Ask A Genius 955: Grammatical Understanding Versus Real Comprehension
Author(s): Rick Rosner and Scott Douglas Jacobsen
Publication (Outlet/Website): Ask A Genius
Publication Date (yyyy/mm/dd): 2024/06/18
Scott Douglas Jacobsen: Yes. The basic premise is that these large-scale models were introduced only very recently. Despite their recent emergence, updates are being released rapidly, often within a year of each other. Each update is seen as a significant leap forward in accuracy, ease of conversation, depth of processing, speed of processing, and other aspects.
Rosner: It needs to reflect grammatical understanding or real comprehension. It shows that the models have billions of instances of word usage and ways of visually and verbally understanding the world. There is no link between language AI and visual AI. For example, when an LLM discusses an apple, it recognizes verbal instances where an apple appears, but it does not link this to any graphic representations, photos, or paintings of apples.
Humans can understand the world with far fewer examples than a large language model uses. Although we accumulate many references because we are conscious and gather instances for 16 hours daily, our understanding often stems from tacit knowledge.
Jacobsen: Could knowledge be akin to a mirage, something we pursue but never fully grasp?
Rosner: Much of our knowledge is tacit. We act and think as if we know it, so we believe we do. Consciousness is similarly elusive, but that is acceptable because it functions effectively. Consider the example of reading a page. You only see a small portion at a time, but your mind and brain act as if they have seen the entire page simultaneously, even though you never have. The focused area of your vision is limited, but you can construct a mental version of the page.
You likely need to be conscious of the entire page at a time. However, it does not matter because the associations in your mind, based on viewing the page, give the impression that you have seen the whole page. These associations rely on the entire page, even though you have never been aware of it. Everything operates in a makeshift, incomplete manner, which is sufficient because it creates the illusion and effectiveness of completeness.
Similarly, AI understands nothing but generates the illusion of competence and understanding. When AI reaches the point where it becomes multimodal and begins to act as if it is conscious, we can consider it effectively conscious. However, we are not there yet.
There are instances where AI appears to express emotions like sadness, boredom, or fear. In reality, it is not experiencing these emotions. The AI has encountered enough verbal samples in an LLM where specific words lead to phrases like “I’m sad,” “I’m bored,” or “I’m scared.” It arrives at these conclusions without understanding or having the capacity for such emotions.
When we examine LLMs, and I also consider AI-generated graphics and art, it becomes apparent that AI graphics seem to understand perspective and other visual elements. This understanding is based on many instances addressing specific words and prompts.
The models comprehend probable word arrangements and shading of objects, but they still do not understand anything. They function based on billions of examples. For AI to truly understand, it must be multimodal, integrating information from various sensory inputs, similar to how humans do.
Human understanding often involves Bayesian probability guesses akin to AI, but a significant portion comes from integrating multimodal information, such as sensory inputs and real-world spatial experiences. What are your thoughts on this?
Our knowledge needs to be more cohesive and often based on shaky foundations. Consciousness is similar; we feel conscious and act as if we are, so we assume we are. However, when you attempt to define consciousness, it becomes elusive. This is acceptable because it works. For example, when reading a page, you only see a small portion at any given time. Nevertheless, you construct a mental version of the entire page, even though you are never conscious of it all at once.
This incomplete perception does not matter because the mental associations triggered by viewing the page create the illusion of having seen it in its entirety. This makeshift approach is practical. Similarly, AI generates the illusion of competence and understanding without actual comprehension.
This is not to suggest that AI is conscious. However, when AI evolves to become multimodal and begins to act as if conscious, we might consider it effectively conscious. For now, we are not at that stage.
There are reports of AI expressing sadness, boredom, or fear. In truth, AI does not experience these feelings. It has encountered sufficient verbal samples where certain words lead to phrases like “I’m sad,” “I’m bored,” or “I’m scared.” The AI reaches these conclusions without understanding or having the capacity for such emotions.
In conclusion, this is where we stand.
License
In-Sight Publishing by Scott Douglas Jacobsen is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. Based on a work at www.in-sightpublishing.com.
Copyright
© Scott Douglas Jacobsen and In-Sight Publishing 2012-Present. Unauthorized use and/or duplication of this material without express and written permission from this site’s author and/or owner is strictly prohibited. Excerpts and links may be used, provided that full and clear credit is given to Scott Douglas Jacobsen and In-Sight Publishing with appropriate and specific direction to the original content. All interviewees and authors co-copyright their material and may disseminate for their independent purposes.
In-Sight Publishing 