In this issue: - Rehabilitating Bubbles—Bubbles have a bad reputation, but they're better than they look. The classic argument for this is that they lead to the buildout of valuable infrastructure, but that's only part of the story. The buildout would happen eventually, but locking on to a self-reinforcing narrative is also a way to build things that simply wouldn't get done at all without it.
- Scaling—What happens if the next generation of AI models is not such a big leap in capabilities?
- Deployment—AI isn't just adding market cap.
- Usage—The tradeoff between optimizing for average users and targeting power users.
- The Arbitrage Cycle—Strategies follow cycles, and sometimes the stars align.
- Metrics—Sometimes there's an incentive not to know exactly how an investment performs.
Rehabilitating Bubbles
Barring an economic miracle, at some point AI stocks will go through the same cycle as any other theme that's received positive investor attention: there'll be a drawdown, like the one Nvidia had over the summer where its shares dropped just over 30% within two months. There will be a recovery, but not quite to the old record; maybe one of the big labs pauses training for a bit to reassess its economics and focus more on getting existing models cheaper and more widely deployed, perhaps the other labs are grateful for the opportunity to stop shoveling quite so much money, quite so fast, into new models. And then dominoes start falling: financial models get updated; spending gets cut; fabs, datacenters, transmission infrastructure and power plants get delayed or canceled; the assumption that models will continuously improve falls apart, and valuations crater with it.
Or it goes another way: AI replaces a broad swathe of call center jobs, and replaces a variable double-digit percentage of what white-collar workers do all day. But other technologies have done that; Excel and high-level languages have automated a lot of what accountants, analysts, and software engineers used to do all day (while still leaving room for high returns from those once-required skills). So the demand is there, but it's a $20/seat lift for one set of existing SaaS products, and a 10 point gross margin headwind for the other. The 🪄 gets displayed a little less prominently in each software update, and eventually vanishes; users have some preferred model that they're already feeding all of the relevant information, so they don't really miss it.
That's a demand-side story, but the outcome is the same. Spending cuts feed on spending cuts, investors revise their priorities, developers decide they're not going to master linear algebra after all. Suddenly everyone needs a new reason to buy stocks at near-record valuations, and, while they're coming up with such a reason, they start to sell. The market drops, big tech drops more, smaller companies that rebranded as AI bets vanish without a trace, and bigger companies quietly drop their AI narrative. It wouldn't be a crash without a few frauds; whoever's fudging the numbers about their inference costs, or claiming they have their own model when they're using someone else's API for everything, will end up getting caught once they can't raise money to keep things going.
That's when books get written and narratives calcify. Economic sectors reach their peak salience when the industry is either at peak valuation or just starting to decline, and they don't get thought of much after that. (At one time, the most-watched episode of television in history was Who Shot J.R.?, a soap opera episode about the murder of an oil baron. In the 2020s, oil just isn't synonymous with money, and for a similar punch your show would have to be about someone in tech, or just maybe finance.) It's a natural selection effect, but it means that bubbles tend to get a bad rap. Meanwhile, their upside—the rapid deployment of home appliances in the 50s because of electrification that kicked into high gear in the 20s, streaming video taking off in the 2000s because of excessive fiber buildouts in the 90s, or cheaper energy today because frackers spent every dollar they could raise in the 2010s—all get treated as laws of nature. Of course a middle-class family can afford air conditioning; why wouldn't we be able to watch Lazy Sunday online the day after missing S&L; what do you mean "aren't you glad oil isn't $175/barrel?"
Boom, which Tobias Huber and I co-wrote over the last few years, is an attempt to rehabilitate bubbles and make the case that they're good. Not all of them, not always right away, and not without a lot of waste and mismanagement along the way. But, overall, bubbles are good.
And not just because of the spillover effects from overbuilding infrastructure in one cycle and learning to use it well in the next one. That's just a timing issue, and if that were the only reason, it would be a good idea to slow them down so we'd get to the same destination a bit slower but at much lower cost.
