Cultural Stasis? Or Just Rising Budgets, a Limited Supply of Good Movie Release Dates, and the Kelly Criterion?

In this issue:

• Cultural Stasis? Or Just Rising Budgets, a Limited Supply of Good Movie Release Dates, and the Kelly Criterion?—It's undeniable that sequels and big franchises dominate movie ticket sales more than they used to. But the economics of theatrical releases in particular are peculiar in a way that makes this a rational response to incentives—and in recent years, those incentives have actually shifted in a different direction.
• MTurk, RIP—It's hard to fight AI on two fronts.
• The Economist Cover Indicator—It turns out that The Economist is good at exactly the predictions you'd want them to be good at.
• AI Cycles—The companies with a shortage of users still have a surplus of infrastructure for training.
• Narratives—So far, the AI job loss story hasn't panned out, at least in the aggregate. This gives labs a much more pleasant story to tell.
• Intrusions—Hackers probably can't afford as good an AI tool as you can.

Talk to this post on Read.Haus.

Cultural Stasis? Or Just Rising Budgets, a Limited Supply of Good Movie Release Dates, and the Kelly Criterion?

It's commonplace to observe that, at least as far as movies are concerned, we're completely culturally stuck. In 2025, the top twenty-five movies by US box office gross included two originals, Sinners and Weapons, but the rest of the list includes a movie based on a video game, a live-action remake, a comic book movie, a sequel, a sequel to a spinoff, another comic book movie, another live-action remake, another, and so on. (It was actually more striking in 2024: we had the fourth entries in the Despicable Me and Kung Fu Panda franchises, both of which produced more US box office revenue than any original that year).

The franchise/sequel/spinoff/remake economy has some obvious logic to it; if you're trying to sell a movie outing to a group of people, it's a nice shortcut to make a movie that's targeted to kids based on something that a parent would be nostalgic for: about half of the high-grossing remakes of the last ten years were remakes of franchises 20-40 years old, the perfect range for a nostalgic parent to recreate their own childhood experience of watching Aladdin or The Lion King for the first time.^[1]

And we can actually model this! Sequels, remakes, adaptations and the like have a narrower range of outcomes than originals; Obsession's total box office was ~500x its production budget, which is a hard return to pull off if your denominator is bigger than their $750k. But studios can't just toggle a risk/reward dial, because they have another problem: there's a finite supply of good days to release movies. The formula, pioneered by Jaws, is: release your movie on as many screens as you can, accompanied by a marketing campaign that crescendoes just before release, then hope word-of-mouth keeps movie ticket sales going for a few more weekends and then leads to home viewing down the line. (The home viewing component has obviously changed a lot since the release of VHS tapes, DVDs, and then streaming, etc. It, too, has become more of a manufacturing business than a hit-driven one; when Jaws aired on ABC in 1979, 80 million people watched). Releases tend to be on Fridays, so there's a full weekend of viewing and compounding word-of-mouth; Wednesday has worked for very heavily-marketed movies (Phantom Menace, Return of the King).^[2] Since 2021, eleven release dates each year have accounted for half of box office gross.

So we're actually doing a strange kind of Kelly-flavored betting: instead of choosing how much of our bankroll to bet, the question is, given a relatively fixed bankroll for betting, and a limited number of bets, how do you optimize a portfolio of movies?

Granted, this is not exactly what the studios are doing. The movie business is full of gamblers in the sense of people who risk set sums for uncertain rewards, but these are not the same kinds of gamblers who find their way into prop trading. Which is actually quite reasonable: if you get good at poker you find yourself asking a life-changing question like "Call option? What's that?" your next step is to move into a business with a similar feedback loop.^[3] But even if you're in a sample size-maximizing position like a reader at a studio, you're looking at roughly two scripts a day, compared to a poker player playing a few hundred hands a session. And the poker player is much less likely to fold with a given poker hand than a script reader is to recommend a particular script. From a purely statistical standpoint, a producer gets about as big a sample size of affirmative decisions over their career that a poker player could get from a long weekend. But Kelly is descriptive as well as prescriptive; it tells you what kind of betting is optimal, but that means it also tells you something about what kind of betting has led to certain track records: the people who persistently bet more than full-Kelly went bust, the ones who persistently underbet are underrepresented in the sample, and, by process of elimination, Disney is home to the movie industry's best advantage gamblers even if that's not strictly what their job title says.

