In this issue:
- Graduating to Parallelized Strategies—There are surprising parallels between the evolution of programs and the evolution of companies: from single-threaded, single-purpose entities to more complex ones that require tradeoffs and planning.
- AAI—An AI company turns out to mostly use human labor.
- OpenAI—More updates have emerged in the OpenAI story, revealing important details about how business gets done and how news and gossip spread.
- Twitter—Twitter is trying to attract smaller advertisers, whose incentives differ from those of bigger brands.
- Pod Shop Math—A new multi-manager hedge fund tries to solve for forward-looking diversification rather than a backward-looking collection of uncorrelated-so-far strategies.
- Privacy, Security, and Monopoly—Apple is resisting efforts to reverse-engineer their messaging protocol.
- Diff Jobs
Today's issue of The Diff is brought to you by our sponsors, Intercom.
Graduating to Parallelized Strategies
There's an interesting parallel between learning to program and launching a company: early on, you can afford a simple, straightforward approach that solves a specific problem. In the beginning, programmers write simple scripts that only one person needs to use and only one person needs to edit. But eventually things evolve, and now they’ll need to collaborate with other developers, build around other users' requirements, and, eventually, manage a distributed system that needs to stay in sync with itself, but without all incremental effort going towards that synchronization problem.[1]
Early companies run through a surprisingly similar process. If they're growing, it's because they've found some under-exploited opportunity, whether it's a new category that should have existed, a better way to sell something, or a cheaper way to make it. So, early on, any discussion of corporate strategy is very short:
"Should we focus on the one obviously overwhelmingly important thing? Or should we do something else?"
This doesn't last forever, though. Markets and channels get saturated, competitors copy ideas, customers raise their standards, and markets shift. Over time, the original idea turns out to be something closer to a motivating example in some more general category, rather than the true question the company was built to answer.
If you're looking for companies that grew faster in their early days than today's big tech companies, you can certainly find them: Zynga and Groupon in the 2010s, for example, had insane growth curves, as did e-commerce consulting companies in the late 90s[2] and commodity PC manufacturers in the 80s. More recently, GPT wrappers were able to hit ARR milestones far faster than other kinds of software business. But for those companies, the R didn't R; wrapping an API in a nicer interface is a viable business until the API provider does it, too, and building a company on just one platform partly means doing R&D for the platform provider: OpenAI can calculate a postmoney-to-API-spend ratio for wrappers and immediately get a ranked list of what features would be most profitable to launch themselves.
These once-meteoric-now-vaporized companies just solve a first-order problem extremely well and either a) don't realize they need to make the jump to solving multiple problems in parallel, or b) don't manage to. It's tempting to say that the real risk is building on someone else's platform (since that's what killed Zynga, numerous PC manufacturers, and, to a lesser extent, the group-buying businesses that were so dependent on cheap ad inventory). But there are other high-growth cases where it doesn't apply: the pandemic's indirect effects created plenty of digital events and remote work-coordinating companies, many of whom found that demand suddenly collapsed. And, of course, it created a whole industry around rapid testing and novelty PPE, which has seen an even more severe drop in demand.
Companies that can put up steady growth over extended periods do so by identifying the future barriers to scale in advance, and knocking them down. In general, the timeline for dealing with these is much longer than the timeline for whatever the company initially succeeds at, so it requires a bit of an adjustment—now capital and time must be allocated to something with both a more distant future return and a lower return, simply because it's more strategic. Examples abound:
- Netflix could have been a lucrative short-lived arbitrage: buy streaming rights from movie studios that aren't the growth of broadband properly, get lots of subscribers, and then sell the business when it's more of a boring service provider. Instead, they went into original content, which is complementary to the arbitrage business since unique shows are a marketing hook but a familiar back catalog keeps subscribers paying.
- Facebook's peak IRR for investors was probably when Yahoo offered to acquire them for almost $1bn in July of 2006 (they were valued at ~$5m in the summer of 2004, $100m in May 2005, and $475m in April 2006). That acquisition assumed that the company's value was from its young demographic—when media companies go ex-growth, their audience starts aging until it starts dying, and they pay a high premium for youth. But the company produced much more value overall by staying independent and expanding its reach.
- Google hasn't had to pivot its business. They shouldn't. Search monetized by auction-based ads is arguably the single best business ever invented. But it's a vulnerable one, because customers are so sensitive to defaults and because browsers and operating systems have so much room to tweak those defaults. So Google has naturally engaged in a parallel-path process of building a great browser that is itself a lucrative business indirectly ($, Diff), owning a mobile OS, and increasingly fighting for the low end of desktop OSes as well.[3]
- Meanwhile, it seems that every big tech company is doing some kind of analysis of the basic units of compute they depend on, and designing custom chips that make it cheaper on the margin. Sometimes that's about on-device performance for consumer hardware, as with Apple's chips. Increasingly it's about cheaper training and inference for AI. Regardless, it's a growing trend.