What makes them important is that they lead to investments that, in the absence of hypomanic optimism on the parts of both builders and their backers, wouldn't happen at all. A big reason for this is that bubbles are a signal to build in parallel: if Nvidia is making faster chips, it's imperative for a competitive foundation model company to figure out how to take advantage of them. If OpenAI, Anthropic, and Meta are all going to keep incrementally leapfrogging each other, anyone building a wrapper needs to be dreaming up features that won't work now but will work soon. End users in the workplace need to know which of their job functions they can automate now, because those are the ones their employer may automate for them otherwise.
Bubbles are situated in time; the more they promise to change the world, the less plausible it is that they happen twice. And they tend to drag everything adjacent to them towards whatever their operating cadence is. Physical retailers in the late 90s had to get faster at updating prices because even if other physical retailers were still slow, online sellers could update everything they cared to in real time. Today, AI's biggest spenders are trying to persuade utilities to make a similar leap in how quickly they adapt to changing demand.
Speedrunning the development and deployment of a technological advance is a necessarily uncertain process. There's always some scaling factor that ends up gating growth, and these tend to be obviously only in retrospect. Consider Bob Metcalfe's 1995 prediction that the Internet would collapse in 1996. Metcalfe knew a thing or two about networking, and one of his concerns was that TCP/IP couldn't scale. This was actually a lively topic of debate for a while, but the fundamental problem turned out to be something different: the overbuilding of both physical and digital infrastructure assumed behavioral changes that would take a while to materialize. The transition from reading on paper to reading on a screen was fairly fast, especially when the screens made it easier to find exactly what you wanted to read, or to follow breaking news. But the transition to spending through computers—typing your entire credit card number into a computer and trusting that it wouldn't be stolen by nefarious hackers or by the company itself—took a bit longer.
Bubbles always overshoot, and that's one reason they're so hard to bet against. They're also hard to bet on, because they involve a natural form of portfolio concentration: participating in one means betting money, but also reputation. The bubble bets will be the most interesting and controversial thing in your portfolio, so it's also a bet of attention. Market inefficiency is only preserved when there's a good reason not to make the obvious trade; since it's always easier to be a cynical bystander, bubbles tend to be under-allocated capital until close to the peak, though they do eventually overshoot. What we're left with is infrastructure, a shift in wealth distribution, actually useful new technology that's been pushed into as many hands as possible in order to amortize the fixed cost of developing it, and an N-of-1 cycle. And history is determined by a long series of N-of-1 events.
Disclosure: Long META.
Diff JobsCompanies in the Diff network are actively looking for talent. See a sampling of current open roles below: - 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. (NYC)
- An AI startup building tools to help automate compliance for companies in highly regulated industries is looking for legal engineers with financial regulatory experience (eg. SEC, FINRA marketing review, Reg Z, UDAAP). JD required; top law firm experience preferred. (NYC)
- A well funded Google Ventures backed startup founded by SpaceX engineers that’s building data infrastructure and tooling for hardware companies is looking for a staff product manager with 5+ years experience, ideally with AI and data intensive products. (LA, Hybrid)
- Ex-Ramp founder and team are hiring a high energy, junior full-stack engineer to help build the automation layer for the US healthcare payor-provider eco-system. (NYC)
- A well-funded startup that’s building the universal electronic cash system by taking stablecoins from edge cases to the mainstream is looking for a senior full-stack engineer. Experience with information dense front-ends is a strong plus. (NYC, London, Singapore)
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
Scaling
And speaking of bubbles that run for a while and then collide with some fundamental obstacle, delivering on their early promise but undershooting the most ambitious dreams of their biggest fans: OpenAI's current model, still in training, isn't improving as fast as the company expected, and doesn't appear likely to be as big an improvement as GPT-4 was over GPT-3 ($, The Information). There are still plenty of open questions: this is a story of one lab's subjective assessment of the difficulties ahead, and it's entirely possible that different labs have unlocked different ways to deal with the data wall problem. (If you were very well-networked in Silicon Valley during the heyday of Moore's Law, you would have gotten used to hearing that that law, too, had hit a wall, because that's the subjective experience of anyone who doesn't keep up.)
