When humans and AI interact, I see that as a collaborative effort that consists of mutual respect. One way in which to achieve that is by looking at both the similarities as well as differences. This helps to find out who has more strength in one area, over the other. This is very common with humans, right? One person may be able to fix anything, but may lack computer skills. One is trained in law, the other medicine. If we group those different disciplines into a unified whole and expect them to work together, the first step may be to assign tasks or functions based on their strengths. During a meeting, for example, input is welcome by all, but the subject matter expert would take the lead. That, to me, makes perfect sense, so why not apply that to working with an AI? I hope to make that argument below by addressing types of learning, both human and AI.
First, let’s take a look at the human approach. Now I’m not a psychologies or a neuro-scientists, so I’ll speak in more general terms.
Humans learn, basically, from the following.
Experiential | Physical interaction, like touching a hot stove, playing a sport or instrument, etc. |
Academic | Symbolic language, like attending a lecture, reading a book, etc. |
Social | Observation, like learning from watching others. |
Associative | Patterns, consequences, like green means go. |
The first similarity to me is obvious. Data. Whether machine or human, data or input is required. The big difference is how humans and machines process the data.
Human vs AI – Head to Head
The differences in processing data can be viewed as follows:
Data source: | Reality vs Representation |
| Human | Learns through multiple ways (multi-modal). We don’t just learn about an apple, we imagine it’s color, texture, taste, etc. We are grounded in our physical and biological survival. |
| Generative AI | Pure representation of data. It has no concept of what an apple is, but through training, it can associate millions of lines of text and images to be able to represent an apple. |
| Efficiency | Small vs Big Data |
| Human | Incredibly data efficient. |
| Generative AI | Data hungry. Relies on massive statistical averages, rather than a handful of meaningful experiences. |
The Mechanism | Reasoning vs Prediction |
| Human | Cognitive schema’s, internal mental models. Allows us to reason outcome. |
| Generative AI | Statistical next-token prediction. It doesn’t reason. |
Intent | Purpose vs Optimization |
| Human | Internal goals. We want, or need to learn something. |
| Generative AI | Optimization problem. Punishment/reward during it’s training. Rewarded when it’s right, punished, so to speak, when it’s wrong. It doesn’t want to learn, rather it’s mathematically forced to minimize errors. |
Overall Process Flow
| Humans | Generative AI | |
| Input | Senses, emotions, logic, etc. | Text, code, images, tokens, etc. |
| Processor | Why. Building mental models. | What. Identifying statistical patterns. |
| Speed | Slow to learn facts, faster to learn concepts. | Memorizes instantly, slow to understand. |
| Context | Personal history. | Relies on data. |
| Error | Cognitive bias, or just forgot. | Hallucination. (see further below for more). |
I asked Google Gemini about this learning and the response was: “When you talk to a human, you are talking to a historian of life. When you talk to an AI, you are talking to a librarian of patterns.
Why this difference is important
| Humans | Generative AI | |
| The Spark | Curiosity. | Math ( attempts to limit mistakes in a guess using an algorithm). |
| Memory | Selective. Tend to remember the emotional or important things. | Total. Data is stored with equal weights until instructed not to. |
| Logic | Common sense. | Correlation within it’s data. |
The Hallucination Factor
Humans will say, I don’t know, when they cannot recall or visualize the answer. Generative AI is designed to predict the most likely result through probabilities, even if and when it’s way off the mark.
Conclusion
Both humans and AI learn. They both need data. An AI is much more efficient in it’s recall because it doesn’t forget anything, unless the data has been pulled. Imagine sitting in a meeting with an AI that remembers every project, every success, every failure as well as all costs, etc. That would be a HUGE benefit, at least I think so. The scenario that currently comes to mind for me is how AI is rolled out. The AI can look back at the creation and roll out of social media as an example. It would understand all the pro’s and con’s of that roll out and then possibly suggest an approach. Social media was a technical problem to be solved. Generative AI is much more than that. We need to tread carefully. Having just read that, I think the next article will be on inference. This is very fascinating. If you read above carefully, there is no data for an AI to learn how an AI needs or should be rolled out, right? We have very limited knowledge of that. So the AI, like a human, needs to infer. I’ll write about that next.
