Artificial General Intelligence Limitations

Ray Kurzweil wrote the Singularity is Near in 2005 and he described the near inevitability of the creation of super intelligence and a Technological Singularity. Kurzweil built upon the Singularity ideas of Vernor Vinge. In nearly two decades, computers have continued to improve, algorithms have improved and AI has improved. The Deep Learning and Reinforcement learning approaches to AI are very financially successful and have made progress on super human vision systems and the complex game of GO.

Ramaz Naam observed that the intelligence explosion arguments assumed that an AI would be able iterate on creating superior AI. However, if better intelligence becomes more difficult then it limits the rate of improvement.

There are problems that become exponentially more difficult as their size increases. This is for problems like the Traveling Salesman problem.

The chess AI software became better than human chess players. However, the chess software reached limits. Non-deep learning chess programs reached a peak with Stockfish with an ELO (chess rating) of about 3550.

Muzero is the latest deep learning chess program. It can achieve an ELO chess rating of 5400. Something that has an ELO rating 200 points higher than something else should wind 3 out of 4 matches.

Magnus Carlon is the best human chess player with an ELO of about 2850. This would mean Magnus Carlson might win one game out of 1 game out of 55 versus an older Stockfish chess program. If someone had an ELO rating of 2400, then they could give an ELO 1800 player knight odds for an equal match. Knight odds is the better player starting down a knight. The ELO points for material increases as the ELO increases. Knight odds would make a 1100 ELO equal to a 1400. Lower ranked players make more blunders.

In chess, intelligence (ELO rank) advantages decrease as intelligence increase or error rates decrease. Checkers is a simpler game than chess. Computers are able to complete solve the game of checkers. They can play perfect games of checkers. The best human checkers player, Tinsley the terrible, only made about seven mistakes over several decades of tournament play.

Human Intelligence and current Artificial Intelligence have many limitations. Humans are limited in time, computation, and communication, defining a set of computational problems that human intelligence has to solve.

Limited by the Technology of Their Time

Even if the AGI is superior, creating the next superior iteration of AGI could involve time consuming processes to develop machines to make the machines, new materials, and research and experimentation. The CDC 6600 was the world’s fastest computer in 1969 and it had three megaflops of processing power. We are now approaching ExaFlop supercomputers. Our computers are 100 billion times faster. We are still working to improve the software.

Tesla and others are spending billions to develop AI for self driving. One of the critical factors is gathering the video of billions of miles of driving examples. Self driving will be highly valuable AI. It is taking a decade or more of very well funded effort to develop.

Self-driving car software will enable superior mobile robot AI.

Tesla, SpaceX and Elon Musk are making vastly more intelligent choices than their competitors. However, it is taking a couple decades for dominance to be achieved.

There are situations like Chess and Go or processing of certain medical images where new AI has rapidly emerged and displaced the current humans in the field.

We can also have thought experiments where humans go back in time to different points in history. We can imagine the maximum disruption and impact that superior intelligence and knowledge could provide. Displacing lower information or less intelligent competitors would be easy. The rate of techological improvement would be faster than the trajectory before superior intelligence and knowledge was added. It would take time for those who traveled back in time to rebuild to our current level of civilization and continue advancement.

What is Hard and What is Surprisingly Easy

Many people talk about needing AGI or a technological singularity to solve the major problems that we have today. The major problems that people identify are generally all solvable with the proper application of current technology.

Feeding everyone, even with a much larger population. China is spending about $100 billion to build greenhouses that will grow food ten to twenty times more efficiently than outdoor farming. This would also use vastly less water and this has been proven at national scale with simpler plastic sheeting greenhouses in China and with farming in the Netherlands. $500 billion using technology originally created hundreds of years ago can enable us to feed 20 times as many people. The greenhouse farms would also be immune to any forecasted climate changes over the next 500 years.

Air pollution. Air pollution is slowly getting better. All cars and trucks will be electrified and fossil fuel energy is being eliminated. This process just takes about 50 years to fully scale.

De-carbonize the atmosphere. This is again a solvable problem without super technology. Fast growth tree species can grow in ten years or less and each tree can hold ten tons of carbon in its wood. If we massively reduce the land used for farming crops with greehouses, we can use more land for growing trees. The trees can be harvested and the wood holds the carbon until the wood decays. If the heavy machinery for cutting and processing the wood is made electric we can make a relatively low cost, low pollution system for removing carbon from the atmosphere at scale.

Iteratively improving the intelligence of civilization and continually accelerating the rate of improvement is more difficult. It would be possible for everyone to learn the methods and management of Tesla and SpaceX. Factories could be redesigned and rebuilt on a two to four year cycles and could even reach one year cycles of re-invention. This would not even require superhuman intelligence. It is taking the best of what is currently done in two organizations and spreading best practices throughout society.

Conquering Space. We are now overcoming unnecessary delays and barriers to progress. The fully reusable rocket should have been developed in the 1980s with a correct version of the Space Shuttle. This would have required working out vertical landing of boosters and not just landing the Shuttle upper stage. Greater intelligence is not needed. We just need less corruption and less stupidity. We need time and continued execution.

Conquering aging seems to be on track with aging reversal and aging damage repair. The aging reversal industry seems to be on a trajectory and approach which will deliver success.

Many Narrow Superintelligences Could Be Superior to Artificial General Intelligence

There is a large problem space for any potential AGI. Currently, it appears that narrow AI will continue to have advantages over general AI. A narrow can have more dedicated and concentrated data and computational resources to advance problem solutions. Google has created a data and resource moat around their search and advertising solutions. It is not clear that AGI would always beat many narrower AI across all problem areas. A slightly more generalized AI, muZero, was superior in chess, Go and Atari versus prior special purpose game AIs.

We will have to see how narrow AI competes with more generalized AI over time.

SOURCES- wikipedia, Kurzweil
Written By Brian Wang, Nextbigfuture.com