Part 2: The dawn of a new species
Technology was at first only for specialists. As it evolved, it is now used universally by anyone.
Scientists on the other hand started as generalists and ended up as specialists. These opposite directions is because human and artificial intelligence are dealing differently with complexity. In this blog we explore how AI systems already are good as specialist experts, but may become generalists as well.
The renaissance age brought some of the brightest scientists of all time. The so-called renaissance men developed his abilities in all kinds of areas like art, philosophy, religion and science. Consider the following thinkers:
- Leonardo da Vinci
Copernicus is especially known as astronomer who placed the sun rather than the earth in the center of the universe. However he was also educated at canonical church law, worked on economics principles (such as the quantity of money) and worked as governor and diplomat for the Polish king.
Erasmus also known as the “Prince of the Humanists” worked as translator (from Greek to Latin), on philosophic topics (like Free Will), on religious topics (Protestantism), but also wrote about pedagogy (On Civility in Children).
Leonardo da Vinci wasn’t just a painter. He did research on anatomy and physiology and as inventor draw all kinds of technical drawings.
These renaissance men are today more seen as thinkers than as rigorous scientists. From the Renaissance to the 21st century more and more knowledge was necessary to made any progress in science. So generalists became specialists. It’s impossible now to excel as astronomer while also work as an economists and a painter.
One of these modern experts specializes in the artificial intelligence field. These experts are creating specialized algorithms who excel in a specific domain. Examples are:
- Autonomous Driving
- Contextual Code Completion
- Fraud detection
- Stock market patterns recognition
- Object recognition
- Analysis of medical images
These are all real world usages of AI where AI assists human intelligence in a specialized field. Are these AI not experts themselves? This a debate among AI scientists and philosophers. Are these narrow-interest AI more tools or is it a form of intelligence?
In the examples on AI every use case uses a different approach. This may lead to expert knowledge, but not to wisdom. AI delivers already results that surpass human ability. Much like a car drives faster than we can run and a calculator can solve a math problem faster than we can calculate. This is not something we would call general intelligence.
Maybe we should not attach intelligence in human sense too much to AI. It’s not general intelligence, but an intelligent generalist we seek. To become a generalist AI needs to calculate, but also translate a sentence to Chinese, play a game of Chess, drive a car and write this blog.
To reach this level a combination of different AI approaches is needed. But before we explore such a combination of advanced algorithms we go back to the most basic software: command line programs.
In the history of computing, the first operating systems tasks were performed by command line tools with text input and text output, for example the program: “cp”. A program to copy files or directories. You can use it on the command line like this: “cp a b” (copy files from a to b).
There are a lot of possible use cases. For example a destination already exist or only a subset of all files needs to be copied. This can be controlled by parameters which make a simple program more versatile:
An operation system has of course a lot more tasks than copying files. Think of deleting, creating, searching files or all kinds of other tasks. For this the GNU command line toolset was created. This consist of more than 100 utilities with dozens of parameters.
Of course there are thousands of other programs for specialized tasks. Old-skool system engineers know a lot of programs and its parameters by heart. They created scripts and cheat sheets to control them. When new to the command line the number of tasks and how to perform them can be daunting.
IBM recently brought AI to command line. With their program users can explore the man pages. These are manuals with explanations of all the parameters. This is kind of an über program which can make use of hundreds of programs and thousand of parameters. Still this is quite domain specific. If we want to do more tasks with flexible input we need to fill the AI with lots more data than manuals. This data is often labeled so that the program can use millions of parameters.
Shocking the world
Currently, there is program which is fed with an enormous amount of unlabeled and unstructured data from websites and books. The program: GPT-3. This AI uses several machine learning techniques to create all parameters itself. The result is a program with 175 Billion parameters. To use it, you don’t need to follow a manual, but just ask it to perform a task. This AI shocked the IT world.
Let’s consider another program which shocked the world, AlphaGo by Deepmind (Google). This is not a program about text input and text output, but one that plays the game Go. Because the almost infinite number of moves a player can make in this game, it was impossible to compute all the moves. AlphaGo developed a strategy from a finite number of steps to defeat every human player.
The AI researchers developed AlphaGo further in its successor AlphaZero which was more powerful and could play Chess and Shogi as well. This approach is fundamentally different from that of GPT-3 in that is starts from scratch and with the rules of the game uses reinforcement learning to become expert level. The AI plays millions of times against itself, before competing with a human.
Deepmind (AlphaZero) and OpenAI (GPT-3) have thus fundamental different ways of approaching AI. Whereas Alpha Zero can learn by reinforcement learning (playing against itself) with only from the rules and goal of game as start, GPT-3 based on the input of millions of texts which labels itself to produce all kinds of output texts.
Both approaches have no real world knowledge so they don’t have a kind of common sense. This is troublesome to move these approaches forward.
Now there is yet another approach which is called COMET. This is a meta model approach which allow a common sense way of reasoning as explained in the following article
Common Sense Comes to Computers
One evening last October, the artificial intelligence researcher Gary Marcus was amusing himself on his iPhone by…
Combining AI approaches
All approaches are still more fundamental research endeavors than the real world examples we earlier gave. Still they are key for making long-term progress. AlphaZero, GPT-3 and COMET all are parts of the puzzle for general artificial intelligence. The combination of these approaches would create a much stronger AI. A combination of task, goal and reasoning orientated ways of coping with the world.
So instead of experts in their field, combined AI functions more like a network of experts. Much like human scientists cooperate through universities.
Homo universalis became a specialist which created AI experts that becoming AI universalis. This means we may get advice in a way Coopernicus, Erasmus or Leanardo da Vinci would give us. As a specialist it’s namely hard to find cross-links between your own subject area and other disciplines. These links are extremely valuable. It takes mostly a certain of brilliancy to connect the dots in new ways. So we may have a lot of knowledge and insight on a particular subject, but this doesn’t mean we can find links to other fields, as for the simple reason that we don’t have deep knowledge in these fields.
AI is good in taking objects, ideas, methods from totally different fields or unrelated subjects and applies this to all other kinds of use cases and situations. MIT used AI for example to discover links between works of art.
Here is one of the results:
In the current state AI is not capable to give meaning to such links, but it can work as a new kind of recommendation system. Like YouTube or NetFlix recommend a video, these AI systems can recommend a scientific link. Researcher can investigate unexpected links leading to whole new discoveries. In the future scientists and AI working together can be both specialist as generalist.