tl;dr

It seems likely, based on current capabilities of Transformers, that in the near future humans will engineer systems to maximize real time communication bandwidth between AI and all physical systems (including humans) by saturating networked models of hierarchical size between data centers and edge devices with streaming inference computation. This could accelerate AI’s gathering of new training data through rich interactions with external systems, increase the effectiveness of existing alignment capabilities, and amplify human intelligence. Humans can build infrastructure to steer this capability toward net beneficial applications and alignment of existing systems.

In this post, I:

Consider that AI communication bandwidth and predictive power may enter a recursive loop
Reason about implications
Present experimental results and analyze the Transformer's potential to scale AI communication bandwidth
Discuss useful human actions to hedge against the potential risks of this overhang collapse

What's an example of high bandwidth communication?

Borrowing from Connor Leahy’s post, let’s start with three (unqualified) assumptions:

“All human behavior and thought is fully describable by some Turing-complete computation.
Computing hardware and algorithms will continue to improve until they hit some physical limit.
That limit is still very far away, and the human brain is nowhere near it.”

From these assumptions, we’ll construct an example scenario and work toward deriving the following statement: at some point in time, an AI can approximately simulate human behavior far enough into the future, and with enough accuracy, to mask network latency, resulting in imperceptible human-AI communication latency.

A human is in an empty room with a personal computer, a webcam, and a microphone. An AI is connected to a powerful data center across the internet and the not-so-powerful personal computer. The AI can listen to the webcam, microphone, and text input and control a 2D avatar to communicate with the human via speech, text, facial expressions, and gestures.

The AI streams its perception to the data center where it uses a large amount of parallel computation to approximately simulate its future perception and actions. This approximation, a large vector, is sent back to the personal computer. The AI then uses the personal computer to slightly correct this approximation based...

Posts

Wikitag Contributions

Comments