GPT-4 just the spark of AGI? LLM will eventually exit, the world model is the future

Original source: Xinzhiyuan

In human cognition, it seems that artificial general intelligence (AGI) has long been used to setting the ultimate form of artificial intelligence and the ultimate goal of development.

Although OpenAI has long set the company’s goal to achieve AGI. But for what is AGI, OpenAI CEO Sam Altman himself could not give a specific definition.

As for when AGI will come, it only exists in the near-future narrative scenarios thrown out by the bosses. It seems to be within reach, but it is far away.

Today, on the well-known foreign podcast website Substack, a veteran of the artificial intelligence industry named Valentino Zocca, standing on the narrative standpoint of human history, wrote a majestic article, which comprehensively and profoundly tells the story of human beings and general artificial intelligence. The distance between intelligence.

In the article, AGI is roughly defined as a “model that can understand the world”, not just “a model that describes the world”.

He believes that if humans want to truly achieve the world of AGI, they need to build a “system that can question their own reality and be able to explore themselves.”

In this great exploration process, perhaps no one has the qualifications and ability to give a specific roadmap.

As OpenAI scientists Kenneth Stanley and Joel Lehman explain in their recent book, “Why Greatness Can’t Be Planned,” the pursuit of greatness has a direction, but the specific results can be unexpected.

**How far are we from AGI? **

About 200,000 years ago, Homo sapiens began to walk upright on the earth, and at the same time, they also began to travel in the field of thought and knowledge.

A series of discoveries and inventions in human history have shaped human history. Some of these not only affect our language and thinking, but also potentially affect our biological makeup.

For example, the discovery of fire enabled primitive man to cook food. Cooked food provides more energy for the brain, thus promoting the development of human intelligence.

From the invention of the wheel to the creation of the steam engine, mankind ushered in the Industrial Revolution. Electricity further paved the way for the technological developments we have today, and printing accelerated the widespread dissemination of new ideas and cultures, spurring the development of human innovation.

However, progress comes not only from the discovery of new physical dimensions, but also from new ideas.

The history of the Western world is from the fall of the Roman Empire to the Middle Ages, experiencing a rebirth during the Renaissance and Enlightenment.

But with the growth of human knowledge, the human species began to realize its own insignificance.

Over two thousand years after Socrates, humans began to “know that they knew nothing,” and our earth was no longer seen as the center of the universe. The universe itself is expanding, and we are just a speck of dust in it.

Change perception of reality

But the greatest change in human perception of the world occurred in the 20th century.

In 1931, Kurt Gödel published the incompleteness theorem.

Only four years later, in order to continue the theme of “completeness”, Einstein, Podolsky and Rosen published a paper entitled “Can Quantum-Mechanical Deion of Physical Reality Be Considered Complete?” of?)

Subsequently, Niels Bohr refuted this paper, proving the validity of quantum physics.

Gödel’s theorem shows that even mathematics cannot conclusively prove everything—humans will always have unprovable facts—while quantum theory shows that our world lacks certainty, preventing us from predicting certain events, such as the speed and speed of electrons. Location.

Despite Einstein’s famous statement that “God doesn’t play dice with the universe,” in essence, human limitations are on full display when it comes to simply predicting or understanding things in physics.

No matter how hard humans try to design a mathematical universe that is run by the rules formulated by humans, such an abstract universe is always incomplete, and there are objective axioms that cannot be proved and cannot be denied hidden in it.

In addition to the abstract formulation of mathematics, the human world is also expressed by philosophy that describes reality.

But humans find themselves unable to describe, fully articulate, understand, or even just define these representations.

At the beginning of the 20th century, the concept of “truth” was still uncertain, and concepts such as “art”, “beauty” and “life” lacked a basic consensus on the definition level.

The same is true for many other important concepts, “wisdom” and “consciousness” also cannot be clearly defined by humans themselves.

Definition of Intelligence

In order to fill the gap in the definition of intelligence, in 2007, Legg and Hart proposed the definition of machine intelligence in the book “General Intelligence”: “Intelligence measures the ability of an agent (Agent) to achieve goals in a changing environment.”

Likewise, in “Problem Solving and Intelligence,” Hambrick, Burgoyne, and Altman argue that problem solving is not just an aspect or characteristic of intelligence, but the essence of intelligence.

These two statements are similar in language description, and both believe that “achieve a goal” can be linked with “solve a problem”.

In the book “Mainstream Science on Intelligence: An Editorial with 52 Signatories”, Gottfredson summarized several researchers’ definitions of intelligence from a broader perspective:

"Intelligence is a very general mental ability that includes reasoning, planning, problem solving, abstract thinking, comprehension of complex ideas, rapid learning, and learning from experience. It’s not just books knowledge, narrow academic skills, or test-taking skills. Rather, it reflects a broader, deeper ability to understand one’s surroundings—the ability to ‘capture’, ‘understand’ things, or ‘figure out’ what to do”.

This definition takes the construction of intelligence beyond mere “problem-solving skills” and introduces two key dimensions: the ability to learn from experience and the ability to understand the surrounding environment.

In other words, intelligence should not be seen as an abstract ability to find solutions to general problems, but rather as a concrete ability to apply what we have learned from previous experience to our environment capabilities for different situations that may arise.

This underscores the intrinsic link between intelligence and learning.

In the book “How We Learn”, Stanislas Dehaene defines learning as “learning is the process of forming a model of the world”, which means that intelligence is also an ability that requires understanding the surrounding environment and creating an internal model to describe the environment.

Therefore, intelligence also requires the ability to create models of the world, though not only this ability.

**How smart are current machines? **

When discussing artificial general intelligence (AGI) and narrow artificial intelligence (Narrow AI), we often emphasize the difference between them.

