Who Hedges Where?

In this issue:

• Who Hedges Where?—Most funds are not meant to be the only investment someone makes, so it's worth thinking about investors' idiosyncratic preferences. If you're worried about AI, either as a threat to your own business or as a threat to all of humanity, you might conclude that an investment that pays off if AI accelerates is more of a hedge than a source of absolute return.
• Ads for Bots—It's probably load-bearing that companies can turn money into changes in consumer perception.
• Authenticity—If AI wins A/B tests, it can still be a strategically sound decision not to use it.
• The AI Race—OpenAI has slowed, in a relative sense, but that also means that they have more obvious strategic responses.
• Priority Disputes—Mathematicians join software developers in the somewhat queasy category of jobs where asking the right questions is enormously valuable because human workers no longer provide the best answers.
• Mission—You can solve for the problem of mercenary employees by paying less, but that leaves you with the harder problem of justifying these pay scales.

The Diff August 3rd 2026

0:00

/738.977959

1×

Talk to this post on Read.Haus.

Who Hedges Where?

One fun way to look at Situational Awareness is that it's actually doing precisely what it's supposed to do. Suppose you're a very wealthy person who, like many wealthy people lately, got that way by being good at telling computers what to do, and then good at telling other people what to do with their computers, to the point that you have a substantial net worth in the form of your ownership of some sort of tech company. One problem you might have—a top 0.01% of the First World sort of problem—is the risk that as AI improves, and develops new skill spikes, it might suddenly render that business completely nonviable. Most of these people would still be rich, but the least fun form of centimillionairehood is the former-billionaire kind.

So they can either spend lots of time worrying that AI will swallow their economic niche, or they can hedge. One of those hedges is to adopt it, and my guess is that if the Forbes 400 had a personal token consumption leaderboard, Situational Awareness' LPs would be high on the list. And to be on the Forbes 400 in the first place is to have made more money than you could possibly spend on yourself, and to keep going. These people are competitive!

Some of them fit into a different category: they're very bullish on AI in the sense that they can extrapolate a hockeystick chart of capabilities just as well as anyone, and worry about the day somebody Googles "population of Cleveland" and the AI realizes it can't possibly count all those people accurately, so it seizes control of a nuclear weapon or synthesizes a novel pathogen and then proudly announces: "Worked for 4h 7m 33s: Zero." What this kind of investor wants, particularly in a fast-takeoff scenario, is to have the maximum possible financial resources in a condition where AI is going too well, so if there's some way to reduce the risk of human extinction by deploying large sums, they'll be in a position to do it.

So one thing these investors can do is to allocate money to a very levered strategy that is, from their perspective, offering a bespoke hedge against a specific outcome that's bad for them. Tail-risk hedging, as a strategy, has some odd characteristics:

1. Depending on the risk being hedged, investors may want poor returns from these hedges. The more generally costly a disaster, the more likely it is to have some unforeseen negative consequences. So the bigger the risk being hedged, the more likely it's still underhedged.
2. Strategies like this can be accretive to portfolio returns even if they're negative expected value on their own. This goes back to the conditionality point: one reason to continuously roll deeply out-of-the-money index puts is that during times like March 2020 or April 2025, your portfolio doesn't go down much, but the other reason is that selloffs like that are a generous banquet for liquidity providers and stock pickers. Spreads are wide and volume is frenetic, so a trader who can keep making a market will earn fabulous profits, albeit with more variance than usual. And during big selloffs, people don't always game out which companies are actually beneficiaries: Boeing dropped 19% in two days after Liberation Day, but two months after the tariff announcement, it was 25% above its pre-Liberation Day price, because investors figured out that these tariffs were being negotiated by a fairly literal-minded President who was also fond of America's globally-recognized brands—a little nepotistic given that he is one—and that a big order to Boeing was a great way to cut tariffs now in exchange for promising to spend money some time after January 2029. Similarly, Peloton dropped by a third in a month when Covid hit, and was trading at an all-time high by May 2020.

This is all fine in retrospect, though a lot of it hinges on whether that's what investors thought they were getting.^[1] And it raises a broader point, about who hedges and why. The world is net 100% long equity and real estate, and net flat with respect to debt and most derivatives.^[2] So some risks can't be fully hedged away, and others are naturally hedged in the aggregate.

One piece of useful trader wisdom, via The Laws of Trading, is that you should only take the risks you're being paid to take, and hedge out the rest. It's a pretty powerful idea, because one thing it tells you is that if there's some otherwise-attractive opportunity but it's expensive to hedge out an associated risk, that's why the opportunity is attractive. If you buy an index fund, you get paid for holding equities and facing the occasional capital loss that coincides with the risk of job loss because that's an annoying risk that a rational economic actor needs to be paid to take.

Another reason, more applicable to professional money managers rather than the average person, is that hedging risk makes it both easier to measure (and thus compensate people for) returns, and that it squeezes a little more upside out of a research process. You could imagine a structure where portfolio managers put on whatever positions they feel strongly about, and the fund itself hedges their exposure to a particular industry, or to momentum, or whatever, in order to achieve pure alpha. But there are at least three downsides to that:

1. If hedging is done centrally, then portfolio managers aren't appropriately charged for basis risk, i.e. the risk that the hedge didn't perform the way it was supposed to, or whoever was responsible for hedging made some avoidable mistake. And portfolio managers would absolutely loathe a system where they're charged for this loss even though it wasn't their fault.
2. If you have a view on a single stock, and you can't find a way to hedge out the other risks of the position—you don't have a single-stock view, just a view on the factor you're refusing to hedge.
3. In fundamental equities, people tend to have natural biases in terms of what kind of setup they're comfortable with. There are investors who are good at underwriting growth, spotting scams, assessing hidden assets, etc. And this tends to reliably put them on one side of a factor. But, if you're digging through your coverage universe looking for frauds, you will also find that there's some variance in how clearly the non-frauds do their reporting; it's not that exciting to own something on the grounds that its long-term free cash flow is higher relative to EBITDA than its peers, whereas its very exciting to short something when you realize its reported earnings clearly outstrip how much money the actual business makes. Forcing portfolio managers to take long and short positions means getting a little more upside from the research they've already done.

The big goal of risk management is to maximize risk-adjusted returns, first by minimizing the probability that those returns touch -100% and to maximize returns per unit of risk. But one of its other functions is to ensure that investment managers are getting paid in proportion to how much value they create, and sometimes the specific return stream some people are after is a weird one.

Diff Jobs

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

• High-growth startup building dev tools that help highly technical organizations autonomously test and debug complex codebases is looking for someone who can personally execute the SaaS bear case: review the third-party software they use and figure out what to keep, what to drop, and what to implement in-house. If you’re an early career engineer interested in AI transformation, this is a great opportunity for you. (SF, DC, London)
• Well-funded, frontier AI neolab working on video pretraining and computer action models as the path to general intelligence is looking for researchers who are excited about creating machines that learn from experience, not text. Ideally you have zero-to-one pre-training experience and/or are a high-slope generalist who’s frustrated that the big labs aren't doing this. (SF)
• 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-Anduril, Ex-Abnormal Security, Ex-Bridgewater, fast growing startup bringing agentic cybersecurity to 99% of businesses via MSPs is looking for platform and machine learning engineers. Startup experience preferred; what matters most is that you've grown in scope and handled ambiguity over the last few years. (SF)
• AI Transformation firm with an ambition to build an economic world model to run swathes of the private, unstructured economy is looking for Systems Engineers, Platform Engineers, and business generalists who understand how to solve problems.  

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

Ads for Bots

Time magazine has been serving some of its content to AI in markdown format, and is adding information-heavy ads to the articles. On one hand, this feels like cargo cult behavior, since the LLMs can identify ads, don't have any incentive to mix them into answers alongside article content, and probably want the revenue from LLM ads to accrue to them. On the other hand, Time is basically betting a little of their journalistic reputation on the quality of the ads, which gives them at least some incentive to care about accuracy of advertisers. Brands get perceived in particular ways in part because they're paying for that perception, but that tends to get a higher return on investment if the ad is at least directionally right. So for an LLM to capture the zeitgeist and retain backwards compatibility, it should probably retain some way for advertisers to convert money into brand perception, though where that lives in the stack is still an open question.

