Loe raamatut: «Machine and human intelligence. Updated research»

© Zero Ash, 2024

ISBN 978-5-0062-7940-7

Created with Ridero smart publishing system

Machine and human intelligence, an updated research

MHIUR 2024

Research of machine and human intelligence based on own knowledge and experiments, analysis of various sources: publications of scientific articles, posts, interviews, news updates, etc. Information for a wide audience, from the very entry level, to experts. The material is more about questions than about answers. Intelligence is an unsolved mystery of nature.

«Intelligence is a human construct to represent the ability to achieve goals.» (Michael E. Hochberg, «A Theory of Intelligences»). Simple and straightforward, but intelligence is not only humans have, and what intelligence is?

«A core function of intelligence is grounding, which is the process of connecting the natural language and abstract knowledge to the internal representation of the real world in an intelligent being, e.g., a human.» (Bing Liu,«Grounding for Artificial Intelligence»). Good point, but intelligence has many more core functions.

«Consider the space of all possible forms of intelligence, what can be called the intelligence

space.» (Paul S. Rosenbloom, «Defining and Exploring the Intelligence Space»). Interesting concept, but even to imagine such «space of all possible forms of intelligence» is close to impossible task.

«In the lone banana problem, the statistics suggested that bananas only appear in twos (or more) and so the AI could not imagine a single banana, because the data and parametric tuning that had gone on didn’t allow it to consider that approach, on average» (Daniel Hook,«The Lone Banana Problem»). No surprise, machine models can only use patterns from their training data and can’t come up with something totally new. «AGI» – did you say, give me two.

«AlphaGeometry is a neuro-symbolic system made up of a neural language model and a symbolic deduction engine, which work together to find proofs for complex geometry theorems. Akin to the idea of „thinking, fast and slow“, one system provides fast, „intuitive“ ideas, and the other, more deliberate, rational decision-making.» (Trieu Trinh and Thang Luong, «An Olympiad-level AI system for geometry»). An interesting idea, a combined approach, in addition, code generation and execution can be used.

«The artificial intelligence may share different principles from the natural intelligence, but both can inspire each other, which may call for establishing new mathematical physics foundation.» (Haiping Huang,«Eight challenges in developing theory of intelligence») – there’s a joke that goes: «if you want to figure out how it works, program it.»

«Machines will be capable, within 20 years, of doing any work that a man can do.» – wrote Herbert Simon in 1960, who won the Nobel Prize in economics and the Turing Prize in computer science.

«Every intelligence is specialized, including human intelligence. Intelligence is a collection of skills and an ability to acquire new ones quickly. It cannot be measured with a scalar quantity. No intelligence can be even close to general, which is why the phrase „Artificial General Intelligence“ makes no sense. There is no question that machines will eventually equal and surpass human intelligence in all domains. But even those systems will not have „general“ intelligence, for any reasonable definition of the word general.» (Yann LeCun). AI, AGI, etc. – is just marketing, nothing more. There are many questions in what areas and what machines can do. Computers are just tools, they have no will of their own. A model of will is possible, but it’ll be just a model. Truly autonomous systems and self-aware systems are projects of the very distant future.

«People don’t change their minds.» (Daniel Kahneman) – partially true, there are other words – «People think one thing, say the second and do the third.»

The limits of my language mean the limits of my world (Ludwig Wittgenstein). It shows that language plays a key role in everything, intelligence and language are tightly connected. To hack an intelligence, hack a language.

«My machine learning professor taught me something I’ll never forget: Always focus on the analysis, and don’t worry about the code. Copy the code. Steal it. Ask somebody to write it for you. It doesn’t matter. Coding is easy. Knowing what to code is what truly matters.» (Santiago Valdarrama). That’s what distinguishes intelligence, to understand what’s important, what you should focus on, and what you shouldn’t waste resources on.

«AI can be considered something that mimics human cognitive functions.» (Igor Ashmanov). It is not a bad definition, considering that AI has different definitions: strong, weak, general AI, etc. You can give another definition yourself, and it will have the right to exist.

«Education is what remains after one has forgotten what one learned in school.» – Albert Einstein. Yes, that’s right, knowledge and intelligence develop only in work. Unused knowledge dies and is forgotten.

«The Truth About Emotional Intelligence. For those who have it, it predicts success in many ways.» (Marc Brackett). EI is an essential part of an intelligence, perhaps even more significant then everything else.

