Subscribe
Share
Share
Embed
The Diff is a newsletter exploring the technologies, companies, and trends that are making the future high-variance. Posts range from in-depth company profiles, applied financial theory, strategy breakdowns and macroeconomics.
## "Do Firms Need Juniors?" Might be the Wrong Question
A fun question that comes up in light of AI automating pointless busywork, albeit with some supervision is: what will junior team members do in the numerous industries where their job is to do exactly that? Who will hire new lawyers if discovery and legal research get automated? What will junior investment bankers do if "turn this 10-K into a financial model" and "use this model to create a PowerPoint presentation suggesting that the company in question do a secondary" are both commands that some prepackaged tool can do for a lot less than junior bankers make? Will marketing departments really keep their newest employees around purely so they have someone to blame when the company Twitter account puts its foot in its mouth?
These are legitimate concerns, though one thing they illustrate is the surprising durability of juniors in the face of technological change: word processing software, spreadsheets, and compilers have all eliminated tasks that used to function as training for people who were new to the business they were in. In each case, the assignment started on the basis of high-level context, both in the sense of what the desired outcome was and all the tacit wisdom that comes from experience, including the experience of doing the grunt work. The assignments juniors get are essentially informed bets on the payoff from applying undifferentiated hours to some task. The senior person's job is to quickly assess whether they're asking a trick question or not: if the task is "figure out why this specific part of the program is running slowly in some cases," the expected answer probably has something to do with repeatedly looping through the same process redundantly, and not figuring out some esoteric bit-flipping hack that will yield a surprising increase in speed; a new lawyer who's tasked with looking at some contract is probably not supposed to find a clever backdoor, just to check it for the obvious stuff.
But this points to something weirdly bifurcated about LLMs: they're good at the easy stuff, and flake out or at least need a fair amount of supervision when they're doing something complicated—just like the a pretty-good intern, or someone with a few months of experience and a burning desire to keep making a good impression. On the other hand, they have sprawling context and can make connections across different fields, which is exactly what the experienced people are there to do. An LLM isn't an expert on most fields, but it is pretty good at giving you a checklist for the things you might be doing something wrong, which is, at some level of fidelity, reproducing the mental checklist that people develop by doing said thing wrong in the past.
So, depending on the career track, one thing LLMs substitute is the senior people. They're the ones who are supposed to be zooming out to ask big-picture questions, thinking about analogous situations with counterintuitive answers, knowing who to contact for additional advice, etc. And these are all the kinds of questions that, with a bit of diligence and humility, you can get from an LLM. It's actually a kind of fun feedback loop: start skimming some document that relies on more domain-specific knowledge than you happen to have, and every time you're confused, just drop a screencap of the relevant text into the LLM and start asking questions.
What you shouldn't expect is for this to be a direct substitute for having the right expert on hand. But, as with many LLM outputs, the relevant counterfactual isn't a better version crafted by a real person but nothing at all—one of the economic drivers of these kinds of apprenticeship-style job paths is that someone who's smart but inexperienced is going to end up wasting gobs of valuable time. Every year between when your working memory and physical endurance peak and when you get enough context to use it correctly is, in some sense, a waste.
Where does this work best? It's ideal for jobs where there's plenty of publicly-available documentation about the job, both in terms of high-level strategy and low-level implementation. It helps if the business has few barriers to entry, or, at least, if those barriers aren't especially identity-based (even if hypothetically spending three straight years interrogating ever-smarter models about the nuances of law is a better use of your time than law school, it isn't a JD). When the barriers to entry are very low, like "can you register a domain name and form an LLC?" or "can you open a brokerage account?" this dynamic will show up more.
And it won't work at all in cases where there's tacit knowledge that's hard to encode in text form. You can probably read a lot about being a good therapist from LLMs, but they aren't going to be great practice if you're face-to-face with someone whose life is collapsing around them. In jobs that require heavy equipment, it's an interesting piece of trivia that someone might learn how to operate such equipment all on their own, but an autodidact still needs to borrow the social capital of someone who learned things the old-fashioned way.
So what this will really reveal is a bit more about which jobs are valuable, and how. Sometimes, the point of a job is less the skills and more that it's a way to be the human embodiment of some impersonal institution, whether that institution is Goldman Sachs, Salesforce, or the State Department. For jobs like that, the experience of learning on the job is part of the economic asset that's created by that learning.