On one hand, this may make the AI Safety complex relatively less important, but on another there's still an AI bull case at current capabilities. In fact, there are several: if general-purpose models hit a wall, we'll have better predictions for how specialized models will scale, and that means more accuracy in predicting the right training budget for those. The existing backlog for GPUs, and already-completed R&D for specialized inference chips, means that the marginal cost of inference has declined. And org charts haven't changed to account for how much roles change when intelligence is just another system resource like memory. There are many tasks for which the following facts hold true:
- You could do it in about five minutes.
- Claude could do it in about two seconds, if
- You spent some uncertain amount of time, maybe minutes, maybe hours, automating this.
Everything that fits that description is some kind of market inefficiency, and the more focus there will be on getting models deployed everywhere they're useful. That focus on ROI has a lot of return ahead of it, but also implies a bit less activity on the investment side.
Deployment
The WSJ has a good look at one place where AI was deployed fast, to great effect, profiling the 99% decline in Chegg's market value after ChatGPT's release ($). Chegg was in an awkward position: they provided what might politely be called study guides and what's more directly a method of cheating, and, for a while, business was good. But the kinds of schools where Chegg is a useful tool are teaching broadly-available knowledge of the sort that works its way into training data, so ChatGPT was accidentally a great competitor. There's plenty of academic work that LLMs can't handle quite yet, though benchmarks are topic-specific (here's a great one for economics). But the broader access to college is, the more it has to involve schoolwork that the average student is capable of. And on intellectual tasks, that average student is a slower, less accurate ChatGPT, with a bit more personality.
Usage
A semi-common complaint online is that you'll open an app, see something interesting, start to read or watch it—and then the app will refresh your feed and it'll be lost forever. Meta is getting rid of this, at least in Instagram. Interestingly enough, they say that this change actually reduces engagement. As with many other products: if you adopted it early, you're likely to have strong opinions about it, putting you at the tail of some bell curve. As the app grows, more of its growth comes from the middle of the bell curve, where users think and behave differently. But that's not a universal guide to always pander: Meta is doing what power users ask for even if the average user dislikes it, because those highly-engaged users are more likely to create the content that everyone else logs in to see.
The Arbitrage Cycle
Trading strategies follow a cycle driven by both industry dynamics and exogenous shocks. Strategies that work regularly get crowded, which temporarily improves returns and then leads to blowups, and sometimes there's some feature of the real world that makes a particular approach work better, whether it's the China-driven emerging market growth of the 2000s or big tech's compounding returns to scale in the 2010s and onward. Merger arbitrage is lucky enough to be going through both right now: mergers are more likely to go through under a Trump administration, and a few recent deal blowups mean that there's less capital in the strategy than there was a few years ago ($, WSJ). And this matters for the industry's long-term future: arbitrage is overrepresented in the early careers of many successful investors, since it's a good training ground for thinking about both fundamentals and adverse selection.
Metrics
There's a now very old argument that illiquid assets get too much attention from investors because they look artificially good on a risk-adjusted basis. If they're illiquid, there isn't a market price, so the owner of the asset has to make a guess as to how well it's done. And those guesses tend to smooth out volatility in both directions; there's an incentive to understate declines in value when peer companies decline, and a good way to bank valuation ahead of that is to avoid marking things up too aggressively in the meantime,. That means that the same company, run the same way, would contribute less risk to a private equity portfolio than to a public one. Now quants are trying to create better measures of what true risk-adjusted returns are. It's a worthy project, but there are two difficult problems:
- One of the adjustments that needs to be made is the cost of illiquidity. But this can only really be measured by looking at market prices, and the market for illiquid assets breaks more severely when there's a downturn. So any such measurement is very sensitive to what time periods it includes, and how severe the drawdowns were during them.
- There's an incentive problem, where both private equity managers and their limited partners would prefer that stated volatility remain low, at least as long as they can deliver returns in the end.
|