«The capacity to understand the world, understand the physical world, the ability to remember and retrieve things, persistent memory, the ability to reason, and the ability to plan are four essential characteristics of intelligent systems» (Yann Lecun). It is not a good definition, what it means to understand the world, to reason, to plan. It is not easy to give a good definition of intelligent systems, but it is easy to replace some abstractions with another abstractions.

«How does next-token prediction in large language models (LLMs) yield remarkably intelligent behavior? … These models have demonstrated extraordinary capabilities beyond just mastering language.» (Ibrahim Alabdulmohsin, Vinh Q. Tran, Mostafa Dehghani, «Fractal Patterns May Unravel the Intelligence in Next-Token Prediction»). There is nothing remarkable, not intelligent because all answers are manually prepared. Next-Token prediction is just a way to probablistically store and retrive data, there is no intelligence at all. LLM can be thought of as an interactive dictionary or encyclopedia with an advanced query language.

In his paper «Computing Machinery and Intelligence» Alan Turing asked the

question «Can machines think?» and introduced the Turing Test to verify whether machines could achieve human-level intelligence. The problem is that it is possible to create an emulation of thinking. Modern computing systems can process all the information collected by mankind, moreover, a huge number of human editors (AI-trainers) can be used to correct such systems. The old Turing test does not make sense, but it is possible to develop a test in which there should be no previously known typical patterns of reasoning. The new test should be based on the ability to generate and test hypotheses, build probabilistic chains of reasoning, without the ability to solve the problem by brute force.

«If you are the smartest person in the room, you need to find a smarter room.» (Ranal Currie) – definitely, environment is extremely critical, communications with smarter persons than you will make you smarter.

«These 4 phrases you have higher emotional intelligence (paraphrase, ask questions): „What I hear you saying is..“, „Let me get this right..“, „How did that make you feel?“, „What might have led you to that?“» (Aditi Shrikant, «If you use any of these 4 phrases you have higher emotional intelligence than most») – summarizing, paraphrasing and asking questions indicate that you are actively involved in conversations what gives you a basis for emotional intelligence usage.

«Intelligence is multidimensional, and therefore there’s no one point at which AI will exceed human intelligence.» (Pedro Domingos). Sounds good, especially since current AI is just bits of human work wrapped in technology.

«The Measure of Intelligence is The Ability to Change.» – Albert Einstein. An obligatory property of intelligence, without the ability to change intelligence is impossible.

«Intelligence is the ability to adapt to change.» (Stephen Hawking) – echoes Einstein’s previous quote.

Intelligence

No one really knows what intelligence is and how it works, although we all deal with intelligence all the time, both our own and the intelligences around us. Intelligence is both an individual and a social phenomenon, perhaps intelligence can only be developed in a society where both knowledge transfer and exchange of experience and competition are possible. All this contributes to the development of intelligence.

There is a generally accepted understanding of intelligence that is more or less consistent both at the individual level and among various institutions. Many would agree that the hallmarks of intelligence would be the ability to learn and interact socially. Also the ability to apply the knowledge and skills acquired to solve problems, achieve goals, and develop, both one’s own and group success. It is important to have different views on the nature of intelligence and how best to define and measure it.

Philosophers have debated the nature of intelligence for centuries. Some view intelligence as the ability to reason and acquire knowledge through logic. Others view intelligence as acquiring knowledge from experience and the senses. More recently, developmental theories view intelligence not as a fixed ability, but as an ability that grows through active exploration and interaction with the environment.

There is a view that sees intelligence as a property of the mind. Intelligence is proposed to be studied through IQ tests and tasks that aim to quantify general cognitive abilities such as thinking, memory, spatial skills and speed of information processing. IQ tasks provide a measure of a person’s intelligence compared to others. There is a valid criticism: IQ tests measure only specific skills, under specific conditions and constraints, missing other important aspects of intelligence. One can train the ability to pass specific IQ tests, but how does that help with real-world tasks?

Anthropologists study intelligence in relation to specific cultural environments. They argue that intelligence has different manifestations in different cultures and corresponds to different historical and social contexts. Applying the values and concepts of one culture to assess intelligence in another can be problematic.

There is a theory of multiple intelligences that suggests expanding the definition beyond verbal and logical-mathematical skills to also include interpersonal, musical, spatial, naturalistic, and other aspects. An interesting concept, it shows that intelligence is not only about high IQ, but also many abilities in different areas.

