Balaji Srinivasan on AI & the future
Balaji expects many AIs rather than one god-like AGI: models from different cultures, increasingly distilled and decentralized, bounded by chaos and cryptography and built to wait on a human prompt. AI works middle to middle: humans sense the world and prompt, AI acts, and humans must verify, so generation gets cheap while verification gets expensive and AI does best where output is easy to check (visual, physical). For work, that makes every person a CEO directing AI, and turns what is human into the premium good.
15 points · 3 appearances · 2025–2026
How the view evolved
Polytheistic AGI, not one vengeful god → AI's future is distillation and decentralization
In 2025 he hoped for at least American and Chinese AI, plus decentralized open-source AI 'if we're lucky'; by 2026 he expects distillation and decentralization to be 'a very large percentage of the future', since distillation attacks are hard to stop technically or morally.
AI fails at markets and politics: adversarial, time-varying → Humans are the sensor, AI is the actuator
In 2025 he argued AI fails at adversarial, time-varying domains, so the human who senses the market prompts the AI; in 2026 he turns this into a general model: humans are the sensor, AI is the actuator.
AI works middle to middle, not end to end → AI cuts generation cost but raises verification cost
In 2025 verification is the neglected half of 'middle to middle' AI; in 2026 it becomes an economic law (AI cuts the cost of generation but raises the cost of verification), and he backs it with his own practice of in-person, proctored exams.
AI is good for the visual, worse for the verbal → Physical tasks are easier to verify than digital ones
In 2025 AI is better for the visual because visual output is cheap to check; in 2026 he extends the same verifiability logic to the physical world, where tasks have clear end points, so AI works for 'visuals, the verifiable, and the physical'.
Up next
2021–2025 · changed view
“Really these are two halves of the same coin because what crypto really does is its property and property rights and speech rights and encryption and civil liberties outside of the control of abusive states and if your state by the way is mostly convincing rather than coercing then people don't really need to use crypto quite as much I mean you can use it for lots of other things you'd use it for getting around centralized companies not just states because you you know the power of Facebook or a Twitter is also something you can decentralize with crypto which is very important so even if you don't buy the state argument all the corporate argument is important but the more abusive a state is like a Venezuela the more important crypto is it's kind of that last bulwark the check and balance is actually outside the state.”
2019–2025 · changed view
“Are new countries even possible? We know that we can found new companies like Google out of a garage, and new communities — the three billion person community that is Facebook started out of a dorm room. We know we can start new currencies like Bitcoin from a white paper. But can we start new countries similarly? That's the next step. That's what taking the internet seriously means.”
2021–2026 · changed view
“By analogy we have different categories for cash currency stocks bonds mortgages art et cetera all these scarce assets today the fundamental innovation of the blockchain was the ability to represent scarcity digitally because before blockchains you could you could transfer or rather you could you could copy rather than transfer something right I can copy a file to your hard drive what I couldn't do is transfer it in such a way that you knew I no longer had possession of it.”
2021–2025
“Here's a mechanistic explanation for why that happens, okay? So visualize NYT and centralized media as a hub, okay? And then here's a spoke coming off, and this is, let's say, Bitcoin, and then it's another one that's like India, and here's another one that's like Japan or Iraq or whatever, right? These various spokes. And if you are an expert on Bitcoin, you have better local information on that topic, in, like, the Hayekian sense, but not simply about transactions, just other things, events that are happening. And so you have a better source of information than the centralized hub, and so you know that what they're saying is, is false about this topic or it's misleading. But then when you go and read about, uh, you know, India or Japan or Iraq or Hollywood or what have you, you do not have, you're not in that spoke, right? So you don't have local information there, you're relying on the hub, which is like a, you know, a hall of mirrors that's giving you a noisy reflection of what's happening in Japan or what's happening in India, okay? It's reflecting back to you something that's false, but you don't have a direct peer-to-peer connection with India, with Japan, with Hollywood, to get better information.”