Table of contents
- The paradox at the core of this phenomenon is why machines are increasing their intelligence exponentially
- Moravec’s paradox
- From computation to culture
- Humans are good at perception and mobility, but not so good at anything else
- Why robots struggle with sandwiches
- Economic considerations
- Final impression
The paradox at the core of this phenomenon is why machines are increasing their intelligence exponentially
When I started 1SecondPainting , I thought AI had crossed a line of no return.
Machines have improved their capabilities over time. But I (and many others) thought the ability banished simple, monotonous tasks.
The jobs that machines did were, for example, operating assembly lines and assembling automobiles. AI is “for” automating that kind of work, and I didn’t feel uncomfortable with that.
But the events of 2020 changed everything.
Humans have had a long relationship with technology, but in just a few months, modern artificial intelligence has graduated from that relationship. In June of that year, GPT-3 was released to the public. And in December, one of the most shocking advances in generative machine art, VQGAN , landed with a big splash.
For the first time, AI has replicated human creativity. Concepts such as art, design, and language that were traditionally thought of as human nature are no longer perceived in that way. And few understood why.
However, this trend of increasing machine capabilities was actually predicted about 40 years ago . And based on this prediction, we can see where our society is headed next.
In the late 1980s, computer scientist and roboticist Hans Moravec proposed an interesting paradox .
Specifically, the paradox is that it is easy for computers to do what humans find difficult (intelligence), and difficult for computers to do what humans find easy (perception and mobility). .
For example, advanced mathematics ability is the true value of human sharp intelligence. Most people find such an ability very difficult to acquire, and it takes years of dedicated study to be able to solve even a few of the word problems found in standard calculus textbooks. Become.
But machines can do any level of mathematics without difficulty. No mathematics, be it linear algebra or differential geometry, is more than simple arithmetic to a machine.
From computation to culture
The following is probably familiar to readers. Computers have always been, well, good at computing . We made it to do the math.
But now computers have gone beyond just mathematics. Moravec’s paradox is beginning to apply to all intellectual work , regardless of discipline . Its scope has grown to include language, formal reasoning, and artistic creativity.
The GPT-3 already writes better than the average college graduate, and the DALL-E 2 and Imagen are better than the average artist. And new language models such as LaMDA (2) and Flamingo (3) are set to solve complex reasoning problems that intelligent humans struggle with on a daily basis.
Within a few years, AI will be able to produce world-class works of art, full-length books and white papers, and songs of near-infinite complexity. In other words, it will be the role of machines to further develop human culture.
But why did the computer become creative? How did he do in just a few years what has taken humans thousands of years?
Humans are good at perception and mobility, but not so good at anything else
In terms of evolutionary history, the secret to increasing computer creativity is a fairly simple question. The human brain has taken much longer to evolve for survival than for the esoteric pursuits of logic, reasoning, and art.
In other words, mobility such as walking, running, grabbing things, and balancing is too easy for us humans to even think about. Do them unconsciously . Most of our brains have been optimized over millions of years, and humans are literally made to move.
But can a human being write the next Moby Dick or do math on a difficult problem? Because these activities are relatively new in human history, there hasn’t been enough time for evolution to build brain structures that could easily solve such difficult problems. As such, these tasks require an enormous amount of thought, focus, and effort.
Language, art, and meaning, therefore, are not inherently difficult problems. Because humans have always tackled them with very limited tools, these problems have seemed difficult . Even if you give modern AI such a problem, it doesn’t feel the intrinsic difficulty that we do. As Moravec thought, it is easy for computers to do tasks that we find difficult.
Why robots struggle with sandwiches
On the other hand, even very simple things that humans do can be very difficult for machines. Consider, for example, the less well-defined task of making a ham and cheese sandwich.
The ” algorithm ” for building a human sandwich is 1) find the bread, 2) pat the bread on the counter, 3) serve the ham and cheese, and 4) assemble the ingredients. Then put everything else back in the fridge and munch on it.
On the other hand, it is very difficult for computers to do this task, mainly because of perception and mobility.
Imagine the year 2050, when a generic “helper” machine was born to help humans with household chores. Here’s a glimpse of what this machine has to think about to make a ham cheese sandwich.
