In my ongoing exploration of the creative potential for AI — it seems to me as though many of the short-comings of the current systems could be distilled down to a single problem:
AI’s lack personality
Whether it’s in telling a joke, writing a piece of fiction, or role-playing an NPC in a video game (as seen in the technically impressive, narratively disastrous demo below) — I believe it’s the lack of personality that is holding AI back from its creative potential.
This may be easy to define as a problem statement, but it’s incredibly challenging if you start exploring solutions. Which is probably why AI interfaces like ChatGPT aim for a neutral tone by default.
Creating Personas
I’ve experimented with personas, where I’ve asked the large language models (LLMs, like ChatGPT) to communicate in a style consistent with the character backgrounds they’ve been given. For instance, I’ve used an actual person as reference, like Douglas Adams. From there, presumably, it extrapolates a style based on his body of work, and his interviews. But is that enough to capture a personality? What if someone, in their creative work, is deeply driven by some childhood trauma which they would never share on public record? Can you capture their personality without that context?
Or, using an example of a fictional persona, suppose you set up a character with the background of a 17th century pirate? Beyond the cartoon representations of pirate-talk — how deep would that characterization go? How broad of a conversation could you have with that character before you find inconsistencies?
To Blandly Go Where No One Has Gone Before
There is a demo of the game Skyrim, known for its extensive story world, where someone linked all of the NPCs to ChatGPT. Each NPC was given enough context and domain knowledge about their character to play the part. However, beyond a certain point, the character seems to slip back into the bland, slightly philosophical default tone of ChatGPT. The characterization clearly has its limits even when given a well-defined backstory.
ChatGPT and Personality Types
I’ve read a few studies around giving ChatGPT personality tests, including the Myer-Briggs test, and those provide some interesting insights about the ‘default personality’.
I haven’t seen any studies on testing ChatGPT as it adopts a given persona
For instance, if you tell it to adopt an ISTJ personality — does that hold true against a wide range of questions? (I’m aware there are arguments for & against the MBTI classification system, but I’m presuming it is at least consistent in its result.)
The Mindfile & The Personality Vector
There is a concept in transhumanism called ‘the mindfile’ — the idea that you could record enough data about a person to contain enough information about them that they could be faithfully recreated in some form. Borrowing from that concept…
I wonder if we could distill a ‘mindfile’ down to the size of a context window for an LLM — I call this The Personality Vector
Suppose that we have 256 true or false questions, carefully crafter so that the person answering them would consistently give the same answer over time (or at least a reasonable period of time between life events). Let’s also assume you could craft those answers so that everyone would have a unique set of answers like a fingerprint. Each of those true or false responses is a single bit of information so you would have 2²⁵⁶ or 8 BigIntegers worth of data to store — not very much. (The same as a private key for Bitcoin)
For reference, 2²⁵⁶ is close to the estimated number of atoms in the universe.
This seems large enough to encapsulate an entire personality — but it’s also kind of like saying the meaning of life, the universe and everything is: 42. You still need to ask the right questions.
If you allow me the conceit that we could devise 256 questions to accurately capture the attitudes, biases, belief systems, passions and world views of a person — this Personality Vector would be ideal for training AI systems. Arguably, this already exists on Facebook and Cambridge Analytical infamously made a business around it.
Imagine the possibilities this would open, if you could use your Personality Vector, or one that you’ve designed for a fictional character, or the average of two Personality Vectors from known celebrities, and put them into a video game where they are given backstories, societal context and knowledge of in-game world events.
I think this is a topic I’ll keep revisiting as I learn more about AI development and the current state-of-the-art in research.
