In science fiction movies, AI systems like J.A.R.V.I.S. are almost omnipotent, serving as the ultimate assistants to help humans solve various problems.
Behind them is a pursuit of the highest level of AI, a concept known as Artificial General Intelligence (AGI).
The concept of AGI dates back to the mid-20th century when many computer scientists and AI researchers began to ponder how to create computer programs with human-like intelligence. Unlike Narrow AI systems that focus on solving specific tasks, AGI is endowed with broader cognitive and reasoning abilities, capable of learning, adapting, and executing tasks across multiple domains.
However, for a long time, research related to AI has mainly focused on solving specific problems and tasks, and the realization of AGI has always been regarded as a more complex and distant goal.
Recently, Dr. Valentino Zocca, Vice President of Data Science at Citibank, conducted an in-depth analysis of AGI and other related important topics in an article titled “How Far Are We From AGI?” The core viewpoints are as follows:
-
Compared to current Narrow AI, AGI needs to be able to reason and learn across multiple cognitive domains. However, achieving AGI still presents many challenges, such as building world models and performing causal reasoning.
-
Large language models (like GPT-4) excel at solving specific tasks and extracting causal relationships, but lack the ability for abstract causal reasoning. They tend to extract known causal patterns from data rather than generating new causal insights.
-
Some researchers believe that existing large language models (like GPT-4) may be a step towards AGI, but many unresolved issues remain, such as creating world models, achieving self-exploration, and conducting causal inference.
-
Large language models are good at identifying and extracting causal relationships from data but lack the ability to independently reason about new causal scenarios. They have the capacity for causal induction through observation but lack the capability for causal deduction.
-
AI may not truly be able to “learn” but can only extract information or experiences. AI does not form a comprehensive world model but creates an outline.
-
Treating scores as a marker of ability, AI seems to only see a rough overview of the world without truly understanding the essence of the entire world.
-
We should not view intelligence merely as an abstract ability to find solutions to general problems but should see it as a concrete capability to apply solutions learned from previous experiences to different situations that may arise in our environment.
-
Only when we can create a system that can question its own reality, engage in self-exploration, and at least apply causal deduction to build a reasonable world model can we truly achieve AGI.
Academic Headlines has made a simple compilation of the original text without changing its meaning. The content is as follows:
About 200,000 years ago, humans began to walk on Earth while also exploring the realms of thought and knowledge. A series of discoveries and inventions throughout human history have shaped this journey. Some of these not only influenced the course of our history but also subtly impacted our biology. For example, the discovery of fire enabled our ancestors to cook food, thereby transferring heat to brain evolution rather than just to the digestive tract, which propelled the advancement of human intelligence.
From the invention of the wheel to the birth of the steam engine, humanity ushered in the Industrial Revolution. During this transformative journey, electricity greatly spurred the technological advancements we are familiar with. Meanwhile, the printing press accelerated the widespread dissemination of new ideas and cultures, further propelling the pace of innovation.
However, human progress is not solely derived from new material discoveries; it also stems from new ideas.The history of the Western world, from the fall of the Roman Empire to the Middle Ages, underwent a rebirth during the Renaissance and the Enlightenment, emphasizing the centrality of human thought rather than the so-called omnipotent deities. However, with the advancement of human knowledge, humanity began to recognize its own insignificance. For over two thousand years since Socrates, humans have started to “know that they know nothing,” and our Earth is no longer viewed as the center of the universe. The universe itself is expanding, and we are just a speck within it.
However, regarding reshaping our understanding of the world, the 20th century may be the most debated century.
In 1931, Kurt Gödel published his incompleteness theorem.
Just four years later, Einstein, Boris Podolsky, and Nathan Rosen proposed the “EPR Paradox” in a paper titled “Can Quantum-Mechanical Description of Physical Reality Be Considered Complete?” on the theme of “completeness.” Subsequently, Niels Bohr refuted this paper, proving the practical validity of quantum physics.
