mailwal

Doing Business in the Dark Forest

Byrne @ The Diff
By Byrne Hobart • 19 Oct 2026

In this issue:

  • Doing Business in the Dark Forest—It's easier than it used to be to put together disparate clues and uncover secrets. So everyone is, at multiple levels, going to get quieter in any venue that AIs might be trawling through.
  • Stock-Based Compensation—Normally, the way a company ends up with high employee ownership post-IPO is that they've been paid partly in stock all along. But that's not the only way.
  • Infra—The nice thing about the infrastructure bear case is that you can be negative because there's too much capacity, or because there's not enough.
  • Balance Sheets—Capex booms propagate through supply chains until they hit bottlenecks. They do the same thing with balance sheets.
  • An Ownership Society—Maybe it's a good thing if your first exposure to stocks is that it's an alternative to sports betting.
  • The Sonic Boom of Trust—AI lets us uncover secrets, including the secret that there wasn't much of a secret after all.

Talk to this post on Read.Haus.

As regular readers may know, The Diff also has an AngelList syndicate, which raises capital for a selection of promising early-stage companies.

Doing Business in the Dark Forest

The Dark Forest Hypothesis is a recently-repopularized answer to Fermi's question: if we know life can arise on at least one planet, and we know that there are a lot of them, isn't it suspicious that we aren't hearing any radio broadcasts from them? Various answers have been proposed, but one of the ominous ones was popularized in, and named by, The Dark Forest: intelligent life forms have run the same calculation, and think that if they're detected, the species that detects them will worry that they're hostile and launch a preemptive strike—and that it's a good idea to constantly look out for signs of other species in order to get them first.[1]

When I read about it, the first thing I remembered was this Paul Graham essay:

If you ever do find yourself working for a startup, here's a handy tip for evaluating competitors. Read their job listings. Everything else on their site may be stock photos or the prose equivalent, but the job listings have to be specific about what they want, or they'll get the wrong candidates.

During the years we worked on Viaweb I read a lot of job descriptions. A new competitor seemed to emerge out of the woodwork every month or so. The first thing I would do, after checking to see if they had a live online demo, was look at their job listings. After a couple years of this I could tell which companies to worry about and which not to.[2]

That trick still works at companies that are big enough to have a substantial list of jobs, and news stories occasionally break because a job listing shows that a company is planning something it hadn't previously discussed, like Apple hiring someone to build a demand-side platform or Duolingo hiring someone with a background in music education before they'd announced that product. You could also use them to see which e-scooter companies thought they’d won which city contracts. And that's with manual effort, or perhaps some scraping and keyword searches.

The modern version would be to write a scraper, track new posts, extract various trends, and either flag anything interesting for further review or kick it over to a reasoning model and have that model guess what's going on. But once that's a risk—once any competitor who spends a few minutes setting this up can get instantly alerted if a company is considering a new market, vertically integrating, prepping for an IPO, etc., the incentive is to make the job listings more vague. And that's not "more vague," in an absolute sense, but "more and more vague, all the time," until the jobs page basically exists to confirm publicly-available information about roughly what the people at that company do and approximately where in the world their offices are.

This is particularly acute in AI: it's an incredibly competitive space, companies have many degrees of strategic flexibility, many of the founders of big labs despise each other and probably derive equal utility from subtracting a dollar from their competitors' net worths as they do from adding one to their own, and, yes, they are all pretty good at using AI. As the ratio of compute to headcount goes up, the signal in a job listing also rises. It tells you what kinds of intelligence and agency a company wants that aren't captured by AI agents, and where it thinks verification is going to be hard. Increasingly, talent strategy is data strategy. Sometimes, this is blatant: for example, Anthropic is hiring “conceptual reasoning fellows” explicitly to help models improve their reasoning in domains with no checkable answers, like economics, philosophy, cognitive science, etc. Other times, it’s more subtle. But in general, the more AI-centric a company is, the more an accurate list of what they're hiring for basically reproduces whatever strategic plan got presented at the last board meeting. No wonder so many people there are just a Member of the Technical Staff.

