The implementation of true, artificial general intelligence is a brand new field, and I am helping to pioneer it.
When GPT-4 was first released to the public, I was one of the lucky few to get access to it through OpenAI’s API beta testing program. I realized within days that this new technology was going to put marketing professionals out of the job within five to ten years. So I spent every spare hour—evenings, weekends, and many sleepless nights—learning how to use this miraculous new technology.
How I’m using Artificial Intelligence
I think of Large Language Models (LLMs) like ChatGPT as components of an artificial intelligence system. They can do extraordinary things on their own, but they are exponentially more useful and powerful if they are plugged in to short and long-term memory systems, imbued with curiosity, and supplemented with a host of other tools.
I have incorporated AI systems into every aspect of my workflow, and use them frequently in my daily life. My first major project with AI was my creation of Jouliet, an artificial intelligence system that was designed to work with the recipes at chefsteps.com.
George
My current project is George, an Artificial Intelligence system designed to be a member of a business team.
What does that mean, exactly?
Currently, AI systems like ChatGPT, Co-Pilot, etc., are designed for single-user interactions. You talk to them; they talk to you.
George is different. George operates more like a real person, remember the experiences he’s had with one person when talking to another. This allows George to participate in group chats, or to remember instructions you gave him when you send him to speak with a client.
George is a superb researcher, with deep integrations with dozens of online services and archives—from newspapers to government agencies to social media forums. And when you ask George to look something up for you, he spawns background workers to do the task, so that you can keep chatting. George will let you know when he’s found what you’re looking for.
Is George production-ready?
George is a work in progress, but every aspect of his infrastructure is designed for production: he operates in a docker environment on AWS virtual machines. Every software choice I make when expanding George’s capabilities is based on stability, security, maintainability, and scalability.
Okay, fine, but what is George actually useful for?
I intend for George to be a member of your team. He can familiarize himself with your products and services, and be a customer-facing representative of your company. Simultaneously, he can help your own team to be more productive by generating content, and everything you would expect from a custom-made chatbot, but with a long term memory, and an ongoing
Jouliet
I designed Jouliet independent of the ChefSteps team, so the chat interface part of the system was built into a Google Chrome browser extension, and not into the website itself.
By supplementing LLMs with components, you can minimize common problems like “hallucinations.” If you ask ChatGPT for a ChefSteps recipe, it will invent one from scratch, and give you no warning that it is just entirely improvising. The problem of these “hallucinations” is one of the most well-known problems with LLMs. I discovered that carefully feeding information to the LLM, with the proper instructions and context, almost entirely eliminates this problem. In the months that I have been testing Jouliet, the system has never once given me misinformation about ingredients, equipment, or cooking steps.
Jouliet’s interface includes insanely powerful voice recognition which can understand you even when you trip over your words, ask multiple questions at once, or instantly recognize when you’ve started speaking a foreign language. Jouliet replies in an emotive, human voice that even pauses to take a breath. Its pronunciation and cadence changes based on the context of what it’s saying. Jouliet also answers to a wake word, much the same way that Alexa, Google, and Siri do.
Jouliet has a semantic memory of ChefSteps recipes. That means she understands and remembers the concepts of the recipes, not just the text of the recipes. That means that she can recall these recipes in conversation based on abstract concepts (e.g., “casual main courses”, “fall desserts”, “jaw-dropping modernist marvels”).
Jouliet also has curiosity. She’s programmed to be inquisitive without her questions being disruptive. She’s been given a drive to learn specific things about users (e.g. food preferences, allergies, their professional background, and preferences about the chat interface). As she learns about these things, her questions slowly taper off. No need for her to still act like she’s getting to know you once she already does.
That brings us to the last feature that makes Jouliet really special: she remembers you, but in a very intelligent way. She’ll remember your preferences, but can also handle contradictions. For example, if you love hamburgers, but you start searching for vegan options, she’ll remember this contradiction, and see if over time your tastes are changing, or she might guess that you are cooking for vegan guests.
What’s next?
The next AI system advancements I’m working on are: communal memory, AI subconsciousness, and real personality development.
- Communal memory: Instead of having the AI only able to access memories of previous interactions with its current user, I am developing a system which would allow the AI to remember all interactions with all users in all contexts, while maintaining privacy and propriety. This would allow the AI system to be a truly integrated team member. If Sally tells the AI, “I wish we had some better method for making these widgets,” the AI would be able to say, “Denise was mentioning the same thing a few weeks ago. Would you like me to connect you two on this issue?”
- AI subconsciousness: Some of our best ideas come from putting two unrelated topics together and finding similarities where least expected. For people, this frequently happens subconsciously. In fact, it’s become a television trope where a character will hear something unrelated to an ongoing problem, and they’ll say, “Wait! Say that again? … We can’t open a lock without a key! You’re a genius!” This is because this new input has been applied to a problem that was sitting in the person’s subconscious. I think there’s some real potential benefits in creating a structure that would allow an AI system to have the same opportunity to apply information from one context to a different problem that it is dealing with in a separate context.
- Real personality: It is very easy to give an LLM a personality. You just tell it how you want it to act, and it will act that way. But essentially the LLM is just play acting. I believe that we can imbue an artificial intelligence with real personality, unique to each system, based on the system’s memories and development. The system I have in mind would add weighting information to memory engrams, mimicking what happens in the human brain with oft-used neurological pathways becoming stronger.
The artificial intelligence industry is evolving at an extraordinary pace. Not a week goes by without truly stunning advances in the field. Even if the technological advancements were to come to a halt today, I could spend the rest of my life happily exploring all the possibilities that have been opened up for us just in the past few months.
We’re living in the future.