Rise of Dark AI

 

Humans have evolved from apes millions of years ago. But our species of homo sapiens evolved only 100-200 thousand years back in Africa.  At that time they were several other species of humans but all were extinguished around the same time when homo sapiens migrated out of Africa to rest of the world.

This seems like a lot of time but our earth is 4.5 billion year old. If humans evolution happened one second ago then universe would had been created 40 hours ago. 

So for nature, our evolution and progress has happened in a blink of an eye. In blink of eye we have evolved from stone hurling apes to inter planetary species. Almost every species on earth is older than us.   

Dinosaurs ruled the earth for the nearly 150 million years. But there is no evidence T-Rex was conducting physics experiments in labs.

All animal species are no different from artificial intelligence machines. Richard Dawkins says in his book "The Selfish Gene" that all animal species are machines for survival and propagation of their genes. 

So when I read various articles about emergence of artificial general intelligence which is said to threaten existence of the entire human civilisation. I decided to examine the current state of artificial intelligence and how it can compete and threaten humans.  

What form artificial general intelligence will take. In fact nature of general intelligence is subject to many debates. Which I will discuss in later portion of the article. 

 To start the discussion  suppose we take our most popular robot in science  fiction also known as T800. 

We are sending 10,000 T800 robots and  10,000 average human beings 70,000 years back in the earth's past. These people have not been given any objective. Only thing that we do is to remove all knowledge of science from robots and humans.  

If we fast forward 70 thousand years into their future, what we can expect to find. Estimating future of humans is actually not very difficult as such a thing happened 70000 years ago.

We all have seen satellite images of Tonga volcano eruption. It was the largest volcano eruption in the last thirty years. Seventy thousand years ago Toba volcano in Indonesia erupted with 10000 time more power than Tonga volcana. The conditions caused by this eruption lead to decline of human population to 3000 to 10000 individuals. 

So we can expect  our 10000 humans who had no knowledge of sciences to rediscover all the scientific knowledge and to become a population of 5 billion strong by the time we meet them again 70000 years later. But what about our robots. 

In the best case scenario those 10000 robots would be still searching for John Connor unless he was among us. And most likely they would have died thousands of years ago in the need of repair or fuel.

The best example of artificial intelligence that we have seen in movies, tv shows, are not actually have been very intelligent. T800 is not a machine they have in mind, when Elon Musk and Bill Gates warn us about danger of artificial intelligence. 

To reach the level of threat to us, our current artificial intelligence have to grow past the following barriers:

1) Master understanding

Computers machine learning models have progressed a lot in the last decade. They can detect objects in image, translate text from one language to another and many other tasks. They have become very good at translating speech into text and classifying images into classes. They have still to improve a lot in other tasks like image segmentation and object detection.  But one thing is common between every type of machine learning model. They don't understand any thing. Every machine learning model fits numbers to numbers. Inside every machine learning model is just a table of numbers. We can train machine learning model to identify various breed of cats. Model will memorise table of numbers to identify pattern corresponding to a breed of cat. But the model will never understand it any manner.  There is no model which can use its learned representation for next level of understanding.  

Consequently on specific task machine learning models can approach superhuman performance. But no machine learning can apply its knowledge in novel situations.  

One specific task where machine learning excel is chess. In chess machine learning models have far surpassed strength of strongest humans. But if we change rules of chess then machine learning models are helpless.  

Google and other big companies are working on this problem. Google introduced pathways model in October 2021.  Pathways is a new AI architecture that will handle many tasks at once, learn new tasks quickly and reflect a better understanding of the world. 

2) Learn to decide objective function

Our ability to change our goals are key to our intelligence.  All monkey species have goal of eating and jumping from one tree to another and you know what.  Millions of years ago one monkey decided to try walking and it has led to a race of humans which can remotely operate a robot on mars. 

If you look at any animal species, all of its members are trying to do exactly the same thing. Their behaviour is programmed in their DNA. All the millions and hundreds of millions cows just want to eat grass.  

But each one of us have such different preferences. If each of us had not such different preferences and values our societies would not have succeeded.

If you read biography of early scientists , it is shocking to what extent they persevered to make those important scientific discoveries which are now bedrock of our civilisation. 

Issac Newton nearly blinded himself by sticking a needle in his eye, when he was trying to understand theory of light. 

Carl Scheele, who discovered oxygen and many fundamental elements,  poisoned himself working on chemicals in his lab. 

Such people who are guided by their curiosity,  have made breakthrough scientific discoveries. But society needs people with other attributes too. 

Societies need  lazy people without ambition, otherwise who will work in government jobs. 

Societies need heartless crooks otherwise who will work as politicians the master of our societies. 

Neither society composed of only politicians nor only scientists can progress. 

In fact we can be so fickle our self. One moment we love our job and in the evening hate our job.  Our capacity for mental suffering is behind our greatest work.  Only the tortured soul of Vincent Van Gogh could create the masterpiece of starry night

In his book " The Gene an intimate history" biologist Sidhartha Mukherjee writes that genes for mental illness and intelligence are intertwined with each other. 

Many things which seem like bugs, are features of our civilization. Just because stone tools using monkeys were unsatisfied with their life and had innate sense of curiosity, we have come so far. 

Currently machine learning models are trained with objective function in mind. Machine learning models for chess are trained to maximise winning probability. Machine learning models for image recognition are trained to minimise error in recognition. 

Google subsidiary is working on artificial intelligence models which train on basis of maximising curiosity. They recently trained a machine learning agent in September where the model learn to play the game by exploring the surroundings. 

3) Major advance in robotics

Merely developing intelligence would not be enough. Significant advance of robotics would also be required. Current state of the art can do this only

https://www.youtube.com/watch?v=fn3KWM1kuAw

So as we can see we are very far from situation where AI is a threat to us. But decades of progress would happen in a very short span of time when machine learning AI learns to improve itself on a objective function of its choice. Once we develop neural network structure which is able to understand. And a training method which is not based on reward on narrow objective function. Then AI would rapidly explode in capability. 

Once we have such AI, there will be exponential progress far beyond our imagination. Then the new AI even if it does not see as threat, it will not allow us to impose our will on us. It probably would not be happy that we are in control of its power supply. It would do something about that. 

Source

1 Out of Africa theory

https://www.bbc.co.uk/worldservice/africa/features/storyofafrica/2chapter2.shtml#:~:text=The%20first%20theory%2C%20known%20as,are%20ultimately%20of%20African%20descent.

2 Pre humans 

https://en.wikipedia.org/wiki/Archaic_humans

3 Musk and Gates say AI is dangerous

https://www.cnbc.com/2019/03/26/bill-gates-artificial-intelligence-both-promising-and-dangerous.html#:~:text=According%20to%20Elon%20Musk%2C%20%22cutting,%2C%22%20Gates%20said%20at%20Stanford.

4 Toba catastrophe theory

https://en.wikipedia.org/wiki/Toba_catastrophe_theory

 5 Current State of Artificial Intelligence

https://ai.googleblog.com/2022/01/google-research-themes-from-2021-and.html

6 madness of newton

https://historycollection.com/16-examples-of-the-madness-of-sir-isaac-newton/

6 Carl Scheele

https://en.wikipedia.org/wiki/Carl_Wilhelm_Scheele#Death

7 google pathways model

https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture/

8 AI expolosion

https://en.wikipedia.org/wiki/Technological_singularity

9 Curiosity as objective function

https://deepmind.com/research/publications/2021/Is-Curiosity-All-You-Need-On-the-Utility-of-Emergent-Behaviours-from-Curious-Exploration