Some industries already hire this way, at least for some jobs. Consider pod shops: their central task at this point is to execute a labor market arbitrage, where they lock in the best investing talent on the best terms they can and then harvest the alpha. They're market-timers: Ken Griffin flew ex-Enron traders in to talk about what they did, and he interviewed LTCM employees after their blowup, too.[3] One of the ways for a platform to outperform is to either identify the best talent in an emerging strategy that will be able to absorb more capital soon, or to call the cycles as different strategies go in and out of fashion. If you get the best SPAC trader to sign a two-year noncompete right as the SPAC boom is taking off, it's a captive source of alpha. Which means that actually revealing such a plan is a way of saying "I, a very smart guy, hereby encourage all of my similarly smart friends to start bidding against me for the talent I like the most." So, they tend to go easy on the job listings, to slow the information leakage.

A similar process is happening in the world of investment pitches. A classic way to write a stock-pitch newsletter is to have a few bullet points—some mix of a cheap multiple, this-is-a-real-business validation, a description of a catalyst, maybe some management action that makes the hypothesis look more plausible—and then to have the paywall kick in right before the ticker. So, you can dump those bullet points into your chatbot of choice and, generally, find exactly one company that clearly matches.

A subtler variant of this is what you might call the "Straussian Singularity": it will keep getting harder to drop esoteric hints, or to do actual dog whistles in text.[4] It used to be a status marker to drop an apt literary reference, which fell out of fashion as a form of intra-elite jockeying once that elite didn't consist mostly, but not entirely, of people who would recognize a Horace quote in Latin. Now, that kind of thing feels pretentious and try-hard; you want your Shakespeare allusion to be easily missed, but potentially recognizable, so a reader catching it doesn't denote a foregone conclusion. But suddenly, anyone who recognizes that a reference is being made can just plug the relevant quote into their chatbot of choice and find out exactly what the reference is. In that case, once the reference is recognizable, it's hard to deny; an overall Straussian reading of some essay is going to be fuzzier (Straussians can only be a school of thought for as long as they keep arguing with each other). But we are all leaking more information in our public statements than we used to. For now, it's a question of motivation, but over time, people will start assembling broader AI-driven dragnets to infer useful facts dribbled out by various sources—maybe a tweet gets deleted months after the fact, and that, coupled with an RFP and a tweak to a footnote in a 10-Q will be the confluence of datapoints that makes a sufficiently diligent observer suddenly confident about an imminent merger, product launch, or lawsuit. If there's a positive return from this kind of thing, these privatized sigint processes will probably keep reinvesting. They'll provide baseload demand for inference through the constant process of iteratively figuring out which things are capable of being figured out, and then figuring them out.

  1. Depending on what assumptions you make about how close to light speed it's possible to travel—we could theoretically hit only 3% with current technology, albeit with immense engineering challenges and costs—a message about a planned attack might show up right around when the attack does. This is a pretty far-out sci-fi scenario, which makes sense given that it comes from a science fiction novel. But it's also an annoyingly good argument. Though it would be funny if the real answer were the Spooky Shadow Puppet Hypothesis, which is that intelligent life is out there, but rare, and nobody's traveling anywhere near the speed of light so we'll never encounter each other. However, there's only a brief blip between developing intelligence and having enough of it to produce a Cixin Liu, at which point it freaks out and goes dark—or, possibly, can't come to any species-wide agreement about whether this risk is real or not and annihilates itself fighting over it. ↩︎

  2. This paragraph continues on for a bit, talking about which technologies he did and did not find intimidating. While the essay wasn't designed to do this, ending it that way actually engineered some virality: close with some mild trash-talking of the numerous Java and C++ programmers at the expense of the then far smaller cohort of people who got to write Perl and Python for money, it'll generate lots of defensive, angry comments. But it's also a punchy way to end a talk or essay, and people mostly weren't thinking about virality in this way at the time. If you look at engineered-to-go-viral stuff from this period, it feels clumsy. But it worked, because people hadn't built up an immune system. There were still email chain letters circulating back then! ↩︎

