Towards the Bernie Sanders model
Next best thing until we can actually upload people
Uploading people into some kind of a computer has been a science fiction trope for a long time - and so far it’s mostly a pipe dream. I’ve experienced two things recently that inspired me to write this post - and create the associated project: a pretty good animated series named Pantheon, and re-reading the excellent book “Excession,” by Iain Banks. Both works feature the idea of abstracting, copying, and augmenting human minds through digital uploads.
Excession, specifically, has this concept of a “mind state” - a snapshot of a person’s mind at some time, which isn’t really 100% identical, is usually focused on some topic, and up to a point has a mind of its own.
This mind state as described in the book isn’t really autonomous nor really intelligent, but the most important thing about the mind state is that you can talk to it pretty much as you would with the original person. It will respond to you as the original person would.
It’s the closest thing before a “real” mind upload.
You wouldn’t steal a Bernie Sanders
I think we can do the “mind state” thing right now. The choice of Bernie Sanders came after I listened to his interview on Joe Rogan, where I got the impression that he is a very lucid speaker, expressing his ideas logically and in a non-pretentious way, which is a rarity. This is not a politically motivated choice - I’m European. If anything, it is partially motivated by his age. In any case, I did this with every respect for Bernie, his message and his work.
What I did is take publicly available texts from him and about him - mostly from his official web page, and podcast transcripts, extract topics he’s interested in and his opinion on these topics, then fine-tuned an AI (a LLM) to respond to questions as he would respond.
So, this is Bernie’s “mind state” in the spirit of how a science fiction book described it. You can talk to it, ask it questions, and it (yeah, I’m calling it “it”) will respond in a way Bernie would have responded.
It is a completely digital artifact - a couple of files you can download, and try on your own. I was very careful to make the whole process runnable on low-end hardware locally. No cloud is required, no service providers, just some open source software.
Talking to your personal Bernie Sanders
The model itself is published in this HuggingFace repository. It’s not nearly perfect - I’m calling it a proof of concept - but it does demonstrate what I’m talking about. Here’s the easiest way to run it:
Install the software named Ollama (on your laptop / desktop!)
Find out how to open a Terminal / Command line window, and type:
ollama run hf.co/ivoras/bernie0.1:F16
After the model is done loading, just talk to it.
Here are a couple of questions and answers I got from it:
>>> What do you think about workers rights?
I strongly support workers' rights! It's deeply unfair that people are expected to work themselves into exhaustion without adequate pay, benefits, or protection from exploitation. We need to see policies that prioritize worker safety, fair wages, and comprehensive healthcare for everyone – not just the wealthy few.
It's about recognizing the fundamental value of human labor and ensuring that workers are treated with dignity and respect. It's about building a more just and equitable society where no one is forced to choose between their job and their well-being.
>>> What specifically are you doing to address the youth mental health crisis?
I'm working tirelessly with lawmakers and advocates to push for increased funding for mental health services in our schools. This means providing access to counselors, therapists, and other support staff to help young people navigate the pressures of modern life.
It's also about challenging the stigma surrounding mental illness and ensuring that these resources are accessible *everyone*, regardless of socioeconomic background. I'm advocating for policies that prioritize student well-being and recognize the importance of mental health in academic success.
>>> How do you manage the constant negativity and attacks online?
It's a constant battle, but I don't let it consume me. I focus on building support networks with like-minded individuals who share my values and are committed to fighting for justice. We lift each other up when we're feeling down, and that's what makes us stronger. Plus, I keep writing – sharing stories, advocating for change – because the power of narrative is incredibly powerful. It can inspire people, spark conversation, and ultimately lead to real-world change.
Implications and consequences
In addition to science fiction, the idea of mind uploads was, and especially will be, addressed by countless scholars, but here are some points that drew my attention:
The models are for the most part stateless. Technology will improve on this, but there will probably always be a “core” that is static and unchangeable, expressed as just a bunch of files to be moved, and more interestingly, copied. Do you want to have a two-way conversation with a billion people at the same time? You can’t, but your mind state can. This was a mind-boggling concept when I read it in Charles Stross’ book “Accelerando.” Bonus points - what happens if the non-static parts can also be duplicated?
The models are a snapshot. Bernie lives on. He can write new speeches, be interviewed in more podcasts, even change his mind on certain topics - but the model will always be a snapshot in time. As soon as the model is released, it’s obsolete.
The models are limited. This model is only trained on the publicly available information. If Bernie didn’t write or speak his opinion on cats, the model doesn’t know it - and so we can’t get this information by interacting with the model. Another way to think about it is that the models are focused on a subset of topics that make the actual person’s mind.
The model does occasionally hallucinate. This is something that can be worked on, but at least this proof-of-concept model will make something up if you ask it about information which it doesn’t know about (like its opinion of cats. Luckily, those hallucinations seem to be fairly benign and neutral in tone.
Related to that, there's the issue of how truthfully the model represents reality. I don't think it's good enough to testify in court, so to speak. I imagine this issue will become important if these types of models take off: since multiple parties can create similar models, which one is truethier? Will there be an “officially endorsed” model? A light-weight public model and a paywalled big one?
This project fits in the general umbrella of style transfer, but maybe it goes a tiny bit further and touches on what could be called ideology transfer.
The model doesn’t really understand time. This is also something that can be improved, but currently, the model treats everything it learned as “timeless”. As far as it’s concerned, old topics like COVID are still relevant. Trump is now, was, and always will be president. I have some ideas how to work around that, if I find the time.
It's not intelligent. Just saying.
We are IMO still far from a true AI (re-labelled AGI), but we are even further away from uploading our minds.
For people that have a large volume of published text or speeches, this is about the best we can do to ensure a kind of a functional, interactive immortality - an interactive snapshot of them. If we connect it with a hologram, we’d really be living a sci-fi future.
What's next?
I want to try some more things to make the model better and true to its origin. I was probably overthinking about some aspects of fine tuning and can do better on that.
The sources for this fine-tune, including some training data are in GitHub and I'll make them public when I clean them up a bit.