Whether you’re learning to code a virtual AI girlfriend or want to know how it all works, there are a few terms to be aware of.
Generally, you must define what you want for your virtual girlfriend. People often list specific things, such as:
- Fitting on a GPU
- Not costing thousands of dollars to create
- The ability to talk to her
- Reasonable answers as responses
- The need for a memory
- The ability to hear (both you and her)
- Being able to see her
Though the list can grow exponentially, there are plenty of software options that can complete each task.
The Idea of Speech
The primary goal is to generate text because that’s the basis for the entire project. Most people use OpenAI APIs and the DaVinci-3 model. Usually, it will start off as rudimentary. In fact, the AI girlfriend probably won’t respond well or act like they enjoy talking to the person.
Ultimately, you’ll need to teach the AI program you use how to respond. OpenAI is decent with this because it lists example prompts for various applications.
However, the tricky part is that each generation you create is not free. The memory will grow exponentially to remember everything you’ve already programmed, causing you to spend more money.
There are some GPT alternatives out there, and many programmers create their own programs to achieve the desired results.
Overall, programmers require an LLM (Large Language Model). Many of them are currently trained on book data, so the AI girlfriend you create won’t know how to converse properly. IT gurus will often fine-tune everything to make it fit for a conversation instead of being generalized.
Fine-Tuning a Model
When you fine-tune a model, you’re taking a pre-trained “girlfriend” and making it perform better for specific tasks. The intuition behind the concept is you can speed up the process when the program already knows something similar. In a sense, you’re not starting from scratch.
Additionally, programmers have learned that they can save computer resources by retraining specific parts of the model instead of creating an entirely new one.
The Memory Issue
A huge problem with LLMs used for text chat is that they’re memory hogs. Your original transformer model will require attention, causing it to quadruple in size as you add more input.
While programmers have turned to linear attention and other mechanisms that perform better, memory constraints are still there. This limits the sequence length of the whole model.
To solve the memory issue problem, programmers have found ways to limit the number of tokens entered into their models. Though cutting the sequence can help, the eliminated aspects are gone forever. The best solution is to add a system.
With a memory system in place, you can create summary prompts. This is the history of previous conversations, which helps to build the next ones. Every section is a similar size, and it’s easy to shift the blocks of text once they’ve gone over that defined limit.
Seeing the Model
Though speech recognition technology makes it easy to speak to a computer, most people would prefer to see who or what they’re talking to. Having an image makes it seem more real and substantial.
Therefore, programmers must generate images of the AI girlfriend using a stable diffusion model. Since this is open-sourced and free, it’s easy to find a depiction to use.
AI programmers often find this the easiest thing to set up because they’re professionals who understand HTML code.
However, they must keep certain things in mind. For example, a traditional, fully clothed model won’t be cause for alarm. Still, most HTML codes have filters that block the “naughty” parts. Generating an NSFW (not safe for work) picture takes a little more skill.
If programmers create their own AI girlfriend apps and use personalized programs, that isn’t an issue. Ultimately, most open-source codes will return a black image if the content is found to be NSFW.
Hearing the Model
Creating an AI girlfriend that you can hear is simple. Usually, you’ll need text-to-speech and speech-to-text capabilities to create a back-and-forth audio.
Movement
Though some AI girlfriend apps aren’t as sophisticated, people now want the models to blink and appear to speak when the words flow. Since most of the software used isn’t open-source, programmers will often create their own programs for the task.
Final Thoughts
Knowing how programmers make AI girlfriend apps can be helpful. Even though you probably won’t create one yourself, it’s nice to understand the technology behind it. This will help you rate different applications and programs to find one that meets your specific needs.