| Articles & Essays | Lab.Project Int.Aug | Project Int.Aug

What is Intelligence Amplification?

About this Project

This website chronicles some of my work on attempting to build a tool to amplify intelligence.

If that sounds grandiose, I think you'll be convinced on its feasibility after reading some of the essays listed above. I also document my technical progress and observations towards this goal here.

Intelligence amplification is an attempt to raise the current cognitive abilities of cyborgs. But don't you need neural implants to be a cyborg? No, as I argue here. In fact, I'm skeptical that such implants will ever be able to give anything more than a constant speed up (from lower latency of I/O) over external augmentation, which washes out for years long projects. Why? Well, for a start, we don't even know how to restore function to paralyzed or otherwise afflicted individuals. More important, attention, heat and distributed representations. We seem unable to learn without forcing things through the trickle of a few bits/second of attention.

To be creative, you need to learn in a way that forms connections over many regions. Building a device that integrates with all of your memory, knows just what to encode where, while considering heat limitations, seems a very tall order. Consider another angle: do you think neural implants are near? Many people think this but think virtual reality without any external devices is far away or that memory rewrites and erasing are also far away. But both those technologies are needed long before a non I/O type interface will ever be feasible!

I've been working on this project on and off since about 2014. There have been stretches where I could not work on it and other stretches where I was plain stuck. The latter is to be expected—research is just search and like all search, requires exploring down a teeny, tiny fraction of the branches in a tree of possibilities. Most of these will end up as dead-ends and require backtracking; you can't hope for better than steady, stalling progress with limited reversals.

Building intelligence amplifying tech is also difficult due to the mix of skills it requires. You need machine learning and in particular, natural language processing to be able to get a bit more (but still poorly) structured representations of text. You need tools that are fast and responsive for very large amounts of text, eliminating many Double Decker Bidirectional neural network LSTM architectures.

You need good UX because the system will be interactive. For example, after trying many iterations on interaction, I decided that to achieve the best balance of ease vs power, a notebook format and a simple language that composes certain high level functionality in order to allow flexibility without having to anticipate and explicitly code all functionality would be most ideal. Finally, a project like this requires understanding intelligence on a deep level in order to better workout a framework to structure things within. Because the human is not addressing an oracle, things like learning to think, effective ways to decompose a problem and knowing how to design queries within the language (and making a good language for this) are important too!

Finally, I want it to work on machines of modest means (low RAM, no GPU, okay processor). The learning must also happen from scratch on your machine and must start being useful within an hour. These tight constraints have made the project far harder than I originally thought it would be but I'm happy to say, I've made steady progress despite the many obstacles—within and without the project—that I've faced.