- Where is the refrigerator relative to the helper body?
- How much resistance can you expect to encounter when trying to open the door?
- How far can you lean forward without losing your center of gravity?
- Approximately how much does the container you are reaching out weigh?
- What kind of force do you need to use to keep the container you are holding from slipping off?
- What is inside the container? How does your movement affect the content? If you apply too much force or force too fast, will you crush them?
- How strong should the container be on the counter?
…The above is the work before assembling the sandwich .
It takes an enormous amount of engineering ingenuity to make a robot walk , open a door, make a sandwich, and do other common tasks. Human children, on the other hand, can accomplish them with little effort.
This is at the heart of Moravec’s remark that “it is difficult or impossible to give (machines) the skills of a year-old when it comes to perception and mobility.” Our brains subconsciously process vast amounts of information related to perception and mobility, which makes them the easiest tasks for humans in the world.
From the above considerations , I hope you have a vague understanding of why intelligent models focused on language and reasoning have evolved so quickly compared to robots.
However, the consequences of the above considerations lie further ahead. In addition to the problem of scope (where we work on language and reasoning), there is also the economic pressure to quickly solve intellectual problems on the market.
Now that machines are acquiring creative skills, the global economy may once again grow exponentially as machines take on creative jobs.
Part of the pressure to force computers to perform intellectual tasks is that this pressure facilitates some of the most economically impressive feats in history. A universally capable language model with sufficient bandwidth would instantly multiply the financial value of any nation-state by a factor of ten. All entertainment, negotiation, R&D, and business will be automated, and the quality of each task will be orders of magnitude higher than what humans can do.
The reason for the arrival of such a future has been detailed in a previous article, so I won’t go into it here. But suffice it to say that when comparing, say, a super-intelligent language model with the expected economic impact of a sandwich maker, the choice of venture capitalists is clear.
Moreover, it is much easier to use software than hardware to do repetitive tasks. Software can be deployed instantly to millions of devices with the click of a button. Hardware, on the other hand, requires physical upgrades, maintenance, and logistical considerations. As you continue to use hardware, it becomes more costly, more options to consider, and consequently harder to maintain.
Humans will eventually refocus on perception and mobility. But such a return will not occur before intellectual tasks such as language, art, and reasoning are largely resolved. And by the time that regression occurs, AI will likely be at the forefront of research and development.
(*Translation Note 6) In a blog post published on May 3, 2022 on the personal website run by Nick Saraev, titled ” Future Socio-Economic Impact of Artificial Intelligence, ” GPT-3 and DALL-E 2 The following two scenarios are discussed regarding the impact of the spread of creative AI on the global economy.
Two forecast scenarios for the global economy envisioned from the spread of creative AI
|(Scenario 1) If the spread of creative AI is not properly regulated: A serious unemployment problem will arise , especially in the creative industry . Furthermore, while wages for immature workers will fall , wealth will concentrate in some occupations that develop and manage AI . As a result, large economic disparities will arise, and political situations around the world will become unstable.
|(Scenario 2) Appropriate regulations for the spread of creative AI: Appropriate AI regulatory policies that impose a higher tax rate on products produced by AI than those produced by humans are implemented . And the tax revenue obtained from the policy will be used for relief measures for those who have suffered unemployment or wage declines due to the spread of AI. As a result, we can enjoy the economic benefits of the spread of AI in a moderate manner without creating extreme economic disparities . The economic disparity itself will not disappear, but it is not bad either.
Scenario 2 can be interpreted as an AI- powered utopia, but Saraev warns that there is no guarantee that governments will be able to implement appropriate AI regulations. Final impression
In short, Moravec’s paradox is the reason for the rapid evolution of AI in the fields of language and art. It is also why the fields of robotics and locomotion have lagged far behind our expectations.
Given enough time, it is clear that AI will eventually surpass human intelligence in all relevant areas. During the grace period before AI fully surpasses human intelligence, let’s not forget that the skills we humans have most naturally are often the hardest skills for machines to learn. And let us plan for the near future, where AI is still inferior to humans.