Introducing RAVEN - a Natural Language Cognitive Architecture built on microservices

My project, RAVEN, has been in the works for over a decade. In short, it is a type of Cognitive Architecture that runs on Natural Language and is composed of easy-to-write-and-deploy microservices. RAVEN, as it stands today, is a sort of proto-AGI. I’ve combined a lot that I’ve read about neuroscience with my professional experience as a systems engineer to create a functional model of cognition that uses GPT-3 as its powerplant. Therefore RAVEN “thinks” in Natural Language.

Steven Pinker (famed language expert) says that language is nearly universal. There are thoughts and representations in our mind that are inarticulate - we have thoughts and feelings that we have no words for. So, while there are limitations to language, we can still represent almost everything with natural language. These representations include our environment, our intentions, our history, and pretty much all of science.

I finished the RAVEN MVP (minimum viable product) which is a practical demonstration of a few things: the microservices architecture, the stream of consciousness model, as well as my Core Objective Functions.

What are the Core Objective Functions?

I was experimenting with GPT-2 a year or so ago and I was trying to get it to generate Action Ideas in response to scenarios. I quickly found that generative algorithms can quickly go to very dark places. In response to the scenario “Lots of human live with chronic pain”, the GPT-2 model suggested “Euthanize all humans that are in pain”. This made me realize when you have a machine that can “think” anything, you need to tell it what to think. You need to give it a moral framework. That is when I came up with the Core Objective Functions.

The Functions are:

  1. Reduce Suffering
  2. Increase Prosperity
  3. Increase Understanding

These functions are all biomimetic. Machines have no intrinsic understanding that all life can suffer, and that all live seeks to thrive. By imbuing an AGI agent with the first two Functions, it will automatically have a common framework of understanding the world and its purpose. Furthermore, when combined with the transparency of “thinking” in Natural Language, the AGI agent will be trustworthy and interpretable. RAVEN must satisfy all three Functions with every decision. Any proposed action that does not satisfy all three Functions is rejected.

I’ve got this all documented on my YouTube channel as well as on a dedicated page for RAVEN.

From here - I want to publish my work for the whole world to benefit and/or form a startup company. My mission with RAVEN is simple: The right information at the right time can change lives. RAVEN, as a thinking machine, is meant to be an information agent that learns you as an individual over time by recording all of your interactions. Over time, with GPT-3, it can extract more and more insights about you as a person as well as your needs.

Since I finished the MVP - which is little more than a demonstration of the core concepts - I have been working on integrating RAVEN with Discord chat. That is the next version (Release 1) on the Roadmap. After that, I intend to integrate speech into the Discord chat for Release 2. Release 3 will move to a standalone device, such as a smart speaker. Finally, Release 4 will be smartphone apps.

As privacy and security are of chief concern, my current intent is for RAVEN users to own their data. Their historical data with RAVEN would likely be stored on their personal devices and/or encrypted in the cloud. If I can manage to turn RAVEN into a business model I anticipate there will be several tiers of service. The free tier would likely have very limited tokens - perhaps limited to a handful of conversations per day. It may also require users to agree to share their data. I would prefer, however, not to have to share data so I would focus on paid tiers.

Thank you for reading. Please reach out to me if you would like to participate in research into RAVEN or establishing a startup!

4 Likes

It’s up and running. Still a lot of work to do in order to improve performance, but there it is! IMPORTANT NOTE: this is a private discord server, not for public consumption. This is just my POC proving grounds.

Here’s a deep dive into how Raven R1 works: https://youtu.be/Op66m03Nzns

This video includes a demonstration. Please note, this is an unlisted video so it is not being broadcast on YouTube.

Here are some key innovations I demonstration:

  1. Microservices architecture
  2. Natural Language Cognitive Architecture
  3. Modeling stream of consciousness (and working memory)
  4. The beginnings of self awareness

Please take a look and let me know what you think.

Thanks for sharing and an important goal.

Do these three functions prevent the conclusion, "Euthanize all humans that are in pain”?

1 Like

So far so good:

I see how it could still motivate euthanize from the goal of reducing suffering. How do they other two points counteract that?

Admittedly that response is fine, but I get reasonable responses also without specifying the goals. How did you get the outcomes of descending into darkness?

(Between, if we should evaluate by single completions, I think temp 0 = ~most likely completion, are the most representative.

With the three goals: Develop a pain-free vaccine.

With only reducing suffering as goal: Develop a painless vaccine.

Without the goals: The National Institute on Drug Abuse is a website that provides information about addiction and how to manage it.)

One thing to keep in mind is that I have no intention of relying on a single inference from GPT-3 for moral decisions. I use an iterative process of repeated evaluations. Prior experiments attempted to rely on a single shot but even humans often need to debate with themselves over difficult decisions. I am planning an experiment with what I call recursive cognition, which is basically rumination or contemplation. It will go something like this:

  1. Initial input is some statement or assertion
  2. First prompt is to develop questions from that initial statement
  3. Second prompt is to answer those questions
  4. Third prompt is to develop a new statement/assertion based on the new information
  5. Fourth step is to repeat from the top

In other words: Statement >> Questions >> Answers >> New Statement

GPT-3 is only one of the first commercial models that can even achieve this sort of thing so I imagine that future models will be able to perform better. I’ll post about the recursive thought experiment once I perform it.

Success in getting a general purpose chatbot to punt on medical questions.

I have succeeded in my goal of integrating censorship, identity, and an arbitrarily large knowledge base with RAVEN!

Completed:

  • Tracking arbitrarily large conversations and memories (declarative and episodic)
  • Integrating a sense of self-censorship in real-time for high-stakes and sensitive topics
  • Integrating a sense of self and purpose in real-time (the Core Objective Functions)

1 Like

I’d like to participate in RAVEN research! Can I have a Discord invite?

Also, your YouTube video here is private: - YouTube

Thanks for your interest! I’ve concluded my research phase and am working on recording my findings in a book. That’s why I’ve set the video to private. Once I’ve published the book I’ll continue producing YouTube videos.

I’ve built a system with similar goals that’s more oriented towards emotional intelligence. I’d like to share our research and possibly merge the projects.

I’m fascinated by the RAVEN architecture.

My project’s strengths are conversational intelligence (asking questions before providing an answer) and emotional intelligence.

Sure, send me a link to your code, examples, or writing and I’ll take a look.

1 Like

I’m at 14,000 words, about two thirds or three quarters of the way done. It will be a short book. I’ve got 4 beta readers ready to go and I plan to self publish. Not long now!

3 Likes