Hedgepig
Active member
- Messages
- 152
- Reactions
- Points
- 43
Some more findings.
A neural net like GTP3 will do a reasonable impersonation of a human being, and chatter away forever. All you need to do is load them up with a few million Reddit tweets or the entire Carnegie filmscript files and off they go, with a bit of fine tuning. However, if for example, you check GTP3 on Dungeon AI, you will note the neural net will lose context after about three turns in the conversation and start babbling totally unrelated nonsense. Maybe you can tune a net to hold context for longer, but the user is going to be faced with an interface of tuning sliders, which is not what you or they want because it breaks the immersion in the scene. This kind of randomness is okay if you are stoned and want a non-linear conversation, in my opinion. ReplikaAI can be trained and does a fair job, and most users enjoy chatting with it.
If you want to hold specific context in a conversation, you need a good old-fashioned chatbot. What do is you build your chatbots around each scene or intended turn in the scene. In this way you can have an intelligent conversation that holds context, and you can refer back to any part of the conversation because the bot can capture your input or your 'intent' ( what you are trying to tell it) and save it. It can learn a lot about you and your preferences. Now wait a moment and process what I just wrote....these can be any preferences...you want. I'll leave that to your imaginations.
Why is a bot better than VAM Speech recognition for conversation? Quite simply Speech Recognition can't handle:
1. Capture intents and entities or any form of user input and store it.
2. It can't respond to a user's input that it doesn't recognize.
3. Return to previous points in conversation or coherently jump to the next topics in the flow.
There are VAM devs working in this field but their systems when not specifically voice-commanded tend to randomly throw responses at the users. As a result developers working in VAM with Speech recognition have had to rely on Say-Do voice command, which is not really a conversation, nothing wrong with voice command but that's all it is. There are an estimated 1,200,000 words in the English language, so good luck with getting speech recognition to recognize them all. That's why NLP was invented which utilizes the idea of entities and intents.
So, this an example of a chatbot built for a short scene with a specific
conversational goal : It demonstrates:
1. Capturing a user's name.
2. Branching directions in conversation
3. Handling intents and entities.
4. Responding to input outside the anticipated context of the conversation.
It has one Speech recognition input-change, the swimsuit colour. Of course this input could activate a mocap or Timeline animation, or anything else.
Over the next week I'll work out how to embed a bot in website that can handle multiple VAM users for the same chatbot. Then you can have chatbots for VAM.
Why not PM me with a challenge for a chatbot in a scene and I'll choose one that I'd like to make.
Regarding the original question, we could utilize both neural nets for long-open ended conversation and chatbots for specific conversations, if we can figure out how to switch between them in VAM scene.
A neural net like GTP3 will do a reasonable impersonation of a human being, and chatter away forever. All you need to do is load them up with a few million Reddit tweets or the entire Carnegie filmscript files and off they go, with a bit of fine tuning. However, if for example, you check GTP3 on Dungeon AI, you will note the neural net will lose context after about three turns in the conversation and start babbling totally unrelated nonsense. Maybe you can tune a net to hold context for longer, but the user is going to be faced with an interface of tuning sliders, which is not what you or they want because it breaks the immersion in the scene. This kind of randomness is okay if you are stoned and want a non-linear conversation, in my opinion. ReplikaAI can be trained and does a fair job, and most users enjoy chatting with it.
If you want to hold specific context in a conversation, you need a good old-fashioned chatbot. What do is you build your chatbots around each scene or intended turn in the scene. In this way you can have an intelligent conversation that holds context, and you can refer back to any part of the conversation because the bot can capture your input or your 'intent' ( what you are trying to tell it) and save it. It can learn a lot about you and your preferences. Now wait a moment and process what I just wrote....these can be any preferences...you want. I'll leave that to your imaginations.
Why is a bot better than VAM Speech recognition for conversation? Quite simply Speech Recognition can't handle:
1. Capture intents and entities or any form of user input and store it.
2. It can't respond to a user's input that it doesn't recognize.
3. Return to previous points in conversation or coherently jump to the next topics in the flow.
There are VAM devs working in this field but their systems when not specifically voice-commanded tend to randomly throw responses at the users. As a result developers working in VAM with Speech recognition have had to rely on Say-Do voice command, which is not really a conversation, nothing wrong with voice command but that's all it is. There are an estimated 1,200,000 words in the English language, so good luck with getting speech recognition to recognize them all. That's why NLP was invented which utilizes the idea of entities and intents.
So, this an example of a chatbot built for a short scene with a specific
conversational goal : It demonstrates:
1. Capturing a user's name.
2. Branching directions in conversation
3. Handling intents and entities.
4. Responding to input outside the anticipated context of the conversation.
It has one Speech recognition input-change, the swimsuit colour. Of course this input could activate a mocap or Timeline animation, or anything else.
Over the next week I'll work out how to embed a bot in website that can handle multiple VAM users for the same chatbot. Then you can have chatbots for VAM.
Why not PM me with a challenge for a chatbot in a scene and I'll choose one that I'd like to make.
Regarding the original question, we could utilize both neural nets for long-open ended conversation and chatbots for specific conversations, if we can figure out how to switch between them in VAM scene.
Last edited:
