wagmit

Network School is back. We’re up and running in the beautiful country of Kazakhstan, which combines high tech, low cost, brilliant people, a competent state, and plenty of space to build. So: if you want to build a techno-optimist society with us, apply to Network School.

What comes next for AI? Perhaps physically specific intelligence, more than artificial general intelligence. Here’s a draft thesis. (1) First, AI is now entering the trough of the hype cycle. Every tech, no matter how amazing, goes through this. It’s actually one of the best times to invest and build once the space begins to thin. (2) Second, AI is both highly useful for search, summarization, visualization, and prototyping *and* a highly annoying source of spam, scams, and slop. Too many will focus solely on the downsides. (3) Third, the exact plateau of AI is quite unexpected. The fact that current models seem to have topped out at “far better web search and summary” is not where most would have pegged it in 2022. Disrupting Google is a historical accomplishment, yet it’s not the machine god. (4) One thesis for why models plateaued in this way is that LLM-style digital AI really *is* just repeating, rather than truly thinking. It is downstream, not upstream. Now, you can get much further with this than most thought. But you can’t get all the way to novel thoughts. (5) That leads to another point: as I’ve said before, digital AI does it middle-to-middle, not end-to-end. Because the bottleneck for digital AI is prompting and verifying. (6) However, *physical* AI — in the sense of robotic guidance — *can* feasibly do things end-to-end, albeit after the investment of a lot of energy on edge cases. For example, self-driving cars now truly can take you from point A to point B. They do it end-to-end. (7) So, extremely well-specified and economically valuable physical world problems like “drive from point A to point B” are where we should see significant AI progress. Call this physical specific intelligence. (8) Chinese robots in particular will nail task after task, in many form factors beyond just cars or humanoids. Sidewalk robots and delivery drones are already live in China, alongside self-driving cars, and their hardware sector is accelerating. (9) One reason to be more short-term bullish on physically specific intelligence rather than artificial general intelligence is that the physical world is real while the digital world can be fictional. @drfeifei made the related point that the textbook concept of simultaneous location and mapping (SLAM) is actually an explicit physical world model, something we’ve all known is lacking in the digital realm. (10) In particular, if you have N different robots all doing SLAM (or equivalent) and uploading their sensor data to a central database, they are all in a sense learning from the same physical world and each other. The millions of training miles driven by self-driving cars show that this strategy works. (11) Contrast this to the digital world. Text can be fake, it can be fictional, and indeed it is often so. And this is doubly so as AI text permeates the web. (12) So N robots sensing the physical world will converge on a consensus reality. A Unitree robot and a Tesla robot will perceive the same objects. By contrast, N agents in the digital world just keep ingesting text and images, which can and will be inconsistent. (13) In the absence of digital signatures, the online data used for AI training will be not just internally inconsistent but fake in subtle or deep ways. Indeed, widespread AI models ensure that much digital data (outside hard digital borders) will become fake, or slop, or AI agents talking to each other. (14) Algorithmic breakthroughs could obviate everything I’ve just said. But at least right now, the field seems wide open for physically specific intelligence. (15) The industrial version of physically specific intelligence is factory robots. The consumer version though is a garden of intelligent things: robot cars, dogs, drones, fridges, and the like that can talk to you and understand you. Amazon Alexa, but for everything?

This is the livestream for Network State Conference 2025. xitter.cf/i/broadcasts/1OyKAjkbd…

This year's Network State Conference is on Friday, October 3 in Singapore. Our speakers include Vitalik Buterin, Bryan Johnson, Ben Horowitz, Brian Armstrong, Amjad Masad, Ranveer Allahbadia, Arthur Hayes, Nas Daily, Noah Smith, Andrew Huberman, and the governments of Singapore, Dubai, Abu Dhabi, and El Salvador. That means many of the most successful founders, creators, investors, and policymakers in the world are now interested in the idea of bringing online communities into the physical world. Our conference also features writers, community organizers, and intellectuals from around the world (and across the ideological spectrum!) who are aligning around the concept of fusing offline states with online networks, with ideas ranging from pro-tech regulations to special innovation zones to startup cities. All this capital and talent going towards the area means that you too can join a startup society that suits you — or found one yourself — even if you're starting out with nothing besides an internet connection. And by building new societies with defined social smart contracts where everyone has explicitly consented to governance — and anyone can opt out — we can defend capitalism, democracy, internationalism, self-determination, and liberal values in an increasingly illiberal world. That's the goal, at least. We can only give broad sketches in a tweet. But if you're interested in the long-form version, register to attend in person or remotely at ns.com/conf2025. Because we’ve started internet companies, we’ve started internet currencies, and now we’re starting internet communities. Come to the Network State Conference to be there at the start.