Authenticity

One of Facebook's early differentiators as an online community was actually that it looped back to pre-Eternal September norms: everyone's identity is associated with a place of higher education and they're identified by their real names. This technically meant that they omitted a feature from MySpace, where people would spin up joke accounts, accounts for their pets, etc. Which meant that Facebook was actually more social, while MySpace was more media. Social was the right bet. So, Snap will not highlight purely AI videos in their Spotlight recommendations. AI content clearly wins some A/B tests, since it keeps getting more common, but viewers are generally stumbling on it, not seeking it out intentionally. Snap is trying to do a little of what Facebook did long before; be the first social media platform people check when they want something real.

Disclosure: long META, SNAP.

The AI Race

The Diff has argued before that AI leadership naturally seesaws because the reward for having the last hit product is being relatively more bottlenecked by compute. That means less infrastructure for training the next model. But there's another dynamic, too: when there's a new form factor of sorts, like agents, whoever needs to catch up can throw a lot more inference at it, especially for early users. So the labs can, with fairly short turnaround, launch a product that brute-forces its way to being strictly better than competing version. At least until that approach wins and there's some other context where small market share in some category gives the challenger room to outspend the incumbent on inference. We will probably keep seeing those stories, [like this one in the WSJ suggesting that OpenAI's deceleration has bottom-ticked]((https://www.wsj.com/tech/ai/how-openai-lost-its-ai-crownand-the-fight-to-win-it-back-7d069695?st=KhR6t6&reflink=desktopwebshare_permalink) ($, WSJ). It's incredibly hard to be the top AI lab, so no one stays that way too long.

Priority Disputes

Two different GPT-5.6-assisted proofs for the same problem were published hours apart. One way to look at how AI develops at the level of some specific task is that there's a point where AI is mostly incapable of handling things on its own, there's a point where the problem in question is mostly solved, and, somewhere in the middle, there's a time when being able to ask the right question and coax out the right answer is immensely valuable. For coding, that's been a long stretch—as pre-AI programmer salaries indicated, the world generally suffered from a shortage of code. In pure math, where open problems are surrounded by token-rich research on adjacent problems, or failed attempts at solving the original, there has turned out to be a lot of low-hanging fruit. And even though researchers are worried that their field is getting solved, they're also in the very privileged position of knowing what the important open questions are, possibly having a sense for which ones haven't gotten much attention lately, and far above average skill at evaluating results. (This might go a little more slowly than usual for a cultural reason: usually, someone who says "I think I've solved a big open problem, but I'm not 100% sure I understand my solution" is a crank sending an email to an academic, not an academic talking about how they can't see any specific flaws in the Lean code the model produced.)

Mission

A weird thing about the labor market for AI research is:

1. For many of the people involved, AI is the crux of human history, and the decisions made by labs today determine whether our future is incomprehensibly pleasant or we're all wiped out.
2. They are scarce complements to the biggest capex buildout in history, and it doesn't make sense for employers to economize on them, so they're also some of the best-paid people in history.

Squeeze these together and you get high turnover between labs, partly because people slightly switch their priorities and partly because the labs themselves will sometimes make decisions (or mistakes) that someone worried about existential risk would be sensitive to. Anthropic is worried that people are joining because they're paid well, not because they believe in the mission, and presumably that's true for some of them. But Anthropic has a pretty easy answer to half of this: they can ensure that their employees aren't purely mercenary by having a policy of never making the most financially generous offer. At that point, they have to optimize for the fuzzier goal of being the most safety-conscious—and the trouble with that is that if the strategy works, and they get talent more affordably, this means that they'll be able to produce more capable models which occasionally create novel, sometimes terrifying, safety issues.