- Meanwhile, some companies do this exactly backwards. Businesses built around commercializing an open-source product are a fascinating case study in creating the standard first and then finding a way to monetize it. This can work extremely well in some cases; capitalist incentives work wonderfully with the collectivist ethos of open-source when every user has a different feature they insist would make the product better—the net result is a product that can do whatever any specific user would want, but is too complicated for any of them to use, creating immense demand for either implementation consulting or hosted services, all provided by for-profit open source companies.
A fun paradox of history is that leaders who preside over generally good times get few, short books written about them; the big topics are the people who are in charge during, or are the direct cause of, absolute chaos. Similarly, there's not much to say about the business acumen of someone who rides a trend when it's a good trend to ride: they were either very lucky or very smart when they chose their life's work, and after that choice made one fairly obvious decision after another. Things get interesting when the future gets murky, and when the best option is to choose among differently disappointing alternatives—one high-risk plan to reduce one constraint, another operationally-intensive effort to mitigate some other problem. But that's ultimately what growth is; the straightforward, single-variable part is brief, but in the long run, the real world is multivariate and uncertain, and every well-run great business is diversifying into a merely good or actively mediocre one just to stay alive a bit longer.
Disclosure: Long AMZN, MSFT, META. (There is a correlation between buy-and-hold companies and anecdotes that illustrate the difference between durable and one-time growth.)
This starts out with comparatively minor concerns, like the tradeoff between doing something server-side or client-side—you're dealing with questions of performance and cost, but also the question of whether or not you're losing users who are on devices less powerful than what you're using, and the further question of whether or not you care. Over time, though, the synchronization question gets more complex. When Facebook celebrated their billion-user milestone, it was celebrating the moment they were statistically confident they had 1bn users, not the actual instant of the billionth user joining. That's the path of scaling. You go from displaying key metrics like number of users as an integer to estimating it as a float. ↩︎
One of the surprises in eBoys is just how low valuations were for scalable growth companies. In 1997, eBay was a profitable business growing over 10x annually; Benchmark was able to buy in at 3.5x forward revenue. That deal was done midyear, so it was closer to 15x run-rate revenue. Even in today's funding environment, it's not hard to imagine a company with the metrics and network effects of eBay raising money at roughly 10x the valuation. ↩︎
My kids aren't old enough to do anything especially sophisticated with computers, but the only desktops they've used are Chromebooks, and Google almost certainly realizes that a $150 computer that can access Disney+ is going to be the beater computer that kids use to discover all sorts of interesting digital things. And part of what that means is that to my kids, a computer that has a caps-lock key where the search key should be will feel a bit weird. Making search a fundamental interaction model at the hardware level is the kind of decision that has zero payoff right now but that will keep revenue per user ticking higher for decades to come. ↩︎
A Word From Our Sponsors
Intercom is a complete, AI and human customer service platform—and the only platform to combine a help desk, AI chatbot, and proactive support tools.
It also comes equipped with a fully featured inbox, ticketing, help center, and messenger, making it not only complete—but also the most powerful support platform. On the planet.
Elsewhere
AAI
The typical use cases for AI are situations where a high level of flexibility is useful and significant marginal costs are tolerable. This is also the use case for human beings; if a process is repeatedly done in exactly the same way, it's been automated already. But AI businesses get higher multiples than labor-intensive ones, which creates an arbitrage opportunity: Bloomberg reports that Presto Automation is being investigated by the SEC over "AI" products that were actually powered by outsourcing to the Phillippines. It's been a running joke in AI spaces for a very long time that "AI" often means outsourcing to countries with a low cost of living. But it's increasingly feasible to use low-cost workers as training for fully automated systems, gradually titrating the human element from ~100% to ~0. When these companies pitch themselves as AI businesses, it's a financial form of the prophetic perfect tense: they will be automated eventually, but getting there requires manual work today.
OpenAI
It's always hard to report on boardroom drama, because every claim goes through so many filters: board members talk, their lawyers tell them how to talk publicly, they make public statements that preemptively frame how other people's public statements will be interpreted, and all this gets filtered through journalists' biases. (And then The Diff adds its own spin.) Nevertheless, a many-bylined NYT article, i.e. one in which many longstanding relationships were cashed in, covers runup to OpenAI's decision to oust Sam Altman, and the subsequent fallout.