Gödel’s incompleteness theorem indicates that even mathematics cannot ultimately prove everything; we will always face situations where some facts cannot be proven. Quantum theory, on the other hand, posits that our world lacks certainty, and we cannot predict certain events, such as the speed and position of electrons, even though Einstein famously stated, “God does not play dice.” Fundamentally, our limitations have transcended the scope of merely predicting or understanding events within the physical realm. Even if we strive to construct a mathematical universe governed entirely by rules of our own design, this abstract universe will still harbor undeniable facts.
However, beyond mathematical statements, our world is also filled with philosophical statements describing reality, and we find ourselves unable to describe, fully express, understand, or even define these realities.
Similar to the uncertainty of the concept of “truth” in the early 20th century, definitions of other concepts such as “art,” “beauty,” and “life” also lack fundamental consensus. However, these are not isolated cases; other concepts like “wisdom” and “consciousness” are also caught in this dilemma.
To bridge this gap, Legg and Hutter proposed a definition of intelligence in 2017 in “Universal Intelligence: A Definition of Machine Intelligence”:“Intelligence is the measure of an agent’s ability to achieve goals in various environments.” Similarly, in “Problem-Solving and Intelligence,” Hambrick, Burgoyne, and Altmann argue that the ability to solve problems is not just an aspect or feature of intelligence but is the essence of intelligence. These two statements are similar in that achieving goals can be linked to problem-solving.
Gottfredson summarizes the views of several researchers in “An Editorial with 52 Signatories”:“Intelligence is a very general mental ability, including reasoning, planning, problem-solving, abstract thinking, understanding complex ideas, quick learning, and learning from experience.” It is not merely book knowledge, narrow academic skills, or test-taking tricks. Rather, it reflects a broader and deeper understanding of the surrounding environment; a capacity to “grasp” or “understand” things, or to “conceptualize” responses.
This definition introduces two key dimensions: the ability to learn from experience and to understand the surrounding environment, thus extending the concept of intelligence beyond mere “problem-solving skills.” In other words, we should not view intelligence merely as an abstract ability to find solutions to general problems but as a concrete ability to apply solutions learned from previous experiences to different situations that may arise in our environment.
This highlights the intrinsic connection between intelligence and learning. In “How We Learn,” Stanislas Dehaene defines learning as “the formation of a world model,” which means intelligence also requires the ability to understand our surrounding environment and build an internal model to describe them. Therefore,intelligence also needs the capacity to create world models, even if this capacity may not be comprehensive.
When discussing General AGI versus Narrow AI, we often emphasize the differences between them. Narrow AI (or weak AI) is very common and successful, often outperforming humans in specific tasks. A good example is when AlphaGo defeated the world champion Go player Lee Sedol 4-1 in 2016. However, an event in 2023 highlighted some limitations of Narrow AI. In a Go match, amateur player Kellin Perline won using tactics that AI failed to recognize.This illustrates that AI lacks the ability to recognize uncommon strategies and make corresponding adjustments, just like humans do.
In reality, at the most fundamental level, even inexperienced data scientists can understand that every machine learning model AI relies on, even the simplest models, needs to strike a balance between bias and variance. This means AI needs to learn from data to understand and generalize solutions, rather than merely memorizing them. Narrow AI utilizes the computational power and memory capacity of computers to generate complex models relatively easily based on a large amount of observed data. However, once conditions change slightly, these models often fail to generalize.
This is akin to proposing a theory of gravity that only works on Earth based on observational results, only to find that objects weigh much less on the Moon. If we use variables instead of numbers based on our knowledge of gravity theory, we would understand how to use the correct values to quickly predict the gravitational strength on each planet or satellite. However, if we only use numerical equations without symbols, we will not be able to correctly extend these equations to other celestial bodies without rewriting them.
In other words,AI may not truly “learn” but can only extract information or experiences. AI does not form a comprehensive world model but creates an outline..
The commonly understood definition of AGI is that an AI system can understand and reason at human levels or higher across multiple cognitive domains.This sharply contrasts with the current Narrow AI systems that specialize in specific tasks (like AlphaGo).AGI refers to an AI system that possesses comprehensive, human-level intelligence capable of traversing different domains of abstract thinking.
As mentioned earlier, this requires our ability to create a world model that is consistent with experience and allows for accurate hypotheses about predictions.
