|
In this issue: - Social Media as a Test of Human-in-the-Loop Generative AI—Social media platforms have, in various ways that very much play to type, started incorporating generative AI in a way that encourages their users to consume it and to share it. The evolving norms around this show that we think of AI as a tool, and that we attribute problems with its use to the user, and that's a good norm to codify.
- Holdcos—The optimal size of an advertising company is somewhere between one part-time worker and a globe-spanning conglomerate with a team of 100,000+
- Cash and Control—On the margin, it's possible to trade a little less equity upside for a little more control, but past a certain limit that extra control constrains the counterparty's expectation about the upside.
- War Economies—Russia accidentally engineered a wartime economic boom. Here comes the bust.
- In-Kind Dividends—Buy-one-get-one-free comes to Japanese equity markets.
- Currency—El Salvador quietly drops Bitcoin's legal tender status.
The first practical use case I encountered for generative AI was on social media: the Matt Levine bot, which was built with GPT-2, which produced tweets that were sometimes funny and often relatable. It was a fun little tool, despite the lack of practical uses.
And generative AI does seem naturally suited to social media. Posts are constantly selected for virality, which you can sometimes see in real-time (for example, here is a screenshot of a news story saying that the UnitedHealthcare shooter's backpack included "a jacket and monopoly money," while the caption has the much more viral claim that it was "full of Monopoly money." Fixing that detail of reality makes the story much more viral). Visuals are favored by both users and algorithms, but for some posts there isn't a ready visual. Feed-based sites use a mix of text, images, and video, with static visuals at the ideal intersection where 1) a computer can do a decent job much faster than a human being can, and 2) the cost isn't as prohibitive as it would be with video.
Meta, Twitter, and LinkedIn have all started incorporating generative AI into their products (in addition to hosting a lot of slop that was created externally). And they do it in ways that play very much to type:
- For Meta, generative AI prompts in response to posts are replacing a bit of monetizable screen real estate (follow-up questions on viral posts, prompts to generate AI profile pictures) with something that requires more capex and entails more opex and COGS than what came before, which will naturally annoy investors until the additional usage it creates starts lifting revenue and costs come down with scale.
- Elon Musk constructed Twitter's AI integration through related-party transactions with another company he has absolute control over, which is owned by a different set of investors (though he did give Twitter shareholders 25% of xAI's stock ($, FT). He also sped up the hiring process by grabbing a few experienced engineers from Tesla ($, The Information).
- LinkedIn has added AI to the microcopy for status updates; they're under no illusion about how hard it is to write something work-related, work-safe, and not terminally boring. As a consequence, LinkedIn users are constantly pumping out readable but banal status updates. Since LinkedIn is more professional, it isn't as image-centered as other social networks, so it also doesn't have to worry as much about the cost.
Part of this is business-as-usual for social networks: there's a new interaction model, so they all try it out, whether or not it makes any sense for their users. But it's also an interesting test case in both how AI will be used and how it will be regulated, because all of these uses have something in common: the user is curating the output, or at least being held responsible for it. It's been encouraging to see an emerging norm that blaming AI is never acceptable: if you invent nonexistent precedents to bolster your legal case, cite imaginary historical events, or use ChatGPT to generate fake citations for your report on the dangers of generative AI manipulating public opinion, basically nobody will cut you slack. That's good evidence that, at least socially, generative AI is not viewed as a more advanced form of autocomplete or autocorrect—there's little shame in being "so ducking mad" about something compared to citing completely fake information.
But it also means holding people responsible for vetting what they say. And that's a decent way to constrain some of the broad risks of AI. A software agent is just code, but it's written by someone, run by someone, hosted on someone's hardware, etc. The fact that it can produce arbitrarily surprising outputs means that it's qualitatively different from deterministic software, or even from software that incorporates a bit of randomness into a mostly deterministic process, like a procedurally-generated game that will always generate the same world from the same seed, but whose seeds are selected randomly. An AI agent that thinks-out-loud about the most cost-effective way to host an office pizza party and concludes that it should order the pizza with a stolen credit card is doing something very different from a Dwarf Fortress map that puts a cave spider right next to the starting area. In a game, it's a low-probability outcome from a known distribution; with an AI agent, it's an outcome from a distribution that gets less certain with each step (and as "lay out trees of options, prune some, and expand from there" gets more common as a way to turn token prediction into coherent plans, those distributions will get wider.)
In general when there's an emergent phenomenon, the regulations economic actors impose on one another are a good guide to the baseline rules that the government ought to impose. Markets and courts are always engaged in an exploratory process defining the boundaries of property rights and liability—technologies are indifferent to their own upsides and downsides, but they're used by agents who can pay fines or go to jail, so that's where the regulation happens. And, thanks to widespread alarm about the prospect of AI-manipulated images and AI-generated words having a significant real-world effect, they're also an environment where this content will get disproportionate attention, even if it takes the fairly harmless form of inspirational images (a misattributed quote over an AI-generated image of a sunset has about as much social or artistic value as the same quote over a real image of a sunset, and the artist's contribution probably wasn't going to get compensated either way). So the gradual adoption of generative AI in social media posts is a good sign about our ability to handle it elsewhere, and a great place to firm up the precedents around precisely how we'll do this.
Disclosure: long MSFT, META.
