AI Assistants Will Be Great (Especially for the Biggest Companies)

In this issue:

• AI Assistants Will Be Great (Especially for the Biggest Companies)—Thoughts on the economics of packing as many meetings as humanly possible into a short trip, and how AI agents will make this task vastly better. Big companies will capture some of the financial upside, but while they do that they'll be redistributing access to the now-pricey but soon to be cheap luxury of highly effective administrative assistants.
• Short Selling—Are loose regulations bad for short sellers, because frauds don't go to zero? Or are they a boon because more frauds last long enough to go public?
• Crypto Treasury Gresham's Law—Moving part of a corporate treasury into crypto is too easy to make it a competitive advantage.
• The Joy of Higher Rates—Japan's financial markets finally price in something other than indefinitely flat price levels.
• Labor Substitution—When human labor lacks the personal touch of a statistical language model.
• Pricing Black Swans—Plane crashes are rare, extreme events, but they don't seem to have much of an effect on passenger behavior or airline profitability.

The Diff February 10th 2025 12:30 • Click to play audio

AI Assistants Will Be Great (Especially for the Biggest Companies)

I spent a few days in New York last week, and like many similar trips:

1. I didn't ping everyone I'd like to have met,
2. I didn't get to every meeting I would have preferred, and
3. When it did happen, it was often in a fairly inconvenient way: Uber may have spent a lot of money acquiring customers a decade ago, but I suspect that at least in my case lifetime contribution margin flipped positive during some round of uptown-then-downtown-then-midtown-then-Brooklyn commuting.

All of this can, in principle, be dealt with manually. And there are probably some very conscientious people (or people with very conscientious admins) who have nice spreadsheets of prospective meetings, color-coded by importance and tagged by location, such that they can solve for maximum human utilization during a business trip.^[1] But it's an open-ended problem, of the sort that justify paying very high-priced assistants ($, Diff): to optimize for this in advance you need a comprehensive list of people you'd want to meet, some way to prioritize them, a sense of both where they're usually located and where they'd like to be, and then the patience and judgment to approximate a solution to the NP-hard problem of which meetings to take. (Usually when you look at such problems, you can brute-force them, and that's true here, too, except that instead of a hardware constraint on how many schedules you can simulate, there's a practical constraint that people's schedules change over time, they usually don't share every single slice of availability in advance, etc. So not only is it an NP-hard problem, but it's a distributed system with all the CAP theorem-adjacent problems that entails.^[2]

AI will not solve this, but it's a wonderful tool for trying. Because what I actually want is to describe, in natural language, why I'm in town and who I'd like to meet, to get a list of all the prospective meetings, and then to have my AI agent talk to everyone else's AI agent and figure out the best mix of meetings from there, probably with a rough cut early on and then continuing refinements from there. Ideally, if you run such a program you'll start out with some obvious questions about priorities, and then over time some increasingly granular and weird ones about the last-minute tradeoffs. Your model might be clever enough to bundle together some meetings around meals, and might ask if you know what your physiological limit for continuous face time is. But the vast majority of the work would be negotiations between different models, each of which is also periodically peppering its owner with similar questions.

At the level of writing prompts and interfacing with a calendar and with a travel time API, this is not an incredibly challenging problem. There are probably plenty of Version 0.1s that have already been put together by technical founders who find themselves doing lots of sales. But one of the hard problems is data: the agent needs to know who to contact, and who not to contact; it needs to be aware of email inboxes and outboxes, but also LinkedIn messages, WhatsApp groups, and Signal chats. And it needs to tie that information together with location data to form its initial list.

There are three directions this can go:

1. Someone like Apple or Microsoft will win if the operating system ends up being the data switchboard for agents. (This is part of what made operating systems a good business in the early days of the PC: owning the layer at which applications work together offers a lot of leverage over the applications themselves.)
2. It leads to people consolidating more of their digital lives into a single company's ecosystem, which will almost always mean dropping the Nth-most-used of your N apps until your agent stops missing important messages.
3. Someone builds a manage-your-life agent that's so good that every app with a messaging feature, an address book, or metadata about social connections is forced to interoperate.

That last one is a possibility, but the functionality of such an agent is easier to replicate than the data it needs. (And, more so than other categories, there's a risk in scaling when the product has usage but hasn't figured out monetization. Inference isn't cheap, and a product that does more of it as it scales is especially expensive.) That last constraint holds here: agents are more patient with each other and are more willing to have a long, detailed back-and-forth, so agents will spend more on inference as they spend more time talking to other agents.

An agent for scheduling face-to-face meetings is a good first category because these kinds of interactions are close economic complements to AI, in two directions. One of the problems with firing people and replacing them with software agents is that it shortens the chain of blame; if the agent messes up, there isn't much for users to do other than complain to the company (and probably hear back something involving the phrase "user error.") So there's economic demand for humans in the loop, specifically so there's someone to take responsibility for mistakes. Meanwhile, many of the most valuable applications of user-facing AI today are things that have an uncertain upfront opportunity cost. Will that script take ten minutes or two hours? Doesn't really matter if Cursor does it in a few seconds. Is it worth learning enough about some new technology to even know what you should Google next to dive in further? ChatGPT Pro’s Deep Research mode de-risks this by producing an ad hoc Wikipedia page for any topic you can imagine. AI is saving time, but it's specific kinds of time, and the rise in that kind of productivity means that there are higher returns to having more conversations.

