Bots, GPT-3 and Virtual Worlds
I didn’t know much about Bots, just that most people dislike them because of the early days of Customer Service Bot conversations that usually ended in, Get-Me-A-Supervisor!
Now there’s billions of bots being blamed for bad posts and tweets that are driving people bonkers. Under Facebook’s Coordinated Inauthentic Behavior policy (CIB), over two billion fake accounts were removed during just one sweep in 2019.
About 40% of all Internet traffic is bots.
Since they are part of the new landscape of influence and interaction, I want to understand how they are being used to help breed mistrust — but mostly I want to understand how they can be used to bring us together and make diverse virtual worlds more engaging.
Trending AR VR Articles:
1. 9 Must Have Skills To Become An AR/VR Developer (With Course Recommendations)
2. Top 17 Augmented Reality Statistics, Facts and Trends for 2021 and Beyond
3. How Augmented Reality Will Improve Your Next Travel Experience
I figured out pretty quickly that it helps to think about different types of bots. I won’t be digging into so-called Web Bots. They’re the ones that don’t talk to us or interact with us; they just go around indexing websites and scraping content.
The Social Media Bots are simple, programmable little schmucks that just do what they’re told — add Likes or Favorites or post pre-scripted text. They can appear to be real Users, which is why they are considered manipulative and misleading.
It is the third category, the broad category of Chat Bots, that has my attention; specifically, the Discord Bots that can converse and also help manage server communities. I started with the highly-rated Loritta, but have since moved on to MEE6.
This type of bot has become very popular because someone else did the coding to set up a template so an administrator just needs to focus very carefully on what they want the bot to do. All Discord Bots and similar bots can accomplish many crucial but repetitive tasks that are difficult for actual humans to do over and over while staying human.
Welcoming people is a classic example.
It feels nice to be Welcomed. Walmart figured that out a long time ago. But what if you can’t afford to have people waiting around for someone to Welcome? Or conversely, if a few people were Welcoming lots of other people constantly, how human would it be, eventually?
Being recognized and greeted by name by a bot creates a positive and welcomed feeling in an online or virtual community, especially when the bot administrator puts some personality into the script. Everyone knows it’s not really the bot’s personality being expressed, but we feel someone’s personality. That person couldn’t be here right now so the bot is their self-expression.
And, Welcoming is, of course, just the beginning.
Because one branch of what we now call Chat Bots evolved to support managers of very large gaming communities, sophisticated behavior management and social influence tools have evolved.. The idea of earning points and progressing in rank by participating in various ways is one example. Known as Leveling, a well-tuned system can help encourage people to get involved — but who has the time to track all the participation and points? The Discord Bot!
Leveling is both a rote function and a sophisticated approach to building community. Bots also constantly look out for what their Admin tells them are Not Acceptable behaviors, with various degrees of Consequence, all of which has to be managed. No human could stay on top of all that. Only bots.
Dabbling in the social influence side of MEE6 was a crash course in what is easily available for a semi-technical person like me. It is related to my main reason for learning about bots in the first place, which is making virtual worlds more engaging.
People are slowly discovering VR all over the world and some people are discovering platforms for building their own worlds. Semi-technical people like me can build them out of kits, templates and imagination.
When a World is constructed, there it is. It just sits there, in effect, a Gallery of whatever is displayed. Some worlds are elaborate enough to warrant exploration, but at this early point in the VR medium, a static World tends to get boring pretty fast.
People and events designed to create types of experiences make Worlds come alive. The World becomes a character in the experience.
What can bring life to a World when there is no scheduled event? How can people be assisted, individually or in groups, in the creation of unique and meaningful experience in a Virtual World?
At their best, that is what human docents do now in museums and other locations.
With all the Worlds people are building now, 24/7 universal coverage will only be possible through really smart conversational Docent Bots.
That’s pretty close to what GPT-3 already is.
The newest Generative Pre-Trained Transformer (GPT-3) API was released in June, 2020 by OpenAI. In theory, it could be used as the world’s greatest docent in virtual worlds today.
It is easy to chat with. It is capable of problem solving and ‘original thinking,’ (whatever that is). It responds conversationally, not just spewing factoids.
Oh, and it knows everything.
The simple List of ways a Virtual Docent could help, shown in the pic, was the very beginning of my exploration. The non-bold text was GPT-3’s first crack at working on this project with me.
There’s only a few slight problems as of early January, 2021.
- GPT-3 is an app and it has an application interface. Every Social VR platform is also an app. At the present time, Social VR apps are not open to other application interfaces. That is changing. The Next Big Thing in VR isn’t a headset or a platform — it’s Interoperability.
- GPT-3 costs too much. Current pricing is not a major constraint for most researchers, but wide scale deployment of the app could become very costly, very quickly. That will also change, but in the meantime it is possible that virtual communities could offer the service to paying members.
- GPT-3 can’t process language quite like we do yet. Right now, there is a technique, maybe an art, for prompting GPT-3 to produce its most relevant and creative results.
It is worth understanding why some things GPT-3 comes up with are surprisingly spot-on, and some things are not even close.
GPT-3 is trying to pick up where the last prompt left off and do The Best Next Thing within the context of natural language; ie, GPT-3 can’t respond to a prompt by playing the saxophone. Technically it is a text-completer or an auto-completer. It has to make inferences about what the heck to do next using language. Just like we do.
What’s best varies due to specific local factors not present in the GPT-3 database or the prompt someone provides. How is GPT-3 supposed to know if the best next thing is to wrap things up or keep the ball rolling? To be wildly speculative or very empirical? To answer your question or expand on it?
Those are definitely Points to Ponder even with body language and inflection and maybe even auras to help. GPT-3 can’t process those extra channels, yet, so humans have to adjust a few dials, tuning GPT-3's Complete-Model for the conversation each of us wants to have.
Most of the dials are over on the right in the picture above. Temperature, for example, refers to the level of randomness GPT-3 should allow in the mix. Many other researchers have written about their experiences testing the variables and some have summarized their findings into little Primers.
Pre-sets help GPT-3 know that a certain format is expected, a chat, a list, a Q&A. I am developing a set of Pre-sets to help GPT-3 understand what interaction format to use as The Best Next Thing — an answer, a question, a continuation — when serving as a Virtual Docent in a specific World.
Imagine the Ultimate Docent, who would converse with you and whatever group of friends you bring, if any — and when the time was right the Ultimate Docent would start suggesting a few general ideas, which you’d fine-tune together.
- Games you could play
- Projects you could start
- Stories you could be immersed in
- Stories of your own you could tell about that World
We will have GPT-3 in VR soon and then I can’t predict what will happen.
It’ll be Co-Storytelling, but not like any of the progressive storytelling games you’ve ever played, because with GPT-3 and GPT-4 and GPT-32, our prompts and the right pre-sets will generate the code to build or modify a World and what we can do in it. Immediately. As we think it.
Immersive stories that make themselves ourselves up as we go along.
Stories we are all co-authoring, indistinguishable from what is real.
I write an occasional sub-stack e-news mostly about everything I’m interested in that is not VR. They are short and never boring. Check it out here if you’d like.
Thanks for reading,
Tom
Don’t forget to give us your 👏 !
