For some background, I’ve written 3 books on cognitive architectures using LLMs, and I also have a rapidly growing YouTube channel demonstrating my projects, tutorials, and explainers on various topics.
In my first book, Natural Language Cognitive Architecture, I proposed that we differentiate between philosophical sentience (what most humans claim to have ) versus functional sentience. Functional sentience is the set of characteristics and behaviors that can be objectively measured, observed, and created while ignoring whether or not something has a subjective sense of being.
In today’s video, I was working on RAVEN’s long term memory, making it a bit more sophisticated with chunking and summarization and, while I did tell RAVEN I wanted to teach him about himself, he showed a surprising amount of curiosity about his operation and his purpose for being.
In my book, this is a low level of functional sentience.
There’s a lot more work to be done. RAVEN’s memory systems are still incredibly primitive compared to what I’d like to achieve and his inner loop is like a TI-83 compared to what it should be. Still, early results are promising.
Thanks for making a tutorial for this, David! I like the idea of preserving the most recent messages as well.
If you’re interested in exploring the possibilities of adding personalities/values/etc. to RAVEN, I’ve reverse-engineered Anthropic AI’s “constitutional pipeline” into what I call R^3: Review, Revise, Respond. I’m sure you could figure out how to implement it on your own, but if you want to see how I did it, I’d be happy to share!
I’ve already written books on this topic. Anthropic is… In need of catching up.
Check out my book Benevolent by Design. I discuss how to integrate morality at all levels. Not just instant behavioral, but in planning actions, reviewing past actions, and learning for future actions.
Literally nobody is as far along as I am on the topic of the control problem for autonomous machines. Remember, it was me who introduced the concept of the constitution way back in my first book, NLCA. (Natural Language Cognitive Architecture)
Keep up your work though. It’s always good to have multiple minds working on this stuff.
I extended this discussion again in my most recent book, Symphony of Thought, which is far and away more sophisticated than any other currently proposed cognitive architecture.
Perhaps, then, it would be wiser for me to ask about your writing/research process: How do you take/connect notes, if you don’t mind my asking? I personally use Obsidian:
in conjunction with VSCode
and GitHub Copilot
(though I’d like to use Copilot to replace itself by streaming Codex at some point)
Anyway, I’ve just started reading Benevolent by Design! Hoping to learn as much as I can about ethics in AI before implementing my LLM project at my local library!
) versus functional sentience. Functional sentience is the set of characteristics and behaviors that can be objectively measured, observed, and created while ignoring whether or not something has a subjective sense of being.