  3. Interestingly enough, his focus was on what they'd done well: staying in business until losses hit 90%. ↩︎

  4. I say "actual" because dog whistle as it's used today is almost a koan: it's supposed to mean a line that supporters will read as more extreme than average people. But it's a pejorative, so people use it for politicians they dislike, which, of course, means that it wasn't just audible to dogs. And at the fringes, sometimes people on the left are more familiar with the extreme right, and vice-versa: in 2020, conservatives were less likely than progressives to know what Qanon was, while more Republicans than Democrats have heard of Antifa. This actually creates a perverse and entertaining dynamic where a politician's incentive is to learn about fringe viewpoints they're supposed to be sympathetic with, and signal that they know about them, in order to make the other side look like paranoid conspiracy theorists. In a win for the conspiratorial types, one way for a politician to learn what hints to drop is to ask their staffers! It used to be that Congressional staffers, Democratic and Republican, were a bit to the left of their constituents (there just weren't that many people who were in the staffer hiring pool and actively disbelieved in evolution, for example). But now, Republican staffers have moved to the right, or at least are pretty familiar with its memes. So there's this bizarre memetic merry-go-round, where a right-of-center politician can lean on a far-right staffer to help them say something that will make center-left people look crazy to moderates. ↩︎

You're on the free list for The Diff. Last week, paying subscribers read about why low market concentration means one of several bad things ($), thoughts on espionage at AI labs ($), and the implications of better mechanistic interpretability ($). Upgrade today for full access.

Diff Jobs

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

  • 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)
  • Series A startup building multi-agent simulations to predict the behavior of hard to sample human populations is looking for researchers and engineers (ML, platform, infrastructure, etc.) to improve simulation fidelity and scale the platform to hundreds of millions of simulation requests. Problem-solving and genuine interest in simulation matter more than pedigree. Experience working with languages with an algebraic type system is a plus. (NYC)
  • Ex-Anduril, Ex-Abnormal Security, Ex-Bridgewater, fast growing startup building frontier agentic cybersecurity to make the long tail of businesses defense dominant via MSPs is looking for a senior machine learning / quantitative researcher. Working in low data environments or working in low signal-to-noise environments e.g, mid-frequency quant finance is a big plus. Evidence of data rigor required. (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)
  • AI Transformation firm with an ambition to build an economic world model to run swathes of the private, unstructured economy is looking for FDEs, Platform Engineers, and business generalists who understand how to solve problems. (NYC, Remote)
  • A Fortune 500 cybersecurity company with decades of proprietary security data is running an internal incubation with a pre-seed startup mentality and a mandate to build something new in AI. They are looking for a founding engineer who can ship fast, an engineer with a security background who’d be excited to contribute to OpenClaw’s security efforts, an AI researcher, and a generalist (ex-banking/consulting/PE background preferred) who wants to wear a bunch of different hats. Comp is FAANG+ and cash heavy. If you want to build something new in AI, but also need runway, this is for you. (SF/Peninsula)

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

Stock-Based Compensation

The employee view of stock-based compensation is that it's a part of your pay that's tied, in a noisy and approximate way, to the success of the entire company you work for, which encourages you, in a similarly approximate way, to do what's good for the firm rather than what's good for you. This is a much bigger incentive when you have points of equity rather than a sub-basis point stake in a massive business, but it still matters. But from the company's perspective, equity comp is a way to finance the business by having employees buy equity through foregone salary. The incentive alignment is a nice side effect. It turns out that you can do this after the fact, with a twist: Mukesh Ambani is asking Reliance Industries' 400k+ employees to invest in the IPO of their Jio Platforms unit, and paid their salaries early so they'd have the liquidity. Part of this is an after-the-fact attempt to give them equity comp, but it's also a way to engineer an IPO pop—and to ensure that the workforce risks being underwater on their purchase if they don't do whatever it takes to keep the stock price high. A positive externality might be that this brings more Indian households into the stock market. Right now, under 10% of the population is invested in the market, though this is changing fast.