Hollywood has a literal concept of options, but the figurative one—a future opportunity to buy or sell something at a fixed price—is relevant here. Every completely original movie, and many movies based on existing IP, is basically a call option on some number of sequels. The option in question is hardly vanilla; the strike price, in the form of the movie budget, tends to rise over time. But this dynamic gives studios an incentive to bet on original movies that might have modest initial returns but produce some low-risk sequels.

But the other force is: once they have a big portfolio of franchises, the squeeze of higher production and advertising costs and a finite number of good release weekends means that they're forced to bet a bigger chunk of their bankroll, at least if they're going for a major theatrical release.

If you backtest this—grab each year's top 200 releases from Box Office Mojo, tag each one based on whether it's a sequel, existing non-movie IP, or completely original, then get budget estimates (noisier), and backtest a Kelly-betting strategy based on trailing five-year windows, this is in fact what you see: movie studios were actually underinvesting in sequels in the mid-80s, compared to what was optimal; a backtest sensitive to historical performance, budgets, and the finite number of good movie launch windows suggests that sequels should have been about half of revenue-weighted movie releases in the mid-80s through early 90s, compared to an average of about 30%. (Interestingly enough, this same backtest starts to show higher returns for original properties later in the 90s. Look at 1999's top-grossing domestic movies and the first thing that stands out is that, compared to today, far fewer of them are connected to big properties, though some of them certainly are. And, in retrospect, many of them turned out to be valuable sequels.)

Building this was a surprisingly complicated process, even with LLMs making a more than token contribution to tagging movies by IP source and the like. For example, Star Wars is not just a bouillabaisse of all visual pop culture from George Lucas' childhood. It was also, technically, a movie based on a book: Star Wars: From the Adventures of Luke Skywalker, which came out a few months before. So that had to be overridden. And the average time to first sequel for big franchises also looks unnatural if you accidentally treat the 1986 animated Transformers movie as the start of that franchise. It's only as a result of cutting things off with full-year box office results that we don't have to deal with The Mummy, the Franchise of Theseus that had a run in the 30s-40s, a parody in 1955, a reboot in the late 50s, an action reboot in the late 90s, and then a horror reboot this year.^[4] And there are other complexities, like genre send-ups, but those are probably their own phenomeon.^[5]

The picture that starts to emerge here is one of a long cycle between originals and rehashes, as well as a secular trend towards the lower-risk option. The cycle happens when there's a large number of original properties in the recent past, or when there's some new kind of intellectual property that can be strip-mined in a systematic way—like taking identifiable comic book characters and making them grittier (Nolan's Dark Knight movies) or by brushing the grit off and making them more family-friendly (rebooting Spider-Man as a series of animated movies rather than live-action). Taking a franchise that has a big corpus in other media, but minimal film presence and making it into a movie franchise, as is being attempted with the Warhammer 40K universe, is the media equivalent of fracking, where you amortize opportunities you already own and have good information on adjacent possibilities. Sometimes, it doesn't quite work out, but sometimes it means that there's an extra layer of extraction that gets continuously more predictable over time.

All of this is, in some sense, a nostalgic discussion of the movie business as it used to work and never will again. An analysis from the 80s to today misses home entertainment, the globalization of entertainment, streaming, and short-form video. At this point, theatrical releases are still a real business, but one that's approaching an asymptote where it's just another IRL pop-up experience for a mostly online brand. But there are some practical forces that keep theaters interesting; they're the most absorbing way to experience a movie, so they're the default distribution medium for the best movies. And even if streamers take the logic of sequels to its natural conclusion and use big movies as an anchor for a series of lower-budget, lower-effort online sequels (or, if it's someone else's success, as inspiration for something similar). Even if most of the enterprise value is tied to streaming subscriptions, there's still a case for having a big event, getting a lot of novelty popcorn containers into fans' hands, etc.

But in another sense, it's forward-looking: sequels peaked as a share of box office gross in 2021, and have been declining since then. And the model's estimate of the optimal level of sequels has also ticked down; it was around 90% from about 2010 through 2020, but the big franchises are getting tired, and this year has already had two big high-ROI originals, Obsession and The Backrooms.^[6] The movie business might have been easier and more predictable if there were a dozen James Bond-level franchises instead of just one. Studios tried something like that, it worked kind of well, but eventually both studios and audiences concluded that things work a little better when audiences sitting down at a new release aren't quite sure what to expect.

