Here are the 12 Key Takeaways from Episode 2 of the 11 Episode series on why no AI product could ever think, understand or perceive. Episode 2 sets out the problem which will be answered in the later episodes of the series (3-11).
Episode 2: What AI Really Is
Key Lesson Takeaways
1. An algorithm is a set of instructions which, if carried out correctly, will produce a specific result; and a computer program is an algorithm expressed in a language a computer can execute.
2. Every AI software application is essentially a type of computer program.
3. AI applications differ most markedly from traditional programs in their ability to mimic learning and improve over time--not by manual programming, but by adjusting their own parameters based on new data.
4. According to the expert majority consensus, there is currently no Artificial General Intelligence (AGI), only Narrow AI--though some insiders, like Sam Altman, believe they now understand how to build AGI, and Elon Musk predicts that by 2029 AGI will surpass the combined intelligence of all humans. Time will tell.
5. Many programs in general--and all large language models (LLMs)--use mathematical functions.
6. Current LLMs and many AI applications use Artificial Neural Networks, inspired by and designed to imitate certain broad features of the human brain.
7. ChatGPT, like any similar model, operates somewhat like a glorified autocomplete: it calculates the most probably next word to add to the input it's received.
8. ChatGPT models language by creating vectors (sets of numbers) whose relationships mirror the relationships between words in natural language.
9. To assign these vectors, ChatGPT uses statistical data on how frequently words occur near each other in its training data.
10. Thanks to how LLMs are trained, even their developers don't always fully understand why the models produce certain results--though, in principle, they could find out.
11. Conversational AI doesn't have to work the way our current LLMs do. AI is a broad field, and researchers are actively developing systems that more closely mimic human thinking and reasoning, rather than relying solely on statistical word relationships.
12. It's tempting to think AI truly thinks and understands. After all, AK now produces shockingly good simulations of human conversation and creativity; and artificial neural networks imitate some general features of the brain. This might suggest that AI has captured the essence of cognition. Si, if, as I will argue, no AI product could ever, truly, think, understand, or perceive, the question is, nonetheless, challenging. It is not obvious that AI cannot think!, given the striking similarities between AI and human abilities, and in how they operate.
This calls for sustained, careful, thought; that is, philosophy, and the insights of great philosophers, namely Aristotle and Saint Thomas Aquinas (Episodes 3-11).