Infra

The FT has an overview of the recent selloff in AI infrastructure bets ($, FT). What's interesting is that there are two different dynamics at play:

  1. Supply and demand will never be perfectly matched, and the labs know they've spent a lot more time short of capacity than dealing with an excess. But capital providers aren't as scaling-pilled as the labs, so they view that as a sign that the companies they're investing in will keep raising more money and rolling the dice indefinitely.
  2. Some high-profile problems, like Microsoft abandoning an Nscale datacenter project just ahead of Nscale’s IPO ($, The Information) or Oracle's Project Jupiter, are supply-side problems: to realize their fantastic hypothetical IRRs, the neoclouds actually need four walls, a roof, racks of GPUs, power for all of them, and for every step of this to survive local politics.

So, some datacenter selloffs are bad for everyone in the space. But some of them are a relief to some companies, and a disaster to others.

Balance Sheets

One part of the AI buildout story is that we slowly use up slack in different parts of the supply chain—we run out of new freely-available datasets, so labs have to license them or create the data themselves; fabs run at full capacity, so new fabs get built; datacenters need power, and generating and moving power has bottlenecks and backlogs of its own. A parallel process happens on the balance sheets of different financing providers; VC gets tapped out, growth equity steps in, big tech free cash flow turns negative, we find a layer of the supply chain that can borrow, etc. It would take a lot of foresight to have read the first paper on the transformer architecture and wonder if this would lead to a shortage of electrical transformers some day, and, similarly, it's a bit surprising that one constraint on AI is that it may lead to more insurance claims, and its use might be bottlenecked by insurance companies' willingness to take on big, lumpy risks ($, WSJ). What's especially interesting about this one is that it has more of an effect on frontier capabilities than on the deployment of existing models. The big, nonlinear risks come from very smart agents doing something unpredictable and crazy; a deterministic process with a model in the loop somewhere is less risky. So to the extent that this turns into a drag on AI generally, it's more of a problem for frontier capabilities than everyday use cases. One way this can play out is that the labs themselves will need to be better-capitalized, or they'll need to find more customers who are willing to take on the liability themselves. That makes AI even more capital-intensive, but that's the natural course of things; the bigger an impact it can have, the bigger a liquidity buffer it needs.

An Ownership Society

One of the policy drivers behind how wonky the real estate market got in the 2000s was the Republican party's vulgar Marxist analysis of voting: homeowners tend to vote for Republicans, so if we make it easier to buy a home, we make it easier to become a Republican! Plenty of politician try to distribute money in a way that gets them more votes, but this version turned out to be a pretty inefficient way to turn taxpayer dollars into votes (especially since these policies really ended up borrowing votes from 2008, at an exorbitant interest rate).

But some of that just came down to implementation: encouraging homeownership through borrowing creates some systemic risk, and homes are not an ideal investment vehicle—owning one makes it harder to move to a new city, for example, so it can create a city-level credit cycle. But there are more abstract and mobile kinds of capital, like equities, and they enable a different way to engineer an ownership society. Trump Accounts have provided a mechanism for various wealthy people to airdrop the asset of their choice; this morning, Elon Musk said he'd give every child in America a share of SpaceX, following the Dells, who donated a bit more in cash. It's not clear that the average person will feel that a few hundred dollars compounding away for them makes them part of a different class, but it is still a form of redistribution, and, unlike some of the other ways someone might make first contact with financial markets, Trump Accounts don't encourage rapid trading and risk-seeking behavior. They make the market pretty boring except for the possibility of eventually pulling some money out—but for most people, a shift in that direction is probably a good thing.

The Sonic Boom of Trust

The Diff has mentioned a few times that AI makes many information-wrangling problems more tractable, and that includes going back and catching someone in a lie. But sometimes it means the opposite: Anthropic was able to recover the original seed value for the ECDSA encryption standard, and it's—innocuous! They're just the output of a standard random number generator, with Unix timestamps as the seeds. So, in retrospect, the NSA of the 90s was slightly more trustworthy than we gave them credit for. Since many of the AI-enabled discoveries of this kind are the opposite—e.g. using Pangram to find out that someone didn't write their own work—it's nice to see some symmetry here.