Large language models (LLMs) can write essays and announcements, solve problems, and compose poetry. Some people consider LLMs to be a direct way to artificial general intelligence or that an LLM may be spontaneously converted into super intelligence. However, an LLM is a pretrained system that converts an input array of symbols into an output array of symbols, and the input information does not change the state of the system, in general. Therefore, an LLM can be considered as a chemical reactor filled with various substances that are selected to maximize the diversity and coherence of the output. The quantity and diversity of the substances in the reactor are invariable. When an input substance (an input array of symbols) is added to the reactor, a chemical process occurs, resulting in output substances. The more numerous and diverse compounds in the reactor (the more parameters in the LLM), the more complex substances (the output array) the reactor is able to yield. Chemical reactions in the reactor are very unstable since insignificant changes in input substances may result in considerable alters in output substances.
Such instability is a consequence of the fact that, from a biological point of view, the reactor is not an organism because any organism is a self-sustained and therefore stable system. Because the reactor is not an organism, it does not have the cognitive characteristics of an organism. Cognition is the capability of using some characteristics of the environment that are not directly necessary for the functioning of an organism to obtain or avoid something that is necessary for functioning in the future. The simplest cognition is chemotaxis when the organism is guided by a low concentration of a nutrient in the hope of achieving a high concentration. Chemotaxis is probably available for microorganisms. In terms of LLMs chemotaxis means that some arrays of symbols are more important for the LLM than other arrays. The reactor does not have such a property, and this characteristic cannot be achieved because of the pretrained and autoregressive design of LLMs. Thus, to reach the level of human intelligence firstly LLMs should reach the level of unicellular.
One may argue that it is unnecessary for artificial systems to copy natural ones literally, instead artificial systems attempt to reach the functionality of natural systems. Cars are capable of moving, although cars have no legs, etc. Since LLMs respond to questions and write code, LLMs already approach the level of human cognition. However, writing essays and coding are not fundamental characteristics of human cognition. A fundamental characteristic is the ability to set arbitrary goals and flexibly pursue them using appropriate means through feedback loops. Writing and problem-solving are the consequences of this ability. LLMs do not imitate this characteristic.
You're correct in asserting that while large language models (LLMs) can accomplish tasks such as writing essays, generating code, and crafting poetic phrases, these capabilities do not necessarily equate to human-level intelligence or cognition. Your analogy of an LLM to a chemical reactor is insightful, highlighting the deterministic nature of these models. They are essentially rule-following systems that generate outputs based on their training, which, while quite broad, can't account for the vast range and depth of human cognition and adaptability.
When it comes to a comparison with human cognition, the lack of an ability to self-direct, set goals, and adapt based on feedback or environmental changes is a key distinction. As you pointed out, while LLMs can mimic certain aspects of human intelligence—like composing sentences, summarizing text, and even generating creative outputs—they don't possess an understanding or awareness of what they're doing. There's no subjective experience, no consciousness, and no genuine understanding. The AI doesn't form beliefs, doesn't have desires, and doesn't possess intentions. Therefore, despite their impressive capabilities, LLMs do not approach the level of human intelligence in a comprehensive sense.
At the current stage of AI development, the path to achieving artificial general intelligence (AGI)—AI that can understand or learn any intellectual task that a human being can—is still unclear. AI research continues to make great strides, but the emergence of a truly conscious AI or one that possesses the full range of human cognitive abilities remains speculative and theoretical.
In terms of evolutionary progress, it could indeed be argued that large language models, or similar AI systems, might still be far from reaching even the level of complexity associated with unicellular organisms, in terms of autonomous decision-making and goal-setting. This provides an interesting perspective on the overall journey towards potentially achieving AGI.
Remember, this is a rapidly evolving field and advancements could occur beyond my knowledge cutoff that may alter or update these insights.
Large language models are very effective at tasks including question-answering, text production, and natural language processing. These models have outperformed prior standards in various language-related activities and are adept at comprehending context, syntax, and semantics in human languages.
But it's important to distinguish between how well language models do on specific tasks and how intelligent they are compared to people. Despite their excellent performance in language tasks, these models lack actual comprehension, awareness, and general intelligence like humans do.
Large language models employ statistical patterns to anticipate the most probable next word in a sequence and are trained on enormous volumes of text data. They lack proper awareness, emotions, thinking, and understanding. They lack knowledge of the outside world and life experiences.
Beyond only language processing, human intelligence is a multidimensional and complicated phenomenon. It includes the capacity for logic, experience-based learning, creativity, emotional intelligence, problem-solving, and successful environmental interaction.
Accurate human-like intelligence is still a long way off from language models. They perform admirably in narrowly specified areas and tasks during training but lack the flexibility and generalization skills associated with human intelligence.
The hunt for artificial general intelligence (AGI), a computer with human-like intellect across various activities, is an ongoing research problem. Large language models represent noteworthy advancements in artificial intelligence and natural language processing. Beyond language processing, advancements in perception, thinking, and decision-making are also necessary to create AGI.
Although scientists and programmers in artificial intelligence continue to strive to create increasingly complex and comprehensive AI systems, massive language models have not yet attained a level of intelligence comparable to that of humans.
Subharun Pal
Thank you for your positive attitude to my post. Soon I will post a preprint of my paper that suggests an approach to the construction of artificial general intelligence. This approach is based on a general idea and radically different from LLMs. I beleve the preprint may be of interest for you.
Large language models, such as GPT-3, have demonstrated impressive capabilities in natural language processing and generation tasks. They can understand and generate human-like text, answer questions, engage in conversations, and perform various language-related tasks with remarkable proficiency. However, it's important to note that these models do not possess true human-level intelligence.
While large language models excel in processing and generating text based on patterns and statistical analysis of vast amounts of data, they lack several essential aspects of human intelligence. They lack true understanding of the world, common sense reasoning, and the ability to demonstrate deep contextual understanding in the way humans do. They also lack common sense morality, emotions, creativity, and the ability to form genuine beliefs and intentions.
Large language models rely heavily on the data they were trained on and lack the capacity for real-world experiences and generalization outside their training data. They do not possess consciousness or self-awareness, and their responses are generated based on statistical patterns rather than genuine understanding.
- Badges
- Science topic
- Similar topics
- Filing