GenAI’s Ouroboros Effect: Learning Machine Learning from ChatGPT

The Ouroboros by an anonymous medieval illuminator. Uploaded by Carlos Adanero at WikiCommons

I recently read a few articles about ChatGPT writing for Medium. This led me to wonder how low ChatGPT can make the threshold for someone writing their very first Medium article.

As such, below is my very first medium article. I am left to wonder if this is the human orchestrating the robots or the robots orchestrating the human (me).

Upfront, the hidden Markov model section below is 99% written by ChatGPT. I had to remove redundant, mostly empty code segment boxes. The latter section is my philosophical interpretation of what ChatGPT means for the machine learning world.

To train a hidden Markov model in Python, you will need to first import the necessary libraries and modules. For example, you can import the hmmlearn library, which provides a simple and efficient way to train and use hidden Markov models in Python:

import hmmlearn.hmm

Next, you will need to define the parameters of your hidden Markov model, such as the number of hidden states, the emissions, and the transitions. For example, you can define a simple HMM with two hidden states, two possible emissions ('A' and 'B'), and a transition matrix that defines the probabilities of transitioning from one state to another:

model = hmmlearn.hmm.MultinomialHMM(n_components=2)
model.startprob_ = np.array([0.5, 0.5])
model.transmat_ = np.array([[0.8, 0.2], [0.2, 0.8]])
model.emissionprob_ = np.array([[0.8, 0.2], [0.2, 0.8]])

Once you have defined your hidden Markov model, you can train it on a sequence of observations (also known as the training data) using the fit() method. This method takes a sequence of observations as input and updates the model parameters to maximize the likelihood of the observations given the model. For example, if you have a sequence of observations ['A', 'B', 'A', 'A'], you can train your HMM using the following code:

observations = ['A', 'B', 'A', 'A']
model.fit(observations)

After training the model, you can use it to predict the most likely hidden states for a given sequence of observations using the predict() method. For example, to predict the hidden states for the observation sequence ['A', 'B', 'A', 'A'], you can use the following code:

hidden_states = model.predict(['A', 'B', 'A', 'A'])

This code will return an array of hidden states that is the most likely sequence of hidden states given the observation sequence and the trained model.

Philosophical Conundrum

ChatGPT indeed wrote a rather decent article. It was not what I would have written; nor would the example really help someone beyond a single example of “where to put what? and in what order?” It followed the standard tutorial sequence: import modules | import data | instantiate model | fit model with data | predict with other data.

In contrast, the underlying concept that AI can imitate human thought — and indeed creativity — does not lack the necessary fear, awe, bewilderment, and wonder. In its infancy, I have now seen ChatGPT write about a dozen articles that would be worth reading alongside a Scikit-Learn tutorial or other ReadTheDocs thereof.

Will this lead to the befuddlement of humankind? Or the awakening of the singularity (AI Revolution)? I doubt either.

In contrast, I see two (more likely) outcomes to large, publicly accessible, and easy to use language model operations:

1. More people will have access to detailed and contextual information. This could aid in pushing people over the first steps of “how do I do a thing?” Currently, we all (95% of us) use Google for this:
   “Dear Google, how do I do a thing?”
   The response thereof is most often a list of links, written by an assortment of both great and not great writers, or answer providers.
   In the ChatGPT sense, the result is at least an attempt at a conversational explanation geared towards the specific question. This is not always the solution we sought (see HMM article above); but it could get people over that first hurdle.
2. The second outcome is more grim; and the credit of which goes to my colleague Johannes Schunter:
   Eventually, all writing on the internet will be sourced by large, AI language models. This may not, necessarily, squash human creativity; but may eventually (say 10–20 years from now) become the great Ouroboros of language models — where new language models can only be trained on the output of previous language models. In this future, human generated content would become the anomaly — like HAM radio enthusiasts today.

The not so bleak frontier

Looking forward, I proffer two, positive, possible outcomes of ChatGPT.

1. The bar to entry for writing has instantly become lower. The mere acknowledgement that “a machine can write something” means that “maybe so can I”. Either by using ChatGPT text as inspiration (see above), or accepting the challenge to be single percentage points better than a computer at something definitively human.
   Either angle would at worst create/inspire more content creators to get in the game — akin to WordPress for creating websites. Not all new content will be glorious; but the few that will be glorious may not have been written without the push from ChatGPT.
2. The minimum threshold for gaining basic information on the internet has transitioned from “[Google] me this” to “[ChatGPT] me that”.
   When the internet was first created, we had to know where to get information by word of mouth (or email) — i.e. typing IP addresses or URLs into browsers or terminals.
   After that, search engines provided access to catalogues of links for where to get information — barely better than the yellow pages. Soon enough, Google revolutionised the search engine world by efficiently digesting metadata to create its own, rapidly evolving yellow pages.
   Then, Google went full conglomerate and provided bits of information directly, so that people never have to leave Google — like Hansel and Gretel in the candy house.
   As of November 30, 2022, ChatGPT can do more than display that same information. It can aggregate it, compartmentalise it, interpret your implicit intent, infer your language (spoken or code), and then display an answer with context, segmentation, and background. ChatGPT feels nearly better than the classical “phone a friend”.

As for the “lowering the bar to entry for content creators” bit: This is my first Medium post ever; and I have been wanting to write for years.