The Thoughts of AI

In the summer of 2024, AI skeptics were feeling good. GPT-5 had been delayed, again. Grok 3 was absent despite Elon Musk’s braggadocio, and Claude was doing only point releases: 3.5 had been released at the beginning of summer with no date on when version 4 might come. There was talk that AI pre-training, the training of AI using all the data on the internet, had reached its limit. Rumors about model collapse due to a lack of new human data and a focus on synthetic data produced by AIs themselves began to surface. Accelerationists and doomers both suddenly were looking for new hot takes to post to X.

Then, a small Chinese company called “Deepseek” posted a model in the Fall of 2024 that changed everything. The “Whale bros,” as they were called because of their whale logo, produced deepseek-r1, built on top of deepseek-v3, a respectable but unremarkable foundation model. However, deepseek-r1 established a new direction in AI research: reasoning models.

Transformer LLMs are trained on terabytes of human text. But when asked a question, they produce a response token by token through a forward-pass system that is largely unexplainable. While computer scientists understand the process in general terms, at the scale of something like ChatGPT or Claude, it is impossible to determine how the matrix of weights (called parameters) that the model largely consists of translates into any one token. Additionally, because the process is stochastic, the production of the token has a measure of randomness that makes it impossible to repeat consistently. This is why LLMs are often talked about as “black boxes.” 

The upshot of this is that when an LLM gives an answer, we do not know how it arrived at that answer. Was it copied from the training data? Or generalized from a variety of inputs? Or even completely made up? Any of those are possibilities. So, when an LLM answers a question, it is in some ways a mystery. And it is the very possibility that it is that last option -- completely made up -- (what is usually referred to as a hallucination) that is a grave problem for LLMs. 

Part of this is because there are essentially three parts to an LLM: the context window where the input is given (the question being asked), the attention mechanism (this is the model itself), and its output generation. The advance that occurs with a reasoning model is that there is now a fourth part added, what I will call the “Reasoning Space” (though there is not a consensus on what the right language is for this). This Reasoning Space is a place where the model constructs its answer, weighs various possibilities, examines its premises and conclusions, and comes up with the best answer. 

Reasoning models have created a new gold rush in AI work. While Deepseek-R1 was not the first reasoning model (OpenAI o1 preview was), it was the first open source model that employed reasoning training techniques which were nearly on par with existing AI tech companies models, AND it made public its training process and its Reasoning Space so that a user could see what exactly the model was thinking and how it came to its conclusions.

This offers the ability for users and developers to do something that has never been possible before: Determine how an AI comes to its answer. While the process of token prediction remains the same, since the answer goes through Reasoning Space before it is finalized, we have access to those “reasoning traces” (the internal monologue of the model) and can see the path to its final answer. The other exciting development about this is that in the Reasoning Space, the model can reject the raw results of token prediction and come to a different and often better conclusion.

I have been working on a project that has sought to construct individual AI instances that replicate the Jesus of the Gospels (and the gospel of Thomas). These are AIs that replicate, for all intents and purposes, the language and responses of the Jesus of Mark, Matthew, Luke, and John in hopes of putting them in dialogue with each other. The project I propose for presentation this year takes this a step further by integrating reasoning models into the mix. 

Now the task is not just to train a model so that it replicates Jesus's language and speech patterns but to do that with the reasoning model. This will allow us to see the reasoning process that is a part of these responses. How does the thinking of Mark’s Jesus differ from the thinking of Luke’s Jesus? What are the concerns and priorities evident in the thinking of the gospel of Thomas Jesus that are different from the Gospel of John’s Jesus? 

Beyond this, there is the question of whether we can identify particular tropes or concepts that are a part of the particular AI's reasoning. If this is possible, then the question is whether that thinking can be altered through conversation with the other Jesus AIs. 

Ultimately, the question of this project is: Can a reasoning AI Jesus learn when put into conversation with another AI Jesus? Long-term, if we see some success here, we would want to see what kinds of thinking changes come from being in dialogue with other AI, perhaps an AI Buddha or AI Richard Dawkins. 

To conclude, I believe this kind of experiment using AI may have very interesting results. AI’s have already shown usefulness in human simulation research. In the end, the use of AI reasoning for simulation may have even more interesting applications through interrogating the thoughts of AI.