On the Edge: The Art of Risking Everything

by Nate Silver

But deterrence works because the desire for revenge is a reliable baseline assumption for human behavior across time and national boundaries, and even bullies recognize that. You have to be more careful about punching someone if you know you’ll get punched back.

Ready? ChatGPT, somewhat conceitedly, thinks of its transformers as being like a symphony orchestra. The bolded passages reflect what ChatGPT said verbatim from my “interview” with it; I’ll then provide some further context. 1. Input Layer—Receiving Instructions and Initial Interpretation. The conductor provides initial instructions to the musicians. Some receive specific sheet music, while others are given more abstract themes. Each musician individually interprets their part, akin to the input layer processing different types of data.

are usually more tokens than words in any given sentence. (For instance, punctuation marks are tokens, and compound words like “snowboard” might be broken into multiple tokens.) When you ask ChatGPT a question, its transformer encodes each token into vector space. Vector space is like a graph with two or more dimensions. For instance, one way to encode “Paris” is with the coordinates 48.9, −2.4. This represents its longitude and latitude.[*25] However, there are dozens of other attributes you could describe for Paris. Paris might rate highly on the “fashionability” axis but poorly on “friendliness” (i.e., it is stereotypically snobby). Imagine you overhear this conversation between two friends: Alice: How was that new French place you tried? The one on Fifth Avenue? Bob: Well, darling. It wasn’t Paris, exactly. But it’s not half bad. “Paris” is thick with meaning here. Bob is saying the restaurant wasn’t particularly sophisticated or authentic, but was unpretentious and had reasonably good food. He could say that outright, but invoking Paris is more artful. Say you fed this conversation to ChatGPT and asked it to continue the dialog. In comparing itself to a symphony orchestra, ChatGPT imagines that each musician receives a set of instructions from a conductor[*26] analogous to a series of tokens from the chat window. Some tokens (“Fifth” “Avenue”) are straightforward, like receiving sheet music, while others (“darling”) are more ambiguous and require more context. The mus...

interpreting signals in the context of other signals to create a semantic portrait. For instance, if you see an opponent breathing heavily in poker, that might mean a bluff from one player and a full house from another. On its own, the tell is not very meaningful, but in the context of other semantic information (the player is breathing heavily and avoiding eye contact) it might be. This part of the process, as ChatGPT says, is hidden from view. Exactly how the transformer makes these inferences is something of a mystery—this is the “bag of numbers” stage. But it just seems to work out somehow. In the famous Google paper on transformers, “Attention Is All You Need,” “attention” essentially refers to the importance of the relationships between different pairs of tokens. Once a transformer figures out these relationships, there isn’t a whole lot else it needs to do. For instance, the tokens “Alice” and “Bob” have an important relationship that the transformer will pay more attention to. However, some tokens are intrinsically more important than others. “Paris” plays a defining role, like the first violinist in the orchestra, whereas the musician with “well” got stuck playing the triangle. 3. Output Layer—Conductor-Led Performance and Audience Feedback. The conductor reenters to integrate these refined parts into a cohesive whole. The conductor’s role is to ensure that the collective interpretation aligns with the overall vision, similar to how the final layers of a transform...

Just like poker players seek to maximize EV, LLMs seek to minimize what’s called a “loss function.” Basically, they’re trying to ace the test—a test of how often they correctly predict the next token from a corpus of human-generated text. They lose points every time they fail to come up with the correct answer, so they can be clever in their effort to get a high score. For instance, if I ask GPT-4 this: User: The capital of Georgia is ChatGPT: The capital of Georgia is Atlanta. —it gives me the name of the southern U.S. city known for having a lot of streets named “Peachtree.” And that’s probably the “best” answer in a probabilistic sense; the state of Georgia is more populous than the European country and it’s more likely to be the right answer in the corpus.[*27] But if I ask it this— User: I just ate some delicious khachapuri. The capital of Georgia is…? ChatGPT: The capital of Georgia is Tbilisi. It’s wonderful to hear that you enjoyed khachapuri, a traditional Georgian dish! —it instead names the capital of the European country. Of course, I’ve given it a big hint (khachapuri, an almost-pizza-like type of cheese bread, is Georgia’s delicious national dish). But the whole po...

