Text and Telos

In this issue:

• Text and Telos—Strictly speaking, LLMs don't have a model of physical reality, just a bunch of heuristics for predicting a sequence of tokens. But in practice, their training data is an ongoing effort to create a comprehensive model of reality, with a focus on the parts people care about the most.
• Scaling—Not everyone is spending aggressively on AI capex. But there are enough participants for the race to still be on.
• Retail Investors—The promise and peril of selling retail investors a risk premium slurry.
• Comparative Advantage—The weak connection between aid and economic growth.
• Transaction Costs and Corporate Structure—If your logo's on their business card, there's a sense in which it pays to be on the hook when they get sued.
• DeepSeek Governance—If a business risks what regulators view as an open-ended liability, the safest place for it to exist is within a structure where there are valuable assets to seize.

Text and Telos

If you'd told someone twenty years ago that, by looking at statistical patterns in how people use words, you could build a tool that could accurately correct typos, they would have been impressed. It's always hard to know in the present how an informed person without our present knowledge might have reacted to new insights, but in this case we know what they'd say, which is roughly "You're hired." But if you'd taken that further and said that statistical models based on human-generated text could write an entire sentence, or an email, or a research report, or that they could produce working code (and fix bugs in non-working code), or write the lyrics to a parody of "Devil Went Down to Georgia" about structuring derivatives, you would, at some point, have insisted that this was taking things a bit far. It's very difficult to have intuitions about how far one can push a simplistic model, given how messy the real world can be.

Then again, some abstractions are unreasonably effective at predicting reality. Richard Feynman once noted a striking comparison: quantum electrodynamics produced a theoretical estimate of one variable (the magnetic moment of an electron) that differed from the measured value by a proportion that was equal to making a theoretical estimate of the distance between New York and Los Angeles and being off by the width of a human hair.

It's cheating to compare prose to math, because unless you have very specific theological or philosophical views there's no reason to think that physical reality has some kind of direct correspondence with text. On the other hand, a literate culture is a continuing effort to codify as much of the world as possible in text form, at every possible level, from religious texts and novels that try to codify enduring truths to internal corporate wikis that explain the process for expensing in-flight WiFi.

From the perspective of an LLM, all of this was very helpful, in the same way that to a human being it's incredibly convenient that cows convert basically useless grass into food, beverages, and raw materials for clothes and other implements. Someone's gone to the trouble of processing annoying inputs into useful ones—great!

This is the result of multiple historical accidents. Text gets privileged when communications are slow enough, and sufficiently low-bandwidth, that it's inconvenient to transmit other media, or to send people directly to make decisions. The privately-held but state-influenced trading companies of the early modern period, like the various East India Companies, had to be somewhat decentralized given that a letter from Chennai to London would receive an answer in about a year, if everything went exactly to plan. So questions like "the pepper harvest isn't great this year, what do you suggest we do?" were completely pointless. What did make sense was codifying expectations and practices in both directions—what each endpoint wanted was for the other one to be able to make reasonable decisions without real-time information. The information did get transmitted eventually, but it would make more sense to package together a few different things in the same single communication: here's what happened, here's what I did about it, here's why I did it, and here are the results so far and what I expect to happen next. In other words, limits on bandwidth encouraged people to report facts and then explicitly walk through the process by which those led to decisions.

What happens when bandwidth and latency improve? There's a lot more room for real-time decisions, so it isn't necessary to outline everything in long memoranda. But the memo is still a default form of communication. Even if what's happening in practice is that someone in the Chennai office is Slacking someone at the London HQ, or someone at a London satellite office is hopping on a call with their boss in Chennai, whatever decision gets made will probably be memorialized in some document that, in form if not content and language, wouldn't have been out of place in the cargo hold of a merchant frigate.

Digitization actually re-ran this same text-heavy, context-heavy communications history. In the early days, text was the only thing it was feasible to transmit, which had a two-sided feedback loop (cheaper bandwidth subsidized digital cameras and digital video, and they provided the demand that justified even more bandwidth). Companies that went online early developed an email-heavy communications culture, and that culture persists even though you could, in principle, manage your business by distributing voice messages, short-form video, etc. Once the default mode of communication is established, it becomes a default way of thinking, too, and becomes the place where you get lots of source material. You can try to reproduce the thought process that led to memos and emails, but if you don't have a record of other ways decisions were made, you don't have a good way to build on them.

All of this was just one of those historical contingencies until LLMs followed the typical historical pattern in which general-purpose technologies radically revalue raw materials, and often in a way that reinforces winners. (Land in Manhattan, London, Chicago, etc. was revalued higher when those places industrialized, but revalued much higher than that when they deindustrialized.) We've been accidentally manufacturing a huge corpus of data that aims to solve the exact limitation that some prominent ($, WSJ) AI researchers have: LLMs don't have a model of the real world, just a model predicting text. But a huge fraction of the text that humanity produces is some attempt to describe and model the real world. Not only that, but we're modeling the real world in the terms that people actually care about—we produce a lot of text, fiction and nonfiction, about interpersonal relationships, and comparatively less about how squirrels feel about one another. And we place a much higher value on understanding one another's inner lives than on understanding the inner lives of random animals. There's a lot more written material about mining gold than mining pyrite, too.

None of these textual models are perfect, but to the extent that they approach perfection, it's in proportion to 1) how important the topic and hand is, and 2) how much progress you can make on that topic by putting one word after another. And the world has gotten more written-word-influenced over time (with periodic backsliding whenever a non-written medium like radio or TV shows up). It's still the case that if you want to have influence fifty years from now, you're better-off writing an essay or a book than making a video or recording a podcast, and that if you do try to influence the world through writing, you're building on a far bigger and better tradition than in other media.

