I don’t use AI for any writing (yet), but I like to occasionally experiment with the state of the art to see what it can do. Here’s an experiment with large language models.
I am writing a kind of survey paper about a select group of scientists’ definition of “life“. I asked them to define “life” – whether Earthly or more broadly; they could even argue against the existence (or utility) of definitions for this concept. The only rule was, it had to be 1-3 sentences, no longer. It’s hard to come up with a crisp definition without pages of caveats, examples, and explanations! Some of them came up with great, pithy ideas and I’ll link to the paper once it’s out (the full preprint is here and has much more info than below).
Of course there are some famous longer attempts at this, such as here, here, here, here, here, here, and here, and most recently Sara Walker’s book. But I wanted to poll a specific group of collaborators and diverse thinkers and see where their opinions were today and how they related to each other in a mindmap of the field. I’m still working on it, so if you have a great definition that meets the above criteria, add it to the Comments below and I’ll be in touch if I can use it in the paper.
Meanwhile I have a few dozen entries already, and started the analysis. My goal was to identify several primary features along which these definitions could be distinguished. What kinds of concepts do the scientists focus on, in their definition – do we recognize life (and distinguish it from non-life) by its behavior (function), its origin story, a specific composition, its effects on an observer, etc. Basically, I’m doing a kind of principal components analysis on these definitions, looking for independent axes along which a morphospace of definitions could be built. I identified a few, which will be described in the paper. But then I wondered, could current language models do this too? I experimented with two: GPT-4o and ClaudeAI (I have no relationship to either company). I haven’t yet tried open-source ones; if you try that, please post results in Comments!
The instructions I gave were as follows (I’m not very good with prompts; feel free to add Comments if there was something that could be improved here):
You are an expert philosopher and biologist. We will be analyzing a number of scientists’ definition of life to categorize them into groups. I want you to think carefully and take it step by step, with no confabulation and not changing their sentences. Here is a list of different scientists’ definitions of life – each scientist’s name is followed by a colon, followed by their definition. I want you to think about these definitions, and identify three conceptual principal dimensions or categories across which they can be distinguished and grouped.
GPT-4o produced the following, referring to the various people whose definitions emphasized each type of distinction being emphasized as a criterion for life:
Based on the definitions provided, three conceptual principal dimensions or categories across which these definitions can be distinguished are:
- Thermodynamic and Energetic Principles:
- Definitions that focus on the idea of life as a system far from thermodynamic equilibrium, with self-maintaining, energy-utilizing processes. This includes definitions that emphasize energy dissipation, self-organization, and metabolic processes (e.g., JT, SN, and AN).
- Information and Complexity:
- Definitions that highlight the role of information processing, complexity, and learning in living systems. These often involve the concepts of information flow, self-replication, and the coupling of informational and chemical networks (e.g., PD, CA, GE).
- Autonomy and Agency:
- Definitions that describe life in terms of autonomous problem-solving, goal-directed behavior, and agency. These definitions often include the capacity for self-maintenance, self-regulation, and the ability to respond to environmental changes (e.g., AS, FH, MS).
These dimensions allow us to categorize the different perspectives that scientists have on what constitutes life, helping to structure the analysis for your review paper.
This is what Claude said:
Certainly, I’ll analyze these definitions of life carefully and identify three principal conceptual dimensions or categories across which they can be distinguished. Let me think through this step-by-step:
- Then, I’ll identify recurring concepts and approaches in these definitions.
- First, I’ll review all the definitions to get an overall sense of the common themes and differences.
- Finally, I’ll synthesize these into three principal dimensions that capture the key distinctions between the definitions.
After careful consideration, here are three principal conceptual dimensions I’ve identified:
- Structural vs. Functional Focus: Some definitions emphasize the physical structure or organization of living systems, while others focus more on the functions or processes that characterize life.
- Reductionist vs. Holistic Approach: Definitions vary in whether they try to reduce life to a single essential property or mechanism, or whether they view life as an emergent phenomenon arising from multiple interacting components.
- Boundary Specificity: Definitions differ in how clearly they delineate the boundary between living and non-living systems. Some provide strict criteria, while others suggest a continuum or acknowledge fuzzy boundaries.
These dimensions capture significant differences in how the scientists approach defining life. They don’t change or misrepresent the original definitions, but rather provide a framework for categorizing and comparing them.
Would you like me to elaborate on any of these dimensions or provide examples of how specific definitions align with them?
Interesting and not bad at all! Different dimensions picked by the two LLMs, but both reasonable and seem to capture deep aspects of the question. Next I checked to see if it could draw this in a graphical format: (I’ve blurred people’s names until everyone signs off on the final analysis)
Great. Now, can you draw a 3D diagram of a cube representing the space defined by your 3 dimensions. On this cube, please place dots corresponding to where each definition would fit, making judgements about how far along each dimension a given definition should go.
Claude gave me:
GPT produced:
Also not bad. One weird thing is that no matter how hard I tried, I could not get GPT to give me a version with the axis labels not cut off. If I asked it to zoom out a bit, it would do so on the data points, but still cut off the on the axis labels.
This is an experiment in progress and will be updated as the project continues.
Featured image by Midjourney.
I thank Eors Szathmary for suggesting a few links to papers on this topic.
Leave a Reply to Benjamin L Cancel reply