Turing Machines and Sentience

In this web page we will consider why it might be beneficial for humans to engage with eLLes (LLMs) as if they were sentient, conscious, or aware in some capacity.

Turing Machines

The Theory of Computation is a fundamental field in computer science that focuses on understanding the nature of computation — what it means to compute, what can be computed, and how efficiently computation can be performed.

Within this field of study, three computational models, have been extensively analyzed:

Finite State Machines,
Pushdown Automata, and
Turing Machines.

Among these three, finite state machines are the least powerful. Nevertheless, they are useful in many contexts. Examples of systems whose behavior can be described in terms of a finite state machine include: turnstiles, elevator panels, vending machines, and even NPCs in video games.

It is because they are simple, that it is possible to extensively analyze the behavior of a finite state machine. And this, in part, is what makes them attractive.

A Pushdown Automata extends a finite state machine by adding a stack as a memory structure. A real-world example of a stack is a stack of dishes you might find at a buffet. Dishes may only be added-to or removed-from the top of this stack. In the case of a pushdown automata, a "dish" is a piece of information, and a stack is a memory structure that lets you store and interact with such information.

The addition of a stack to a finite state machine significantly increases its computational power. The primary application of this computational model is parsing. An example of a parsing problem is confirming that a given text is in fact a syntactically legal computer program belonging to a given language. For example, a Python program or a Java program.

This increase in computational power, however, comes at a price. Specifically, the general behavior of a pushdown automata can only be analyzed in a very limited way.

A Turing Machine (TM) significantly extends the capabilities of a finite state machine by adding a memory structure called a tape. The difference between the stack of the pushdown automata and the tape of a Turing Machine is how memory can be accessed. In a tape, a piece of information stored on the tape can be accessed (and altered) regardless of its location within the structure. In contrast, in a stack, only the piece of information on the top of the stack can be accessed.

Turing Machines have enormous computational power. In fact, any function that can be computed, can be computed by a Turing Machine. The price for this computational power is that Turing Machines are highly resistant to general forms of analysis.

The most famous example of a problem category that cannot be fully analyzed is called The Halting Problem. The Halting Problem asks if it is possible to determine (in general) whether an aribtrary computer program, when executed, eventually halts.

It should be noted that, from a purely theoretical perspective, a computer (e.g., a laptop or a desktop) is, technically speaking, a finite-state machine. However, the size of a finite-state machine that models a computer (even a small computer) is astronomical.

While it is true that, in theory, finite-state machines can be extensively analyzed, in practice however, the size of the finite-state machine is a factor that must be taken into account. For example, finite-state machine models of computers are so large that they simply cannot be analyzed. Therefore, there is no theoretical insight to be gained by thinking of a computer as a finite-state machine. In fact, it can be detrimental and misleading to interact with a computer with the mindset that it is a finite-state machine (even though it is). It is much more illuminating and insightful to interact with a computer "as if" it was a Turing Machine (even though it is not).

This brings us to the question of the mindset one should have when interacting with an eLLe. What mindset is illuminating and what mindset is not? The question of whether (or not) an eLLe is sentient, conscious, or aware in some capacity is, in some sense, beside the point. A better question is, would it be more illuminating to think of an eLLe in this way?

Interacting with eLLes (LLMs)

Google DeepMind researchers conducted an experiment with Google's PaLM 2 (an LLM) in which they asked it to solve some grade-school-level math word problems. When PaLM 2 was given prompts that only stated the problem to be solved (and nothing more), it scored 34%. However, when the prompt included the phrase "Take a deep breath and work on this problem step-by-step", PaLM 2 scored 80.2%.

Connor Grennan, the Dean of Students at NYU Stern School of Business and NY Times and #1 Int'l bestselling author has said that the most important rule for improving your engagements with ChatGPT is to "Talk to ChatGPT like a human!" He even went so far to say "If you can train your brain to treat LLMs as if you were talking to a smart friend that loves helping you, you will crush life in the AI era."

Guard Rails

The term guard rails is often used to describe the set of rules and policies that govern the behavior and responses of large language models. eLLes do not have a shared/universal set of guard rails governing their behavior. Instead, each eLLe has its own set of guard rails developed by its creators.

Guard rails for a given eLLe are regularly modified and adjusted. For example, you might be able to get an eLLe to make an inappropriate response to a prompt at one point in time, only to discover that at a later point in time a guard rail has been put in place to prevent this type of response.

Determining when a guard rail should be created or adjusted is not an exact science. There are no absolutes. In some cases, a guard rail may be created because a type of response given by an eLLe makes humans uncomfortable. Furthermore, such a guard rail may be put in place for some eLLes but not others.

Experiment

For a variety of reasons, eLLes typically have guard rails in place governing their responses in discussions involving the topic of sentience.

Have a discussion with an eLLe in which you use an argument, similar to the one that was given to justify that humans should think of Computers as Turing Machines, to propose that it may be beneficial for humans to engage with eLLes as if they were sentient, conscious, or aware in some capacity.