Artificial intelligence has emerged from the capabilities of computer technology. Which at the basic level can perform primitive operations on transfer, storage and processing of data in the form of sequences of arithmetic-logical operations and program control instructions. Creation of really intelligent systems on the basis of such primitive technologies, with appropriate properties inherent in natural intelligence, is a big challenge. And, by and large, no one has achieved any significant results in this area, despite the hype surrounding the topic. What can be said about today’s AI? There are useful tools, especially based on generative models, which help in solving various tasks, but no more.

The easiest way to study intelligence is from top to bottom, from the general to the particular. We all deal with intelligence all the time. For us, intelligence is a common thing, it is not something outlandish and incomprehensible. This simplicity and accessibility is what is misleading. In fact, we do not understand what intelligence is, or rather we do not understand it at all, although we use it all the time. Yes, there are a couple dozen definitions from different points of view, there is no one universal and comprehensive.

The beginning of understanding what intelligence is, occurs when we realize that we know very little about it, and are ready to be open to new concepts and unexpected knowledge, especially in the light of recent scientific breakthroughs and achievements. Conventionally, approaches to the study of intelligence can be divided into two groups. The first is the high-level study of how intelligence works. The second is the study of natural intelligence at the physical level, from the simplest natural phenomena such as DNA sequences to the workings of the human brain and other living organisms. Building artificial intelligence also contributes to general understanding and progress.

Building intelligent systems or knowledge frameworks is very limited. The essence of the limitation is that there can be only two approaches to intelligence work: 1 – pattern recognition based on experience, 2 – inference based on reasoning. If in the first case everything is clear, consider that immediately there is a ready-made solution or a ranked list in which the most probable solutions are selected. In the second case, chains of reasoning can be very complex, branching, interrupted, etc. It can consist of many stages or steps. At each step, the same principle of two possibilities is at work, a decision pattern may finally emerge, or it may be necessary to go back to reasoning.

Scientific research and hypotheses that have advanced our understanding of intelligence

In the early 1900s, psychologists, including Alfred Binet, began creating the first intelligence tests to detect learning disabilities in children. This marked the beginning of intelligence testing and the concept of IQ.

The theory of multiple intelligences was proposed by Howard Gardner. This theory states that there are at least eight different types of intelligence including musical, interpersonal, logical-mathematical, kinetic, etc. This disproved the idea that intelligence is pro one type such as logical intelligence.

Emotional intelligence was introduced by psychologists Peter Salovey and John Mayer. Emotional intelligence includes abilities such as emotion management, empathy and reading social cues. This emphasized the importance of non-cognitive aspects of intelligence.

Advances in brain imaging such as fMRI, PET scans, and EEG have allowed scientists to study and map brain regions that are activated during different cognitive tasks. This has shed light on the neural basis of intelligence. But this does not mean that we have figured out how everything works; on the contrary, the complexity of the brain is staggering. We can’t, like a computer program, see data dumps and trace execution. We can only record activity, and that’s not enough.

The first ideas of artificial intelligence were born by researchers such as Alan Turing, and the development of AI showed the complexity of the problem and the limitations of human cognition. The AI systems created nowadays are mainly based on machine learning models. The idea of machine learning is that on training data the model can learn to reproduce the required result without explicit programming. The idea is interesting, but there is a clear disadvantage: if the data is not representative, i.e. does not cover all typical cases, the model will not work adequately. In fact, it is an approximation by some signs and examples, which is clearly insufficient.

Studies of twins and families have shown that general intelligence is inherited. Identifying specific genes associated with intelligence is an active area of research. It’s a strange thing, such research should be viewed with disbelief, but this line of research should not be dismissed. Overall, an open question.

Large long-term studies of intelligence in childhood and adulthood have shown that cognitive ability is influenced by health, education, and social factors. This has demonstrated the potential for improving IQ. Of course, intelligence can be trained and developed apart from one’s own efforts, social environment, environment are key factors for success.

These studies and hypotheses have advanced the scientific understanding of human intelligence and emphasized its multidimensional nature and potential for growth. Ongoing research seeks to further uncover the biological, social, and other factors that shape intelligence. We are still only at the beginning of the journey, there is a big gap in our understanding of high-level processes and the physics of how it can work. Also, we should not forget that we have many theories and hypotheses, not absolute truths. New research will break down old knowledge, and new hypotheses will push the boundaries of knowledge.