At one level, this is yet another AI story in which the profitable, scaled incumbents get an even better business. But at the user level, it's a force for egalitarianism: the relevant comparison for this product isn't an Outlook calendar or even a typical administrative assistant; it's one of those assistants whose pay puts them in the top 1% nationally (albeit not in the top 1% in the locations where such people tend to work). If the opportunity cost of a scheduling mistake or a forgone meeting is in the thousands or tens of thousands of dollars, it makes sense to pay this well for someone with a low error rate. But the agents won't suffer the classic market for lemons problem, where the really excellent admins only come on the market when whoever they're working for retires. This kind of service job gets nonlinearly more expensive the richer society gets, both because the stakes go up for the people who can afford them and because there are so many well-paid opportunities elsewhere in the economy. A good assistant is not quite the luxury people usually have in mind when they think about how nice it would be to have a lot of money, but what people in that economic stratum know is that the marginal return on money goes down while the marginal return on time stays fairly constant. Thanks to AI, we'll be able to enjoy at least some of the lifestyle that was previously the province of the very rich.

Disclosure: long MSFT, AMZN, META.

Diff Jobs

Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:

• A well-funded startup that’s building the universal electronic cash system by taking stablecoin adoption from edge cases to the mainstream is looking for a senior full-stack engineer. (Remote, Singapore)
• 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)
• An OpenAI backed startup that’s applying advanced reasoning techniques to reinvent investment analysis from first principles and build the IDE for financial research is looking for software engineers and a fundamental analyst. Experience at a Tiger Cub a plus. (NYC)
• A hyper-growth startup that’s turning customers’ sales and marketing data into revenue is looking for a forward deployed engineer who is excited to work closely with customers to make the product work for them. (SF, NYC)
• 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)

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

Short Selling

This Institutional Investor piece speculates that short-sellers will have a hard time in the next four years because of looser SEC enforcement. That's possible, but it's not strictly true. For one thing, part of what's made short selling difficult is the transition from spotting companies that were lying about having a good business to companies that are, in a strict accounting sense, telling the truth that they run a terrible business but then applying some spin to the story.

If companies periodically commit fraud, and the SEC generally catches them, that doesn't necessarily produce alpha. In fact, it can aptly describe a system where the short sellers don't produce any alpha from frauds: one where they all get caught before they go public. And in a world where regulatory action happens randomly, it also doesn't do much for returns other than adding an unpredictable catalyst. The ideal situation for short sellers is being able to predict not just that the SEC will go after fraud, but when, or to engineer this by publicly calling attention to deficiencies in companies' reported numbers. But there are plenty of frauds for which investors raise questions for years without any regulatory action.

The other side of the deregulatory coin is that short sellers will have a far wider opportunity set. Since the other big secular tailwind to short selling was a change in the distribution for extreme price moves in small-cap stocks, i.e. the meme stock phenomenon, a regulatory light touch is arguably better for them. If there are twice as many frauds and they're twice as glaring, short sellers get to diversify over twice as many names, and the next GameStop is a less dire problem.

Crypto Treasury Gresham's Law

The ranks of public companies that invest their treasury in cryptoassets continues to grow ($, FT). Money is, by definition, the most fungible product around, and swapping one kind into another is an extremely common and thus low-cost economic activity. So there's very little stopping every company from doing this if the market pays a premium for it. It's hard for them to get mindshare, though, so ironically in this space there's the same kind of network effect as in cryptocurrency itself; the earliest mover accumulates the most market cap, and trades at the highest premium; a few attempted clones do well, but most stay obscure.

The Joy of Higher Rates

Long-term Japanese government bond rates have reached a 14-year high, though of just 1.31% ($, FT). Higher rates are a drag on the economy, but a worse drag is rates of roughly 0% being too high to stimulate consumption or investment. Some of the overhangs Japan faced were demographic, but their situation was worse than mere aging would imply because of the overhang of bad debts that weren't worked out. It can take an extraordinarily long time for an economy to work past that, and a long time after for managers and investors to drop the mental habits associated with a deflationary economy. But once money has a price and the price is worth paying, there's a new overhang: the set of investments that made sense but haven't happened yet due to investor timidity.

Labor Substitution

Insurer Allstate has found that its AI-generated customer service communications are less jargony and more empathetic than the ones that were written by its own customer service team ($, WSJ). As with many LLM stories, it's interesting to consider the corpus: inside of an insurance company there are going to be plenty of communications that express some level of annoyance with customers, either over what they did that led them to make a claim or over how they respond to news about whether or not they'll actually get their money. But online more broadly, the typical discussion of an insurance claim is very favorable to whoever's making it; people post about this kind of thing in cases where they're confused, or where they suspect that the insurance company isn't paying at a time that it should. And that corpus probably has plenty of examples of people puzzling out strange acronyms or explaining precisely what an insurance company means when it uses what sounds like a descriptive everyday term. The LLM is taking the perspective of an outsider by default, while someone at an insurance company is an insider, so for communications with the outside world the LLM has an advantage.

Pricing Black Swans

Crashes tend to have little long-term impact on airlines' financial performance or share price, at least for major ones. And this makes sense: for all airlines' efforts to differentiate themselves, the actual transportation is a commodity product, and giving customers even a slight reason to prefer a different brand is costly. But disasters are rare enough that they really don't provide much information about the overall safety of a given airline; whether a crash was a singularly unlucky event or whether the airline in question had a long lucky streak beforehand is a question that, at the current rate of crashes, requires a longer measurement period to answer than the time the industry has been in existence. The airlines themselves are happy with strict safety regulations (so long as those regulations don't slow things down unnecessarily), since the last thing they want is for price competition to turn into competing to relax safety.