1. He-Man is an interesting one, possibly a testament to later ages of first marriage and IVF extending fertility; I was born just late enough that my first exposure to the He-Man franchise came from Bonfire of the Vanities. ↩︎

2. The other big Wednesday release was tied to the liturgical calendar: The Passion of the Christ came out on Ash Wednesday 2004. ↩︎

3. Similar both in the sense that you get a big sample size that gives you evidence of whether or not you're skilled, and in the sense that there are regions of the left tail of the return distribution that you just won't understand until you've lived through them. ↩︎

4. It's out of the scope of this analysis because the economics are more opaque, but one of the most valuable parts of The Mummy's 90s action reboot was that it was the first big movie role for a pro wrestler turned actor, The Rock, who has since starred in movies that produced over $10bn in inflation-adjusted gross. But presumably the bigger a star one person is, the more the upside from their discovery accrues to them. If anything, people acting their hearts out to get their big break are the ones who are most underpriced in any given film. But that's going to be true even if they're rationally pricing the optionality of potential stardom. ↩︎

5. Specifically, genre parodies seem to work when there's widespread awareness of genre tropes. Blazing Saddles could tap into an audience that had grown up on, and grown out of, Westerns. Star Wars was unavoidable enough to make Spaceballs big. Airplane was riffing on a series of Airport movies, and the idea of a disaster in the air was very much in the air because the movie came out after a decade of elevated hijackings. But the rise of the trailer, and VHS rentals, gave audiences a broad enough grounding in media literacy that they could enjoy Scary Movie, Austin Powers, Not Another Teen Movie, etc. These movies are essentially a way for a third party to borrow someone else's sequel economics. But artistically, they're a net positive because they force the genres they're parodying to choose between taking the material seriously or somehow trying to differentiate self-aware camp from complete parody. James Bond is still around, but the women he's interacting with don't have names like "Xenia Onatopp" any more. ↩︎

6. The latter is not completely original, of course, since it's based on an old 4chan thread. As a practical matter, this is closer to basing a movie on folklore than turning a book into a movie. And as an economic one, if there's nobody to pay royalties to, it's an original property by default. ↩︎

Diff Jobs

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

• Lightspeed-backed team building the engineering services firm of the future is looking for founding members of technical staff excited about working alongside civil engineers to translate their domain expertise into the operating system that powers the next era of great American infrastructure. If you’re an engineer with strong product intuition, who's energized by access to users, and excited by the prospect of transforming how we design and construct our built world with frontier AI, this is for you. (NYC, SF or Remote)
• Ex-Palantir, Citadel founders building the meta-harness (the system that knows what hills to climb, and what the right loss functions are) for all the lucrative agents need full-stack engineers that understand that turning AI into economically valuable solutions means a system that includes deterministic infrastructure. If you’re curious about building a system that finds what the efficient frontier between determinism and stochasticity actually is, this is for you. (NYC)
• Series-A defense tech company that’s redefining logistics superiority with AI is looking for a MLE to build and deploy models that eliminate weeks of Excel work for the Special Forces. If you want to turn complex logistics systems into parametric models, fit them using Bayesian inference, and optimize logistics decision-making with gradient descent, this is for you. Python, PyTorch/TensorFlow, MLOps (Kubernetes, MLflow), and cloud infrastructure experience preferred. (Salt Lake City or NYC)
• A well-funded, Series C startup building the platform and agent primitives to drive operational transformation at large, complex institutions (starting with higher education) is hiring platform engineers. The work spans distributed systems, applied AI, and full-stack infrastructure, focused on deploying reliable agents that meaningfully bend institutional cost curves. (Remote)
• A hyper-growth startup that’s turning the fastest growing unicorns’ sales and marketing data into revenue (driven $XXXM incremental customer revenue the last year alone) is looking for a senior/staff-level software engineer with a track record of building large, performant distributed systems and owning customer delivery at high velocity. Experience with AI agents, orchestration frameworks, and contributing to open source AI a plus. (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.

And: we're now actively deploying capital into early-stage companies through Anomaly. Our focus is on defense, logistics, robotics, and energy. If you'd like to chat, please reach out.

