Klara and the Sun
What does it mean to be human? What does it mean to be conscious? What would it take to upload someone’s consciousness onto a computer? A programmer’s job is to take reality and define it in precise terms to be represented in zeros and ones. The models we use in programming (such as classes in OOP) need to be human-understandable in order to share them with others. The magic of good code is that other programmers can contribute to it and build on top of it in a reliable manner. But maybe we need to throw all of that out the window. Maybe the only way to represent something as incomprehensible as humans is to use a system that is just as incomprehensible. Enter LLMs.
As a thought experiment, imagine I had a dataset of a hundred personal questions and spent the time to honestly answer each one. If I used this dataset to fine-tune an LLM, how accurate would it be in representing me? Probably not very accurate, but what if there were a thousand questions? A million questions? Maybe at some point the LLM would start to learn some of my underlying value systems and emotional drivers such that it would be able to respond to questions like I would even if they weren’t in the training set.
A few months ago, Justin and I had been noodling on the idea of cloning ourselves with language models. The concept came back to me this week as I was reading Kazuo Ishiguro’s novel Klara and the Sun. He creates a fictional sci-fi future where parents can buy AFs (artificial friends) for their children. AFs are humanoid robots that have their own independent thoughts and emotions, but seem to be generally subservient to their human owners.
I think it would be really cool to create a real life version of Klara: an AI that is not just an intellectual equal but also a social and emotional equal to us humans. Or maybe even upload my consciousness to the cloud like Kayaba in SAO. It would be incredible to create a snapshot of my 22 year old self that I could talk to later in the future. I wonder how much I would have changed by then, and if this snapshot could preserve some of my youthful naivety that I will inevitably lose over time.
The amazing thing is that all of this might be possible now with LLMs. I need to read more stuff in the LLM interpretability space to see if my questions above have already been answered but anecdotally I haven’t seen too much regarding this topic. I’ve seen a few attempts on Twitter to create “clones” of people by using data from social media and text messages, mostly for commercial use cases like coaching and advertising. However, I don’t think that type of data is enough to create the highest fidelity human-like AIs that I’m seeking. This type of “clone” would require a dataset that is customized to one person who is also willing to share a lot of personal data and private beliefs. It would probably take a ton of time to curate such a dataset in terms of how much self reflection someone needs to do to compile accurate answers. I think this would be a fun project to try for myself but I don’t really see any commercial use cases so maybe sometime in the future.