The proximate issue was a very political one: Altman was disappointed that a board member, Helen Toner, had published an article that said nice things about OpenAI competitor Anthropic's approach to AI safety, and negative things about OpenAI's. And, apparently, claimed that other board members were incensed and wanted to remove Toner from the board. At one level, this is gossip, a literal "he said, she said" situation. But it's also an interesting look at the way things get done: one way to get a decision accepted is to present it as fait accompli: the decision's been made, and it's time to accept it. This is an effective way to coordinate people, but creates the temptation to claim that things have been decided when they're still up in the air. Sam Altman's first startup, Loopt, was a mobile app launched at a time when "mobile" meant dealing with carriers rather than paying a lot of money for Facebook ads. The number of carriers is small, and they're competitive enough that a good way to get a meeting with carrier A might be to insinuate that carrier B is about to top an offer from carrier C. Repeat this process for B, C, and A and then C, A, and B, and your startup has created negotiating leverage out of nothing, at least as long as the companies don't compare notes. And this kind of process is responsible for a large amount of human progress! But it's a skill that must be used judiciously.
Another fun note from the story comes as an aside: "When news broke of Mr. Altman’s firing on Nov. 17, a text landed in a private WhatsApp group of more than 100 chief executives of Silicon Valley companies, including Meta’s Mark Zuckerberg and Dropbox’s Drew Houston... "Sam is out,” the text said." Group chats are an opaque medium, but they're an important one. They're the default medium of social interaction for people who type fast but don't get out much. They're an easy-to-underestimate force in media because they're so illegible, but no one got very far ignoring the medium with the highest viral coefficient.
Twitter has struggled to keep large advertisers happy, and is targeting smaller ones in response ($, WSJ). The basic way that auction-based online ads work is that direct-response sets the floor and then branded ads determine how high above that pricing goes. There are some low-cost general-interest products that will always be profitable to advertise at a sufficiently low cost per click, so an ad business can't quite get to zero. But it can get quick close if the advertisers willing to offer a premium choose to leave. Twitter's issue right now is that its audience is, at least some of the time, a valuable demographic that will amplify the message in an ad (especially when they can be confident that if that ad were false, they would have heard about it ($, The Diff). But those are exactly the sorts of advertisers who will drop a platform because its owner said something controversial and didn't walk it back in the appropriate way. The more small-dollar, conversion-focused advertisers there are, the smaller the hit from that kind of PR issue.
Pod Shop Math
Moving up the ladder of abstraction and complexity doesn't just happen in tech companies. The hedge fund business has transitioned from managing portfolios of assets to managing portfolios of alpha generators, and the science of doing so optimally has yet to be settled. This Bloomberg piece on a new fund, Freestone Grove, focuses mostly on how the company is arranging its teams in order to reduce crowding and maximize returns. Having a hundred teams is one kind of diversification, but it's also a way to accidentally implement roughly the same strategy with 100 different slightly-overlapping sets of assets. Freestone Grove is trying to deliberately cap the size of teams, and of assets under management, in order to mitigate this. Diversification is an economic free lunch, but only if it's implemented perfectly, and the nature of leverage and crowding means that a portfolio is never as diversified as it looks.
Privacy, Security, and Monopoly
Beeper, which reverse-engineered the iMessage protocol in order to allow Android users to send iMessages, has been blocked by Apple. Beeper's argument is that their product is a security improvement, since it encrypts messages that would otherwise be sent in plaintext. Apple's argument is that this is bad for security because it involved finding a workaround in iMessage's security, which might be further exploited. In theory, algorithms are more trustworthy than corporate policies, and we ought to rely on them. In practice, even fairly technical users in high-stakes situations sometimes get lazy when they have to handle this kind of security manually, and the biggest improvements in security tend to come from companies forcing users to adopt end-to-end encrypted messaging, HTTPS, and the like. But relying on companies for this also means subjecting things to those companies' own incentives; Apple likes the integration across its hardware businesses, and isn't especially excited to give that up in order to provide a security improvement (and, yes, blue message bubbles) to people who buy from a competitor.
Diff Jobs
Companies in the Diff network are actively looking for talent. A sampling of current open roles:
- A diversified prop trading firm with a uniquely collaborative team structure is looking for experienced software engineers. (Singapore or Austin, TX preferred)
- A vertically integrated PE-backed cannabis company is looking for an Excel wizard with a background in supply chain. (Remote)
- A concentrated crossover fund is looking for an intellectually curious data scientist with demonstrated mastery in analytics. Experience with alt data, web scraping, and financial modeling preferred. (SF)
- An early-stage startup aiming to reduce labor costs by over 80% in a $100bn+ industry is looking for a part-time technical advisor with robotics experience; this has the potential to evolve into a full-time role. (NYC)
- An AI startup with a new product targeting the financial services industry is looking for a frontend engineer with React and Typescript experience who is interested in figuring out the right user interface for something that's never been built before. (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.