Four paths to superintelligence could accelerate one another, says a researcher at Google DeepMind's AI lab
Four paths to superintelligence could accelerate one another, says a researcher at Google DeepMind's AI lab
On November 19, Tim Genewein described how AI might develop after reaching human-level intelligence. He defines superintelligence as a system that outperforms large groups of experts in almost every intellectual domain, completing tasks that would take them years, faster and with fewer resources.
A trained digital specialist can be copied along with its memory, and many copies can run in parallel. Genewein estimates that total effective computing capacity has grown at least tenfold per year over the past decade. He arrives at this estimate by multiplying the rates of improvement from falling hardware costs, rising investment, and more efficient algorithms. More efficient algorithms do the same work with fewer computations. If these trends continue, far more copies of AI systems with today's capabilities will become available. He identifies four paths to expanding those capabilities.
The first is to increase model size, the scale of training, and the computation used to solve problems. Training requires ever larger volumes of high-quality data, while additional computation lets a model spend longer searching for an answer. The second path is to change how AI systems are designed and trained. Genewein associates this path with continual learning from experience, a reduced need for training examples, and the creation of new concepts.
The third path is to bring AI agents that carry out tasks independently together as a team. As in a research institute, specialists could divide a complex task into parts and combine their results. The benefit of adding team members depends on how many resources go into communication and coordination.
The fourth path is a repeated cycle of improvements to AI itself. For example, a system develops a better algorithm, the next version uses it, and that version helps create an even more capable system. Another option is to use high-quality AI outputs to train subsequent models. Within a team, specialization may free up resources to expand the team, allowing an even finer division of labor.
The pace of self-improvement depends on how a system checks its results. A flawed criterion may reward a useless solution with a high score.
“Self-improvement cycles can easily stall or deteriorate when feedback is poor.”
In mathematics, a proof can be checked automatically, so the system quickly learns whether an attempt has succeeded. Genewein expects faster self-improvement in this area. In research involving physical experiments, the pace also depends on the time spent waiting for experimental results.
According to Genewein, if restrictions apply only to training computation, algorithms and the organization of teams can continue to improve. When a restriction is lifted or circumvented, more computation combined with accumulated improvements could produce a sudden jump in capabilities. Tracking this development will require several measures at once: the capabilities of an individual model, the useful work a team produces, and the reliability of the checks on its results.