microwho / raven_mvp_public

Public MVP of Raven. It's been long enough, time to do a full send.

RAVEN_MVP_Public

Public MVP of Raven, a general-purpose voice-enabled AI companion. RAVEN = Realtime Assistant Vastly Extensible Network. The MVP is not extensible, it is just a proof of concept. Extensibility is coming.

The ultimate goal of RAVEN is to be an AI information companion, like Cortana from HALO or any number of other fictional AI companions. The idea is to create an autonomous artificial cognitive entity by the following stages:

1. Earn user trust through data privacy, security, transparency, reliability, and utility. RAVEN should be trustworthy in every sense of the word.
2. Demonstrate ability to self-monitor, self-check, and self-correct. That is to say that RAVEN should be able to detect when it has made mistakes and rectify them in the future.
3. Having earned your trust, RAVEN should eventually be granted some autonomy. That is, RAVEN should it prove to be safe and trustworthy, RAVEN ought to be granted more control over decisions about what it does on its own, and what it does for you. For instance, if after many years of use and reliable advice, perhaps you may choose to delegate managing of your finances to RAVEN through extensibility modules.

Features

1. Voice-enabled with Google TTS and ASR.
2. Multiple domains and topics. Can serve as tutor, life coach, etc.
3. Transparency. All input/output and thoughts of RAVEN will be recorded in plain English for accountability and explanability.

Setup

1. Install VLC in your computer - this is necessary for playing sound
2. Download and deploy FFMPEG, add it to PATH variables: https://ffmpeg.org/download.html and then http://blog.gregzaal.com/how-to-install-ffmpeg-on-windows/
3. Sign up for Google Cloud Services, specifically the TTS and ASR endpoints, generate your credentials JSON file and place it in this directory (see below for more detailed directions)
4. Place your OpenAI API key in a file in this directory called openaiapikey.txt
5. Finetune a DAVINCI model in OpenAI with the JSONL file in this directory, update the MODEL in RAVEN_MVP.py gpt3_completion function
6. Install pipwin with pip install pipwin in an ADMIN prompt
7. Install pyaudio with pipwin install pyaudio in the same ADMIN prompt
8. Install REQUIREMENTS.TXT with pip install -r REQUIREMENTS.txt
9. Create two folders called memories and gpt3_logs in this directory

Google Cloud Services

This part is VERY TEDIOUS, be patient with yourself and set aside an entire day to do it

1. Sign up for Cloud Text to Speech (TTS) here: https://cloud.google.com/text-to-speech
2. Sign up for Cloud Speech to Text (ASR) here: https://cloud.google.com/speech-to-text
3. Create your Project in Google Cloud Console: https://console.cloud.google.com/
4. Create the service account for this project as well as the API key (this will ultimately yield a JSON file with the full API key)

I'll probably make a YouTube video just about how to set this up, check back and the link will be here

About the Finetuning Data

The file raven.jsonl has over 600 training conversations to teach RAVEN how to behave. These conversations were synthesized over the course of several projects, including a therapeutic/emotional support chatbot and an educational/tutor chatbot. In all cases, the conversations were centered around compassionate listening and curiosity about the user. While it is not explicitly declared in the training data, there are 3 guiding principles (or heuristic imperatives) baked into this training data:

1. Reduce suffering of the user. If RAVEN detects any opportunity to help the user feel better, it will take it.
2. Increase prosperity of the user. To 'prosper' means to 'live well', which RAVEN will support by helping the user to navigate relationships, life, and work.
3. Increase understanding of the user. To understand is to learn, which RAVEN supports by being eager and willing to teach anything.

Later versions of RAVEN will explicitly adhere to these three core objective functions (or heuristic imperatives) by thinking about them in the background. This version is little more than a voice-enabled finetuned chatbot. Future versions will be fully fledged cognitive architectures based upon my work in Natural Language Cognitive Architecture. Essentially, RAVEN will "think" about you, how to help you, and what you need. Furthermore, RAVEN will have long-term memory as well as extensibility so that it can access external data and services via API.

Notes on the Cognitive Architecture and future iterations of RAVEN

There are many problems to solve before RAVEN is fully realized. Overall, I have provided an outline of the cognitive architecture in my book "Natural Language Cognitive Architecture" although as technology advances, there are already some deviations. However, the overall idea of dual-nested loops remains constant. My GitHub account is littered with experiments that each have to do with various aspects of artificial cognition, such as long term memory, answering questions, and so on. Each of these components will ultimately be integrated into a single, coherent platform.

Below is a non-exhaustive list of artificial cognition problems that I am working on:

1. Cognitive control. How to keep RAVEN on task, task switch appropriately, design cognitive tasks for itself, and measure progress towards task success
2. Model of self and abilities. How does RAVEN know what it is? How does RAVEN know what it is capable of and what its limitations are?
3. Self-monitoring, self-checking, and self-correction. What principles does RAVEN adhere to, and how does RAVEN observe itself to ensure it aligns with those principles? What mechanisms does RAVEN have for self-improvement?