How far can we go with Artificial Intelligence?
Scenario analysis on Artificial Intelligence
A lot of people suggest like it’s a no brainer that Artificial Intelligence will reach human-level intelligence in a relatively short time frame. Some experts suggest that this could happen as soon as in the coming decades. Others even speculate that we’ll be able to upload our brains into computers and manipulate consciousness at will. Others are convinced that we are at serious risk of AI doomsday scenarios. In particular, the famous Singularity where AI will spiral out of our control, has received a lot of media attention.
However, as of today, we have to admit we have no clear idea of how and when this will happen. Before embracing the most aggressive predictions, we have to consider a number of important elements.
On the hardware part:
Optimistic predictions are often based on endless extrapolation of Moore’s law. However, in the last ten years, this “law” has shown dramatic diminishing returns and is about to break entirely pretty soon. New technologies like optical or quantum computing are still facing daunting engineering issues before they can be seriously considered as viable alternatives to resume Moore’s law.
Compared to “wet computing” (human brain), silicon-based computing is extremely inefficient both in term of energy consumption and heat dissipation. Given those constraints, it’s not obvious that a future human-level computer will be based on silicon. If this is confirmed, we’ll have to rebuilt from scratch our hardware. And today, we have no idea how to do this.
On the software part:
In the last 50 years, it’s embarrassing to admit, but we have done very little progress in our understanding of actual consciousness and self-awareness. Who can claim with any confidence that we’ll be able to make any progress in the next 50 years?
Today, even our most sophisticate AI programs have no idea what they are doing and why. Neutral networks are just fancy labels to describe sophisticate nonlinear statistical calculation. As a result, even the most advanced AI programs can be easily fooled with relatively basic tricks.