I’ve been interacting with OpenSim bots — or NPCs — for almost so long as I’ve been protecting OpenSim. Which is about 15 years. (Oh my God, has it actually been that lengthy?)
I’ve been hoping that OpenSim writing would change into by day job, however, sadly, OpenSim by no means actually took off. As an alternative, I coated cybersecurity and, extra not too long ago, generative AI.
However then I noticed some reporting a few new research about AI, and instantly thought — this might actually be one thing in OpenSim.
The research was printed this previous April within the journal Neuroscience of Consciousness, and it confirmed {that a} majority of individuals – 67% to be actual – attribute some extent of consciousness to ChatGPT. And the extra individuals use these AI techniques, the extra possible they’re to see them as aware entities.
Then, in Might, another study confirmed that 54% of individuals, after a dialog with ChatGPT, thought it was an actual individual.
Now, I’m not saying that OpenSim grid homeowners ought to run out and set up a bunch of bots on their grids that faux to be actual individuals, as a way to lure in additional customers. That might be dumb, costly, a waste of assets, presumably unlawful and positively unethical.
But when customers knew that these bots had been powered by AI and understood that they’re not actual individuals, they could nonetheless get pleasure from interacting with them and develop attachments to them — similar to we get connected to manufacturers, or cartoon animals, or characters in a novel. Or, sure, digital girlfriends or boyfriends.
Within the video under, you may see OpenAI’s latest GPT-4o presentation. Yup, the one the place ChatGPT sounds suspiciously like Scarlett Johansson in “Her.” I’ve set it to begin on the level within the video the place they’re speaking to her.
I can see why ScarJo obtained upset — and why that exact voice is not obtainable as an choice.
Now, as I write this, the voice chatbot they’re demonstrating isn’t broadly obtainable but. However the textual content model is — and its the textual content interface that’s commonest in OpenSim anyway.
GPT-4o does price cash. It costs money to send it a question and to get a response. One million tokens price of questions — or 750,000 phrases — prices $5, and 1,000,000 token’s price of response prices $15.
A web page of textual content is roughly 250 phrases, so 1,000,000 tokens is about 3,000 pages. So, for $20, you will get a variety of back-and-forth. However there are additionally cheaper platforms.
Anthropic’s Claude, for instance, which has examined higher than ChatGPT in some benchmarks, costs a bit less — $3 for 1,000,000 enter tokens, and $15 for 1,000,000 output tokens.
However there are additionally free, open-source platforms that you simply run by yourself servers with comparable efficiency ranges. For instance, on the LMSYS Chatbot Area Leaderboard, OpenAI’s GPT-4o in in first place with a rating of 1287, Claude 3.5 Sonnet is shut behind with 1272, and the (principally) open supply Llama 3 from Meta will not be too far distant, with a rating of 1207 — and there are a number of different open supply AI platforms on the prime of the charts, together with Google’s Gemma, NVIDIA’s Nemotron, Cohere’s Command R+, Alibaba’s Qwen2, and Mistral.
I can simply see an OpenSim internet hosting supplier including an AI service to their package deal offers.
Think about the potential for creating actually immersive experiences in OpenSim and different digital environments. If customers are predisposed to see AI entities as aware, we may create non-player characters that really feel extremely actual and responsive.
This might revolutionize storytelling, training, and social interactions in digital areas.
We may have bots that customers can kind significant relationships with, AI-driven characters that may adapt to particular person person preferences, and digital environments that really feel alive and dynamic.
After which there’s the potential for interactive storytelling and video games, with quests and narratives which can be extra participating than ever earlier than, create digital assistants that really feel like true companions, and even construct communities that blur the traces between AI and human contributors.
For these utilizing OpenSim for work, there are additionally purposes right here for enterprise and training, within the type of AI tutors, AI government assistants, AI gross sales brokers, and extra.
Nevertheless, as a lot as I’m thrilled by these potentialities, I can’t assist however really feel a twinge of concern.
Because the research authors level out, there are some dangers to AIs that really feel actual.
First, there’s the danger of emotional attachment. If customers begin to view AI entities as aware beings, they could kind deep, probably unhealthy bonds with these digital characters. This might result in a variety of points, from social isolation in the actual world to emotional misery if these AI entities are altered or eliminated.
We’re already seeing that, with individuals feeling actual misery when their digital girlfriends are turned off.
Then there’s the query of blurred actuality. As the road between AI and human interactions turns into much less clear, customers would possibly wrestle to tell apart between the 2.
Personally, I’m not too involved about this one. We’ve had individuals complaining that different individuals couldn’t inform fantasy from actuality because the days of Don Quixote. In all probability even earlier. There have been most likely cave individuals sitting round, saying, “Take a look at the younger individuals with all their cave work. They could possibly be out truly looking, and as an alternative they sit across the cave trying on the work.”
And even earlier, when language was invented. “Take a look at these younger individuals, sitting round speaking about looking, as an alternative of going on the market into the jungle and catching one thing.”
When motion pictures had been first invented, when individuals began getting “addicted” to tv, or video video games… we’ve at all times had ethical panics about new media.
The factor is, these ethical panics had been additionally, to some extent, justified. Perhaps the pulp novels that the printing press gave us didn’t rot our brains. However Mao’s Little Pink E book, the Communist Manifesto, that factor that Hitler wrote that I don’t even was aided and abetted by the books they wrote.
In order that’s what I’m most fearful about — the potential for exploitation. Dangerous actors may misuse our tendency to anthropomorphize AI, creating misleading or manipulative experiences that make the most of customers’ emotional connections and cause them to be extra tolerant of evil.
However I don’t assume that’s one thing that we, in OpenSim, have to fret about. Our platform doesn’t have the type of attain it will take to create a brand new dictator!
I feel the worst that will occur is that folks would possibly get so engaged that they spend a number of {dollars} greater than they deliberate to spend.
