now in a hurry to get out of this business, having divested one lending platform and made efforts to extricate themselves from their Apple Card relationship ($, WSJ). There's an interesting culture-clash story about the business itself, which has been written about before: Goldman was traditionally a fairly elite firm, and financing people's Target shopping and Chipotle burritos is less Goldman than managing a Chipotle bond issue or helping Target execute an acquisition. Another element of this culture clash is the speed of exiting: traders like the fact that, if they're wrong, they can quickly exit or even reverse their previous position. Exiting is a matter of closing out positions, and costs automatically adjust when the biggest cost is performance-based bonuses. In a more traditional banking business, though, exiting the trade is some combination of letting loans slowly run off or paying someone else a lot to manage the running-off process themselves.
"Name and Shame"
Japanese business and policy elites are continuing their campaign to pressure companies into increasing their stock market valuation. The latest move, from Japan's stock exchange, is to highlight companies that are complying with new rules on pushing up their share price ($, FT), shaming the laggards by implication. Most of the time, exchanges don't directly do this kind of thing; they're often happy to engage in vague boosterism about owning a stake in economic growth. But when valuations are at extremes, and it's at least partly because managers are indifferent, pushing companies to buy back stock and thus rewarding stock pickers ends up being good for business.
The Diff has previously covered why Japan's market is so cheap and what historical analogies make the most sense for understanding it.
Diff Jobs
Companies in the Diff network are actively looking for talent. A sampling of current open roles:
- A private credit fund denominated in Bitcoin needs a credit analyst that can negotiate derivatives pricing. Experience with low-risk crypto lending preferred (i.e. to large miners, prop-trading firms in safe jurisdictions). (Remote)
- A new fintech startup wants to bring cross-border open banking to LATAM, and is looking for a founding engineer. (NYC)
- A company building the new pension of the 21st century and building universal basic capital is looking for a product manager with fintech experience. (NYC)
- A systematic hedge fund is looking for researchers and portfolio managers who have experience using alternative data (NYC).
- A proprietary trading firm is seeking systematic-oriented traders with ML experience—ideally someone who has displayed excellence in DS and ML, like a Kaggle Master. (Montreal)
Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.
If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.