This will be extended further, in an odd way: humanoid robots have been a sci-fi dream for a long time, probably because it's so easy to slot them into human roles in fiction (that's especially true in a visual medium; a humanoid robot is a costume rather than a prop). That also means that they're suitable by default for humanoid tasks, and that in a human-centric environment, their sensors will approximate most human senses, in human-relevant ways—they'll be "looking" at things that are eye level, "listening" in the same frequencies people do and focusing on the same sound patterns people care about.^[1] This makes them great data-gatherers for the set of things that people either don't or can't write down—they are to multi-modal models what a text corpus is to an LLM. They won't know everything, of course, and there will be more to figure out. But between their inputs and a few millennia of text, there's enough data out there for AI to, in principle, answer most of what we need to know.

Diff Jobs

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

• An OpenAI-backed startup that’s applying advanced reasoning techniques to reinvent investment analysis from first principles and build the IDE for financial research is looking for a backend-focused software engineer with experience building scalable back-end systems. (NYC)
• An AI startup building tools to help automate compliance for companies in highly regulated industries is looking for an AI research engineer with 4+ years of software engineering experience and a knack for translating cutting edge AI research into production level systems using experimentation. (NYC)
• Well-funded, fast-moving team is looking for a strong-willed full-stack engineer to build the best video editor for marketers. Tackle advanced media pipelines, LLM applications, and more. TypeScript/React expertise required. (Austin, Remote)
• A Google Ventures-backed startup founded by SpaceX engineers that’s building data infrastructure and tooling for hardware companies is looking for full-stack and front-end focused software engineers with 3+ years experience, ideally with data intensive products. (LA, Hybrid)
• A company building the new pension of the 21st century and enabling universal basic capital is looking for a mobile-focused engineer who has experience building wonderful iOS experiences. (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.

Elsewhere

Scaling

xAI plans to open a second datacenter in Memphis ($, The Information). This is a good reminder that a race is still a race as long as there are at least two participants who are worried that someone else will win. It's not as clean a thesis as it was when it looked like every big tech company was racing to spend as quickly as possible, and even in an optimistic scenario there will be cases where spending overshoots. But companies also don't want to be left behind, and if there's variance in spending appetite there are two good explanations:

1. The biggest spenders are overestimating demand for future models, or
2. The smaller spenders are either conceding that they'll make more on distribution than on model ownership, or just haven't figured out some of the ROI-maximizing tricks other parts of the industry have.

Retail Investors

Bridgewater's all-weather strategy now has an ETF. The intuition behind an all-weather strategy is sound: there are various asset classes that are more or less levered to different macroeconomic variables—if you're expecting inflationary growth, you want to own commodities; if you're expecting a deflationary recession, you want treasury bonds and safe-haven currencies. Predicting which regime we'll be in is hard, and even if someone has skill in doing so, the sample size of market regime shifts over the course of an entire career is too small to achieve statistical confidence in their ability to do so. Such a product only really has two problems:

1. Correlations are a moving target, and the more investors there are who own an all-weather mix specifically to harvest uncorrelated return streams the more correlated those return streams will be, and
2. A product like this is hard to market, because it's never the perfect choice for an investor with any specific macroeconomic viewpoint or worry. Every subset of the portfolio has a cleaner sales pitch under the right circumstance, but the mix of all of them only appeals to someone who is radically uncertain about the future, but also interested in somehow turning that uncertainty into a financial bet. This ETF partly solves that problem with the Bridgewater branding, and there's no shame in monetizing a book with management fees.

Comparative Advantage

The Economist has a downbeat review of the impact of aid on economic growth ($), arguing that there's little evidence that it makes countries richer (though it can blunt the human toll of poverty). A dark way to read this is that globalization encourages every country to focus on whatever competitive advantage allows it to bring in money from overseas; China focused on cheap manufactured goods, Japan on expensive ones, the US on reserve currency status, etc. And for some countries, their highest-return export is human misery, and their incentive is to maintain that competitive advantage.

Transaction Costs and Corporate Structure

Companies exist at the efficient frontier between the coordination benefits of top-down control and the information benefits of market competition. One of the best illustrations of this is the tradeoffs that show up when a company tries to internally reorganize. That's something KPMG is going through right now as it attempts to consolidate country-level subsidiaries into larger units ($, FT). The argument for a country-level unit is straightforward: different places have different accounting standards and legal rules, so, in effect, each KPMG subsidiary is providing a slightly different set of services. They're running into two problems: first, as they expand into consulting, they need to invest upfront in order to earn more later, and that's harder to swing for some of their smaller subsidiaries. The more subtle problem is that the same legal encapsulation that justifies a decentralized structure means that the smaller units have a weaker economic incentive to preserve the brand name and a stronger one to bend the rules. So it's better for them to have more people on the hook for the direct financial cost of this, in order to align incentives with the entire organization that pays the cost of damaging the intangible value of their brand.

DeepSeek Governance

There's a related version of this happening in China's AI sector: DeepSeek has, of course, gotten interest from investors, but so far it's turned them down ($, WSJ). DeepSeek was launched as a side project for a quantitative hedge fund, and in a purely capitalistic way it makes sense for them to be a separate entity with a separate ownership base. There aren't many direct synergies between LLMs and trading, and to the extent that there are, these can be captured by a more conventional customer-supplier relationship. But in a regulatory environment where there's potential liability for an LLM that spits out the wrong answers about certain historical dates and various leaders, both living and dead, a standalone entity might have an incentive to ship models before they've been properly vetted and preemptively censored. But as long as this company is part of a larger, more profitable one, in a sector that the state is quite capable of monitoring and influencing, the AI business can essentially offer the quant business as an economic hostage that shows the state they expect to behave.