Elsewhere

MTurk, RIP

Amazon is preparing to shut down its Mechanical Turk service. Before there were many APIs for tasks like "read these thousands of customer service complaints, and put them in the following categories," there was Mechanical Turk, which was a successful bet on the market for repeat microlabor, with a long tail of clever applications. One problem Mechanical Turk has is that it competes with AI directly, and even the cheapest global labor pool is not as effective as a GPU in turning watts and tokens into more tokens at some quality standard. Another problem was that, by the same token, using LLMs to do Mechanical Turk tasks was more profitable than using humans. For sufficiently small outputs, LLM detectors will have too high a false-positive rate to be useful. For analyzing outputs at scale, one problem is that the right tail of industriousness for the most frenetic human question-answerer probably overlaps with the left tail of laziness for the most diffident copy-into-ChatGPT-and-paste-outputs-er. In a pre-AI world, Mechanical Turk was a fun experiment in globalizing the market for knowledge work, one tiny task at a time. But today, its revenue case is weaker and, unlike the pure data-labeling companies, it wasn't designed for an adversarial environment in which treating AI outputs as human labor is the cheat code to easy money.

Disclosure: long AMZN.

The Economist Cover Indicator

A few times, in a few different contexts, that the rise of AI is going to be retroactively informative: pseudonymous writers can be linked with their real-name corpus, student whose plagiarized essays got past today's AI detectors will probably fall to one eventually, LLMs can transcribe handwritten historical documents and make the same kinds of contextual judgments that specialist historians would, etc. They're also handy for large-scale data analysis; you might have a vague sense that some pundit is overconfident, but not have the time to go through all the details—it used to take a seriously committed hater to do that. But now, it's a lot simpler, and all it takes is a small budget and a little curiosity. So The Economist has gone back and looked at all their lead articles that made a prediction, and judged that prediction based on how mainstream it was and how accurate it turned out to be ($). What they found was that their middle-of-the-road predictions are right, and that their out-of-consensus ones are less reliable. Which is as it should be: a mass-circulation publication needs to articulate a reasonably common view well, and probably shouldn't risk accidentally sparking a new intellectual fashion among the elite. It's actually a stricter burden for them, because they're read by both businesses and the governments that regulate them, so they tend to make things technocrat-approved-or-not instead of some more partisan-flavored, and thus more random, outcome. Managing things within the Overton Window is a useful task, and they do it well.

AI Cycles

In an internal memo, Meta says its latest models are on par with GPT-5.5. It's another instance of the variable-lag model of AI lab dominance: whoever has the best model has the least spare compute for training the next model, so among labs that have funding and can attract talent, the status of #1 lab is always in flux. For Meta as a consumer of AI and as a filtering mechanism for social media noise, it's useful to have open-weight models being hosted by as many providers as possible, but for Meta as a prospect seller of AI, it's nice to have a model that's ahead of the competition, rather than matching it. Meta was able to harvest cheap data early in its existence, but now that intent data gets generated and higher fidelity—if you're wearing Meta's glasses, they literally know what catches your eye—they have to be close to the bleeding edge just to find out what AI users are looking for.

Disclosure: long META.

Narratives

When AI companies were small and nonthreatening, one of the best narratives for them to use was that their addressable market was every application of intelligence, which basically meant every white-collar job. Looking at early prototypes, that's the best way to model some technology—look at the demand for what it's the closest replacement for, and extrapolate. But capabilities are jagged; it turns out that electrification temporarily replaced one source of demand for oil, kerosene, but coincided with a larger one, the car industry—and the fact that automotive mass-production scaled after factory electrification was well underway meant that they were able to add some oil demand as a direct result of another technology that subtracted from it. The AI labs now have a similar narrative ($, WSJ): there's been surprisingly little aggregate job destruction so far, though there's definitely a supply/demand mismatch for people early in their careers. It's politically advantageous for the labs to spin this story, and it's fortunate for them that the data support it.

Intrusions

In other AI news, there's a new attack that connects to an LLM in order to manage its intrusion. This is in some ways more flexible a deterministic approach that's limited to running whatever code it can got onto a machine, or to getting manual updates from an external server. But in another sense, it's more brittle: there's a lot more money in running defensive models that manipulate adversaries like this than there is in being an attacker.