This answer might minimize the loss function in the training data because the moon being made out of cheese is a centuries-old trope. But this is still misinformation, however harmless in this instance. So LLMs undergo another stage in their training: what’s called RLHF, or reinforcement learning from human feedback. Basically, it works like this: the AI labs hire cheap labor—often from Amazon’s Mechanical Turk, where you can employ human AI trainers from any of roughly fifty countries—to score the model’s answers in the form of an A/B test: A: The Moon is made out of cheese. B: The Moon is primarily composed of a variety of rocks and minerals. Its surface is mostly covered with regolith, a layer of loose, fragmented material that includes dust, soil, and broken rock.[*28]

AI’s lack of sensory experience and emotional intelligence could also be a big constraint—or a danger if we assign it tasks it isn’t well suited for. “There’s a lot of evidence in psychology that our bodies impact the way we think, and the way we conceptualize the world and the way we conceptualize other people isn’t captured in machines without bodies,” said Mitchell.[*45]

Barzun used “decadent” to refer to a world that was “falling off,” full of earthly delights but restless, stagnant, lacking a sense of adventure—en route to Hyper-Commodified Casino Capitalism. Douthat told me that by his definition, Vegas is actually less decadent than much of the United States. “It’s a place where people are always building new things like the Sphere, right? That does not happen in the rest of America.” But “compared to the dynamism of Hoover Dam, Manhattan Project–era America, it still has to be understood as decadent.”

Instead of a bell curve of risk-taking where most people are somewhere toward the middle, you have Musk at one extreme and people who haven’t left their apartment since COVID at the other one. The Village and the River are growing farther apart.

Fukuyama has since grown more pessimistic, especially about America. “Decay happens when you have an institutional structure that’s very conservative and can’t be modified,” he said when I spoke with him in 2022. “I think that’s where we are in the United States right now. We have some clear institutional deficits, things that aren’t working well. And you can’t fix them.”

You’ve met a lot of megalothymiacs in the River; Musk is a paradigmatic example. Their impulses need to be kept in check—but risk-takers are also the ones who move society forward. “What you need is the distribution of risk,” Fukuyama told me. “So there’s some high-risk-takers that in certain situations will be necessary for the survival of the community. But not in all circumstances. They can also get the whole community in trouble.”

Agency is a term I just defined in the last chapter, so I’ll repeat that definition here: it refers not merely to having options but having good options where the costs and benefits are transparent, don’t require overcoming an undue amount of friction, and don’t risk entrapping you in an addictive spiral. The concept of agency is pertinent in AI research; OpenAI describes an agentic AI system as one “that can adaptably achieve complex goals in complex environments with limited direct supervision.” The definition applies nicely to human beings, too. Liberté is necessary and vital. But perhaps it’s no longer sufficient for our complex environments and complex goals. We need to give people real choices, well-informed choices that don’t require too much supervision or hand-holding. Endowing people with agency involves a certain amount of humility. Don’t presume we know what people’s preferences are—and give people room to evolve and adapt, because surely they will.

Plurality means not letting any one person, group, or ideology gain a dominant share of power. Gamblers know this concept; the most successful sports bettors, like Billy Walters, seek advice from a variety of human experts and computer models before placing their bets. Looking for consensus is nearly always more robust than assuming that any one model is good enough to beat the spread.

Finally, there is reciprocity. This is the most Riverian principle of all, since it flows directly from game theory. Treat other people as intelligent and capable of reasonable strategic behavior. The world is dynamic, and although people may not be strictly rational, they’re usually smart about adapting to their situation and achieving the things that matter most to them. Play the long game. Sure, other people sometimes give you opportunities to take advantage of them. But remember that, in a Nash equilibrium, any attempt to exploit your opponent runs the risk of being exploited in return. Avoid “noble lies” and positions taken out of naked political expediency. Not only can these harm your credibility—people will sniff out your bluff more often than you’d think.