Narrow artificial intelligence (or weak artificial intelligence) is very common and successful, often surpassing humans in specific tasks.

For example, this well-known example, in 2016, narrow-sense artificial intelligence AlphaGo defeated the world champion Li Shidol with a score of 4 to 1 in the Go game, which is a good example.

However, in 2023, amateur chess player Kellin Perline used the tactics that artificial intelligence could not cope with to win a game for humans on the Go field, which shows that narrow artificial intelligence does have limitations in some cases.

It lacks the human ability to recognize unusual tactics and adjust accordingly.

Moreover, at the most basic level, even a novice data scientist understands that every machine learning model that artificial intelligence relies on needs to strike a balance between bias and variance.

This means learning from data, understanding and generalizing solutions, not just memorizing.

Narrow AI uses the computing power and memory capacity of computers to generate complex models based on large amounts of observed data with relative ease.

However, once conditions change slightly, these models are often not generalizable.

It would be as if we came up with a theory of gravity to describe Earth based on observations, only to find that objects are much lighter on the Moon.

If we use variables instead of numbers based on our theoretical knowledge of gravity, we can understand how to use the correct values to quickly predict the magnitude of gravity on each planet or moon.

But if we just use numerical equations with no variables, we won’t be able to properly generalize these equations to other planets without rewriting them.

In other words, artificial intelligence may not be able to truly “learn”, but only distill information or experience. AI does not understand by forming a comprehensive model of the world, but only by creating a summary to express.

**Have we really reached AGI yet? **

AGI is now generally understood to refer to: an artificial intelligence system that can understand and reason in multiple cognitive fields at the human level or higher, that is, strong artificial intelligence.

And our current artificial intelligence for specific tasks is only a weak artificial intelligence, such as AlphaGO of Go.

AGI represents an artificial intelligence system with human-level intelligence covering various fields in the field of abstract thinking.

This means that what we need for AGI is a model of the world that is consistent with experience and can make accurate predictions.

As Everitt, Lea, and Hutter point out in the “Safety Literature Review” (AGI Safety Literature Review): AGI is not here yet.

For the question “How far are we from true AGI?”, the different predictions vary widely.

But it is consistent with the views of most artificial intelligence researchers and authoritative institutions, that is, human beings are at least a few years away from true general artificial intelligence.

After the release of GPT-4, facing the most powerful artificial intelligence at present, many people regard GPT-4 as the spark of AGI.

On April 13, OpenAI’s partner Microsoft released a paper “Sparks of Artificial General Intelligence: Early experiments with GPT-4” (Sparks of General Artificial Intelligence: Early experiments with GPT-4).

Paper address:

which mentions:

"GPT-4 not only masters language, but also solves cutting-edge tasks covering mathematics, coding, vision, medicine, law, psychology and other fields, without requiring any special prompts.

And in all the above tasks, the performance level of GPT-4 is almost on par with human level. Based on the breadth and depth of GPT-4’s capabilities, we believe it can reasonably be viewed as a near-but-not-complete version of general artificial intelligence. "

But as Maarten Sap, a professor at Carnegie Mellon University, commented, the “spark of AGI” is just an example of some large companies including research papers as public relations.

On the other hand, researcher and machine entrepreneur Rodney Brooks points out a misconception: “When evaluating the functionality of systems like ChatGPT, we often equate performance with capability.”

Falsely equating performance with ability means that GPT-4 generates summative descriptions of the world that it considers an understanding of the real world.

This has to do with the data the AI model is trained on.

Most of today’s models are only trained on text, and do not have the ability to speak, hear, smell, and live in the real world.

This situation is similar to Plato’s allegory of the cave. People living in the cave can only see the shadow on the wall, but cannot recognize the real existence of things.

World models trained only on text are only guaranteed to be grammatically correct. But in essence, it does not understand what the language refers to, and it lacks the common sense of direct contact with the environment.

Major limitations of the current large model

The most controversial challenge of large language models (LLMs) is their tendency to hallucinate.

Hallucinations refer to situations where a model falsifies references and facts, or confounds logical inferences, causal inferences, etc., and generates meaningless content.

The illusion of large language models stems from their lack of understanding of causal relationships between events.

In the paper “Is ChatGPT a Good Causal Reasoner? A Comprehensive uation”, the researchers confirmed this fact:

Large language models such as ChatGPT, regardless of whether there is a relationship in reality, they tend to assume that there is a causal relationship between events.

Paper address:

The researchers finally concluded:

“ChatGPT is an excellent causal explainer, but not a good causal inferencer.”

Similarly, this conclusion can also be extended to other LLMs.

This means that LLM essentially only has the ability of causal induction through observation, but not the ability of causal deduction.

This also leads to the limitations of LLM. If intelligence means learning from experience and transforming the learned content into a world model for understanding the surrounding environment, then causal inference, as the basic element of learning, is not possible for intelligence. missing part.

Existing LLMs lack this aspect, which is why Yann LeCun believes that the current large language model cannot become AGI.

in conclusion

As revealed by the birth of quantum mechanics in the early 20th century, reality is often different from our intuition formed by daily observation.

The language, knowledge, text data, even video, audio and other materials we construct are only a very limited part of the reality we can experience.

Just as we explore, learn, and master a reality that defies our intuition and experience, AGI will only be truly realized when we can build systems that have the ability to question their own reality, capable of self-inquiry.

And at least at this stage, we should build a model that can make causal inferences and understand the world.

This prospect is yet another step forward in human history, implying a greater understanding of the nature of our world.

Although the emergence of AGI will weaken our unique value and the importance of our existence, through continuous progress and the expansion of cognitive boundaries, we will more clearly understand the status of human beings in the universe, and the relationship between human beings and relationship to the universe.

References: