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?