There is a game which is popular in incubation environments. It works particularly well in drawing insights from deep-tech founders. What is the one thing which you know which nobody else knows today?
I played the game a few weeks ago with a potential co-founder and I came up with three personal insights. I’m going to share them here, in three short articles, over the next three weeks.
Artificial General Intelligence is closer than you think
Before you dismiss the AI guy announcing that he believes in AGI read a little further. I did not believe in AGI until a few weeks ago.
Generally speaking, I understand intelligence to be an adaptation to environment. Of course, overfitting a specific environment is also not ‘intelligent’. So my offhand definition includes an aptness to re-adapt to environmental changes.
My understanding of human intelligence does not allow for supernatural influences. Nor do I think that it is so complex that we are 100s of years from understanding it. I think it basically comes down to the wiring diagram, which is specific to the human animal and how it operates in the environment we live in.
Numerous sci-fi stories have been written exploring the ability of the human to live in vastly different environments. I think that we would struggle. And, other stories have looked at putting human-like brain architecture into ape and other bodies. I think that they would struggle.
In a nutshell. I don’t think that our intelligence is magic. But it is very specific, so it might be very hard to abstract and embed in technological systems.
For a long time my belief has been that AGI is technically possible but unlikely in the near future.
I want to revise that position now. I think that we will have a technology which current practitioners will recognise as AGI by the end of this decade.
A ten year horizon is of course a very safe horizon. I’m actually leaning towards the appearance of this technology by mid-2025, which is only four years from today, but I’m a little on the fence on this one.
Here is what I think it will look like
I think AGI is going to be like the CPU when it first appeared. First there were integrated circuits (IC) and the transistor. IC’s are like the breadboards you can still buy in electronics shops today. The revolutionary change which the founding of Intel, out of Fairchild Semiconductor, wrought was the integration of the transistors into the IC’s and the reduction of scale.
In the early years of CPUs people really struggled to program them. But they were general purpose processing units. That is the only thing which is important here. It was possible to move away from specific solutions for problems to general (scalable) architectures which could solve many problems.
I think we are already there in our ability to build a similar AI architecture. You can conceive of it as a chip. I suspect that behind the doors of research labs this is rather done in software for now. Hardware is expensive, software is seen as being cheap. I think this is a mistake in the long run. Software becomes unmanageable and has many hidden costs. We need somebody to put their neck out and build an AI chip. It will take inputs (data streams) and have dedicated learning and prediction circuits. It will be easy to swap in and out specific models – similar to signal interrupts (SIGINT) in a CPU. And the first generations will be a mess. But they will provide a stable architecture for people to build upon, that’s why hardware will win-out over software here.
I guess that anybody reading this is already familiar with Google’s TPU‘s and might even know of some of the attempts by Nvidia. These are fantastic, but do not really address the general part of AGI. I’m more interested in what a future version of ARM might produce. The core of general here is in how routing is handled between the different specialised modules.
I have two interests in this space. I would love to write that core algorithm / architecture. That would be something like the human thalamus plus the neocortical wiring diagram equivalent. And, I can’t wait to see how we begin to plug-and-play these chips.
Let’s revise my prediction. I think we will have version 1.0 within four years from now, and ubiquitous usage by the end of the decade. Now that is something which you can act upon!