Diff JobsCompanies in the Diff network are actively looking for talent. See a sampling of current open roles below: - The treasury management arm of an established stablecoin project is looking for a research economist to bring a macro perspective at the intersection of traditional monetary theory and digital currency innovation. Advanced degree and portfolio of high quality research output preferred; no crypto experience necessary. (Remote)
- An AI startup building regulatory agents to help automate compliance for companies in highly regulated industries is looking for account executives or consultants with direct experience selling into large financial institutions (banks, asset managers, insurance companies). (NYC)
- YC-backed AI company that’s turning body cam footage into complete police reports is looking for a tech lead/CTO who can build scalable backend systems and maintain best practices for the engineering org. (SF)
- A hyper-growth startup that’s turning customers’ sales and marketing data into revenue is looking for a product engineer with a track record of building, shipping, and owning customer delivery at high velocity. If you like to build, this role is for you. (NYC)
- A growing pod at a multi-manager platform is looking for new quantitative researchers, no prior finance experience necessary, 250k+ (NYC)
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. Elsewhere
Holdcos
Advertising is an odd industry because there's a continuum from being a freelance creative type, owning a small service business, and running a global conglomerate, and at each level the constraints can be surprisingly similar. One of those constraints is that the human capital in the industry is incredibly mobile, and since the human capital is most of what they're selling, this makes it hard for any company to build a durable advantage. Individual people can, but they take it with them. There are some benefits to firms combining, though; they can diversify across different practice areas, for example—if agency A is good at digital direct response and agency B is good at TV spots, A and B together probably handled the rise of YouTube, ad-supported streaming, and CTV better than they would have apart. And a diversified group of agencies is also a hedge, allowing any one group to underperform for a bit without falling apart. (This naturally requires management finesse, because what it means in practice is that whichever group is doing the best is slightly overpaid, and vice-versa, so they have to be good at identifying cycles and secular changes.) If the holding company model does work, and it seems to, what's the limit to which it can scale? We may find out soon enough, as Omnicom and Interpublic consider a merger ($, WSJ), which would create the largest holding company. (They’ve since confirmed the deal.)
One way to explain that opportunity is by omission. The current largest holding company is WPP, which didn't start out that way; its name is short for Wire and Plastic Products, but it was taken over by Martin Sorrell, who had previously helped build Saatchi & Saatchi into a holding company. WPP grew, mostly by acquisition, to become the largest ad holding company, and shares returned 15.1% annualized over his tenure, but he was ousted in 2018 due to assorted scandals (returns are flat since then). Whether or not he did that great a job managing it, any time a company is run by one CEO for multiple decades, and doesn't have a succession plan in place, it's very hard for it to stay on track—the next few years are a fun process of discovering precisely how much of the company's operations and strategy were contained entirely in the head of a now-ex CEO who isn't inclined to return phone calls. So there's room to scale a bigger holding company because the biggest is not in a great position to grow.
Cash and Control
In other merger news, Ubisoft is considering a going-private deal, and weighing how to trade off between equity and the founding family retaining control. There are some trades that can only clear at certain prices, and this tradeoff is one of them. In theory, there's no limit to how much controlling shareholders can trade their own economic upside for their control, but in practice at some point this looks like a setup that allows them to indefinitely take advantage of the equity outside investors contribute without any accountability. You can trade a little financial upside for a little more control, but 100% ownership and 0% management authority is a bad deal.
War Economies
For a while, Russia's economy was booming despite the war in Ukraine, but it's hard to sustain high growth in the face of real resource constraints like losing buyers for exports, losing suppliers for imports (including some components critical to those exports), and losing parts of the working-age population to war. The whole idea behind managing a country's economy during a war is to shift more spending away from consumption and towards the production of military equipment with as little friction as possible. This inevitably leads to disruptions of one kind or another, and one of the simplest ways to handle it is to accept a bit of inflation—deficit-funded military spending means that people in non-military sectors get lower real incomes, which engenders exactly the production shift the state is aiming for. Accidentally provoking a boom in consumer spending and a bubble in housing ($, Economist) is a side effect of pushing that much spending through the system, but it's not the end goal.
In-Kind Dividends
Sometimes companies will offer shareholders some kind of not-purely-financial benefit, like Berkshire Hathaway's intermittent shareholder discount at their jewelry store subsidiary or AMC's various shareholder perks. Japanese tech conglomerate Rakuten is now offering shareholders a mobile phone plan that's free for a year, which has sent their shares up almost 7%. Rakuten does not show up on the list of most-shorted Japanese stocks, so this doesn't look much like an attempt to engineer a squeeze. But it does make sense as a way to slightly turn over their shareholder base, and tilt it more towards retail. Japanese households have high cash balances, and Japanese stocks are statistically cheap, and this is an interesting way to close the gap. It's also a way to get a slightly stickier retail customer base, both for the mobile plans and for the stock: some customers will stick around because they're satisfied with their new plan, and some will feel like a bit of reciprocity is in order. Corporate Japan has a long history of supplier/customer relationships that are cemented through cross-shareholding, and now transaction costs are low enough to do that with individual investors rather than just other companies.
Currency
El Salvador has quietly walked back its decision to make Bitcoin legal tender ($, FT), as part of a $1.3bn deal with the IMF. For a long time, Bitcoin itself has made more sense as a reserve asset like gold than as a transactional tool like the dollar (and in fact the gold standard typically involved transactions backed by gold, and sometimes settled in gold, but usually not direct exchange of some quantity of gold for some quantity of goods and services). The country has limited state capacity and has invested plenty of it in law enforcement rather than its financial system, so adopting it as legal tender was always more symbolic than practical. As with many crypto enthusiasts, the first stage is thinking that cryptocurrency is a good solution for any problem involving the transfer of value, and the next is deciding that, like any other technology, it has some great applications and, in other cases, is strictly worse than the alternatives.
|