But this also means that there will be an interesting shift in skilled labor: some of the intellectual components are getting commoditized, though there will still be an important role for human thinking and human wisdom for a long time. But if those parts get commoditized, the intangibles will be relatively more important, so the economy will have higher rewards for tall, charismatic, good-looking people who are now less constrained by the need to know what they're doing. So this may be a rare case of a technology actually being relatively beneficial for old people, rather than young people. We're still in an interregnum where knowledge and judgment are getting commoditized, but haven't yet been fully commoditized, so right now there's more upside in using LLMs to avoid dumb repetitive work and make yourself smarter, rather than as a pure substitute for brains. But in the equilibrium that eventually establishes, there will be lots of intelligence on tap, and for jobs that don't push people to their intellectual limits, firm handshakes and funny icebreakers will be relatively more important and smarts and knowledge, sadly, a bit less.
## Elsewhere
### Solve for X + X
Elon Musk has merged xAI and X (formerly Twitter), into a single entity. A few points here:
1. The old Musk model was that he had two main entities that were definitely worth something, Tesla and SpaceX, and then a bunch of side projects (Neuralink, Boring Company, electing Trump, impregnating enough people to get an electoral vote's worth of descendants, etc.). You could bet on the side projects, but that bet was implicitly backed by the actual viability of one of the main focuses. Now, he arguably has three: the combined Xs are worth about $100 billion.
2. Some AI companies are basically pure-play model businesses, like Anthropic and Mistral. They may have a consumer-facing application, but investors are mostly betting on the technical side of things. This combined entity joins the equally-crowded ranks of companies that have both an AI model and an existing consumer app they can use to distribute it. Since app users provide continuous feedback, and also implicitly set a standard for how much AI users of that app should be using, this is strategically valuable even if all the inference is very expensive.
3. On the other hand, there's a very real sense in which this deal doesn't mean much, other than as a bit of financial performance art: at last, companies that lent to Twitter have exposure to an AI business that will consume vast amounts of capital in the hopes of an upside scenario in which they won't participate! Finally, xAIs' equity investors can diversify out of a pure bet on artificial intelligence and get exposure to an AI business that carries some of the weirdest ads in existence! Mostly, the deal is a formality: if you back an Elon company, you're buying tracking stock, but you're really betting on Elon.
### Refactoring
DOGE plans to rewrite code that runs Social Security in a few months. COBOL has had some interesting economic aftershocks: it was an early default, so many of the things that were most obvious to automate (payments, plane tickets, payroll) were initially automated in COBOL. Once these core systems worked, it became a very bad risk/reward to interfere with them in any way, and, as a result, many of them were written by people who have long since retired or died. Rewriting legacy code turns out to follow a weirdly nonlinear timescale: if you do it, you're either years too early or decades too late.
### AI and Distribution
For pure-play AI companies, the usual angle is: build a model, build a chatbot to demonstrate it, and sell API calls to developers who build specific implementations. But the biggest tech companies have a larger surface area: they have existing user relationships, and collect data that their competitors dont. So, Apple is planning to offer a healthcare-focused AI agent that uses the data it collects from Apple's various devices. The default data companies collect is actively user-generated by way of apps and browsers, but in Apple's case they've been building up a corpus of passively-shared data like heart rate and sleep time. So, among other things, this is a good validation of the tech instinct to hoard as much data as possible, knowing that there will either be better ways to analyze it in the future or at least more helpful ways to present it.
### How Models Think
Anthropic has released a fun paper looking at which parts of a model are active when solving particular problems. Lots of interesting detail here: the models have some language-agnostic concepts, for example, and aren't just thinking in whatever language a question's asked in; they aren't just focusing on one token at a time (with rhyming poetry as the demonstration case); and they have surprisingly human-like heuristics for "mental" math. One of the most interesting discoveries is that the model apparently has a default of not answering questions, and only provides an answer when there's something that supersedes that. This is further evidence that we've implicitly given LLMs a world model, but mostly for the parts of the world we really have questions about: if LLMs usually answer your questions, but their default behavior is not to do that, then the questions must be overwhelmingly likely to reference the same kinds of information LLMs are already trained on. They don't know everything, but they know a surprisingly large fraction of what we want to ask them.
### The AI Business Model
One valid critique of AI-generated images is that they're trained on the work of human artists, but don't pay those artists. Similarly, a model that can answer "in the voice of" some living person usually isn't giving that person a royalty. But for chunkier collections of assets, that's not the case: there are AI-generated trailers for real movies on YouTube, and some studios are just asking for a cut of ad revenue rather than asking for a takedown. This is narrowly-scoped, and the studios already have relationships with YouTube because it's such a great venue for getting legitimate trailers in front of potential filmgoers. And the trailers are ads, not content, so it's not a serious business risk if they're being made by third parties. But this offers a general sketch for where long-term AI business models might end up: for creators, it can end up being just one more medium and one more way to get paid.