In line with the views of most AI researchers and authorities, achieving true AGI will still require years, although predictions about when it will appear vary widely. In the article “AGI Safety Literature Review,” Everitt, Lea, and Hutter mention:“We asked many researchers, and they believe AGI may emerge between 2040 and 2061, but there are significant differences in their guesses, with some believing it may never appear, while others think it could happen in the coming years.”In summary, it is certain that AGI has not yet emerged among us.
In a recent paper titled “Sparks of Artificial General Intelligence: Early Experiments with GPT-4,” Microsoft noted:
“We believe GPT-4 is part of a new generation of LLMs that exhibit more general intelligence than previous AI models. We discuss the continually improving capabilities and impacts of these models. We demonstrate that, in addition to being proficient in language, GPT-4 can solve novel and challenging tasks involving mathematics, coding, vision, medicine, law, psychology, etc., without requiring any special prompts. Furthermore, in all these tasks, GPT-4’s performance is very close to human levels and often far exceeds that of previous models like ChatGPT. Given GPT-4’s powerful capabilities, we believe there is reason to consider it a version of a near (but still incomplete) general intelligence (AGI) system.”
What’s the issue? Microsoft is a partner of OpenAI.
A New York Times article quoted Carnegie Mellon University professor Maarten Sap as saying, “This is an example of how some large companies use research paper formats for PR purposes.” Researcher and robotics entrepreneur Rodney Brooks emphasized in an interview with IEEE Spectrum thatwhen evaluating the capabilities of systems like ChatGPT, we often “mistake performance for ability.”
To put it another way, treating scores as a marker of ability, AI seems to only see a rough overview of the world without truly understanding the essence of the entire world.
AI faces a significant issue regarding its training data. Most models are trained solely on text and lack the ability to speak, hear, smell, or live in the real world. As I previously mentioned, this situation is somewhat akin to Plato’s Allegory of the Cave. In that story, people can only see the shadows on the cave wall and cannot directly experience the real world. Even if a world model could be created, their world would only be a purely textual one, syntactically correct but semantically incomplete. This environment lacks the “common sense” generated by direct perception, making it appear bland.
What Are the Major Limitations of Large Language Models?
Another highly controversial challenge faced by large language models (LLMs) like ChatGPT or GPT-4 is their propensity to generate hallucinations. Hallucinations refer to instances where these models fabricate false citations and facts, sometimes producing even nonsensical content. The reason for hallucinations lies in their lack of understanding of the causal relationships between events.
In the paper “Is ChatGPT a Good Causal Reasoner? A Comprehensive Evaluation,” the authors conclude: “ChatGPT has a serious problem with causal hallucinations; it tends to assume causal relationships between events regardless of whether those relationships actually exist.” They further point out: “ChatGPT is not a good causal reasoner but a good causal explainer,” emphasizing again its ability to extract connections while being unable to infer those connections by constructing an existing world model where those connections naturally exist. Although this paper focuses on ChatGPT, it can be extended to any LLMs.
Fundamentally, we can find thatLLMs excel at identifying and extracting causal relationships from data but lack the ability to independently reason about new causal scenarios. They possess the capacity for causal induction through observation but lack the capacity for causal deduction.
This distinction highlights a limitation: systems can identify causal patterns but lack abstract causal reasoning ability. It cannot generate new causal insights but merely interprets causal connections from data.
However, if intelligence requires learning from experience, and learning translates into creating a world model we can use to understand our surrounding environment, then causal deduction constitutes a fundamental element of learning and, thus, a fundamental element of intelligence, which is precisely what existing models lack. This is one of the key steps toward AGI.
As demonstrated in the early 20th century, the reality often differs from the intuitions formed by our everyday observations. Just as early 20th-century physicists struggled to understand quantum mechanics due to its contradiction with human intuition, the AI systems we are currently building are limited to a small portion of reality, even narrower than the range we humans can experience.
Just as we ultimately understood a real world that contradicts our daily experiences, we can only truly achieve AGI when we can create a system that can question its own reality, engage in self-exploration, and at least apply causal deduction to build a reasonable world model..
This prospect may mark a new phase in human history, as we gradually acknowledge that humanity’s significance in the universe is diminishing.
Compiled by: Yun Jing