Graham - what probability would you assign to AGI-5 in 5 years?
0.1%? 1%? 10%?

Five years ago, it looked like human level visual recognition was impossibly far in the future.  Many said that hitting that performance bar would require fundamental new insights into AI.  It did not.  All it required was minor improvements to existing ANN algorithms - combined with scaling up to appropriately large amounts of neurons and training experience.  There are still challenges in vision - video, 3D object localization, etc - but almost nobody in ML now believes that solving these subproblems will require huge new insights.

Likewise five years ago, few believed that ANNs could be used to build a complete complex AGI.  Pattern recognition is important, but there is much more going on in brains - complex sequential planning, attention, episodic recall, etc etc.  Many believed AGI would require a mix of different algorithms.  Then the deepmind guys trained an ANN to play atari and showed that you could achieve human level play across a wide spectrum of games using deep reinforcement learning across an appropriate initial architecture.

All of this has been achieved by small teams of researchers running models with 10 million neurons or less on just a few GPUs.  It has actually required new insights to get this far - many of them in fact - but all fit within a general optimization theory of intelligence that has been solid for a while.

We can achieve AGI-5 by:
*scaling up to millions of GPUs each running billions of neurons (embedding in a video game, for example)
*scaling up the number of researchers to thousands, augmented with new global opt/research automation tools
*creating a new framework to organize the code/design exploration process and manage incentives.

The above combination will accelerate research by a factor of about a million, and will lead to a reasonable probability of AGI-5 in 5 years.

Accomplishing the above will not be easy, but there is literally nothing else more worthwhile.  Even if it only has a 1% chance of leading to superintelligence in 10 years, 1% of a 100 trillion dollars (lower bound on the value of superintelligence) is still a trillion - and justifies almost any investment.

Everyone is making a bet, like or not.  If you dont invest in AI (and in the right project) in the next 5 to 10 years, you will find your wealth/job/livelihood eroded away regardless.