1

In the current applied ethics debate, a serious push to anthropomorphize or understand AI in human terms is underway. What many scholars point to is how LLMs allow AI to replicate and respond to our psychic and emotional states of consciousness. But this leads to the question: can we conclude that these are acts of simulation or mimicry? Do we model for imitation or simulation?

There appear to be no clear answers, but the goal should be to provide an organic cosmology rather than a mechanistic one.

6
  • 1
    In order to prompt clear answers from the community, could you please explain how you distinguish the concepts "imitation" and "simulation" - and what is meant by "mimicking"? - Why do you speak about "cosmology" in your last sentence? Notably, what does mean "organic" cosmology? Thanks.
    – Jo Wehler
    Feb 26 at 21:01
  • @JoWehler agreed
    – 8Mad0Manc8
    Feb 27 at 6:28
  • "What many scholars point to is how LLMs allow AI to replicate and respond to our psychic and emotional states of consciousness. " Any "scholar" who says such a thing doesn't know what an LLM is. LLMs are statistical autocompleters.
    – user4894
    Feb 28 at 0:58
  • We rush to this reductionist account and hastily conclude that LLMs are simply spewing out information from a mega database. But one will become trapped in this overstatement and will deny all evidence to the contrary. This is bad science and metaphysics. Feb 29 at 3:07
  • I’m voting to close this question because it's a question about Engish definitions.
    – J D
    Feb 29 at 12:40

4 Answers 4

1

Why do you think it is either? An LLM can in seconds write thousands of words on virtually any topic you care to mention- which human is that mimicking or simulating? It's not a capability of any human I know.

1
  • 1
    +1 And why do you deny it's both?!? On a good day, sometimes bots fool me... ; )
    – J D
    Feb 26 at 20:35
1

a serious push to anthropomorphize or understand AI in human terms is underway

I would say, as humans our ONLY why to understand anything is "in human terms". That human terms are quite impressive but still limited.

As Wittgenstein famously said:

If a lion could speak, we could not understand him’ because a lion’s form of life is so alien to ours that we cannot seriously claim to know what the lion means by what the lion says.

I think this is the key concept to understand AI. To understand it as an alien form of intelligence.

The anthropomorphization creates this false discourse where the one side claims that there is something human there and the other side is busy explaining why it can not be.

LLMs and other AI systems are already extremely impressive in what they can do but currently still limited. (Most likely this will change within a few years or more likely month).

Reading a humans emotional state from facial micro-expressions (thing "lie to me") and from the sound of their voice and from the language they use in chat is already something that these system are good at and surpass humans.

Simulating human emotions is also what these systems can do and at some point will likely surpass what human simulation of emotions ("acting") can do. Even without there being a single bit of "real" human emotion there.

Now all this is used by the critiques of AI to show the supposed limitations of the systems where the real challenge is to see them as this alien form of intelligence that inevitably will turn our world upside down, one way or another.

And when talking about "imitation" one should take a step back and think about how much of human live is imitation. How we are trained from birth to mimic the behavior of others, how we are trained to conform to social norms. How many people go through live "acting" to be someone they want to appear as, or that they think that be expected to be.

We live in exciting times, where actually for the first time in history the science of philosophy might be one of the most important ones to shape our future.

1

Artificial Neural Networks

https://deepgram.com/ai-glossary/perceptron

The Perceptron stands as a foundational algorithm in the realm of artificial intelligence, offering a first glimpse into the potential of computational models to mimic biological neural networks. Introduced by Frank Rosenblatt in 1958, the Perceptron was conceived at the Cornell Aeronautical Laboratory. Its development emerged from the aspiration to design machines that could automatically learn from experience, somewhat mirroring human cognitive processes.

The Perceptron is often considered the most basic unit when discussing neural networks. Just as a biological neuron receives signals, processes them, and produces an output, so does a Perceptron. While individual Perceptrons are limited to linearly separable tasks, their true power becomes evident when they are interconnected in multi-layer architectures. This structure, known as the Multi-layer Perceptron (MLP), forms the basis of many modern neural networks and was a precursor to more advanced deep learning models. The Perceptron’s foundational role in laying the groundwork for subsequent breakthroughs in artificial neural network architectures is undeniable in this context.

Sentiment analysis stands as a cornerstone of Natural Language Processing, functioning to discern the underlying sentiment or emotional tone enveloped within textual data. Its applications are varied and encompass domains such as gauging public sentiments on contemporary issues, vigilantly monitoring the reputation of brands in real-time, and deriving insights from customer feedback.

In the initial phases of sentiment analysis, single-layer Perceptrons were predominantly deployed, especially when the analysis was binary – typically distinguishing between positive and negative sentiments. Given a labeled dataset comprising both positive and negative reviews, a Perceptron would embark on its training journey. The objective was clear: proficiently categorizing incoming reviews as positive or negative. The model would extract features from the text, focusing on indicators like specific words tinged with sentiment, the frequency of particular terms, or even seemingly unrelated factors like the length of the review. Over time and with adequate training, the Perceptron would fine-tune its weights, enhancing its ability to discern sentiments in unseen data accurately.

However, as the domain of sentiment analysis evolved, it became evident that sentiments were not strictly binary. Textual data could convey a spectrum of emotions, from starkly positive or negative to neutral, or even a blend of multiple sentiments. Catering to this complexity required a more sophisticated model, ushering in the era of Multi-layer Perceptrons for sentiment analysis. MLPs, with their inherent ability to tackle multi-class problems, emerged as an apt choice for this refined level of sentiment analysis. These models would undergo training on datasets meticulously labeled with a multitude of sentiment classes. The MLP adjusted its weights through the training process, aiming to master the art of predicting a broad range of sentiments for fresh, unseen data. Their layered architecture and non-linear processing capabilities enabled them to capture the nuances in sentiment that a single-layer Perceptron might overlook.

Perceptrons and MLPs played crucial roles in the early days of neural-based NLP. While more advanced architectures have largely superseded them for many NLP tasks, their foundational contributions to the field’s progress are undeniable. Their ability to learn from data and make predictions, even if somewhat rudimentary compared to today’s standards, marked the beginning of a shift from rule-based NLP systems to data-driven, neural ones, paving the way for the contemporary NLP landscape we’re familiar with today.

We think apes mimic other apes using mirror neurons. But do we really think artificial neurons mimic biological neurons? This is a semantic problem of how to describe the effort to build artificial neural networks and learning machines. The article seems to be a very good high level discussion of the evolution of machine learning.

1

Let's agree on the terms first.

Mimicry is the act of copying or imitating the behavior of another. Simulation is the act of creating a model of a system or process.

In the case of AI, I think that it is doing both mimicking and simulation. AI can be trained to mimic human behaviour by learning from data sets of human conversations and interactions. However, AI can also be used to create models of human behavior, which can then be used to simulate humans.

1
  • +1 Nice distinction.
    – J D
    Feb 27 at 18:24

Not the answer you're looking for? Browse other questions tagged .