Zabaware Support Forums

Zabaware Forums => Ultra Hal 7.5 Beta - Now powered by OpenAI GPT-3 => Topic started by: Spitfire2600 on February 01, 2021, 05:03:23 am

Title: This. Is. Amazing.
Post by: Spitfire2600 on February 01, 2021, 05:03:23 am
Robert, I love this.

How did you make this possible with VBScript? This is always where I thought Hal should go, it's a totally different experience. I will write again after I play around with my Hal a little more but right off the bat, the conversations are clear, concise, and Hal always replies with a relevant response. I haven't checked the Q/A abilities, but I feel on top of conversations that Hal should be able to answer most questions, given this API. AND!
As far as I am aware the GPT-3 is capable of low level programming advice, meaning we could use Hal to make Hal even better. This would help plug-in creators as well, theoretically, assuming the neural net is aware of how UHP files or VB are set up. Please look into this as well if you get a chance. This could help so many people.

The downside is paying for the responses and the lack of customization, I know this is a Beta. But I wish/hope even the local brains could use this neural net. (I'm totally aware of the computational limitations) But I can dream. Perhaps not the entire neural net, but is it possible to eventually create some miniaturized form of this for the offline Hal? GPT-2 was opensource, I think? (Unaware of the computing power needed for that one)

How is the Neural Net set up, with one big online Hal brain, or does each user have their own neural net per online brain? Have you fed the massive online conversational database to the neural net? Is our online neural net explorable in anyway, via the Brain files or UHP files?

I'll check prices and chat some more, but this has REAL potential, this could be more than just a chatbot, Robert, this is a life assistant possibly.

Thank you so much for your work here Robert!!! Long Live HAL!

Spitfire_2600
Title: Re: This. Is. Amazing.
Post by: Medeksza on February 01, 2021, 02:00:19 pm
Glad you love it so far!

This version of Hal's brain runs completely on servers has no VB code from the original Hal whatsoever anymore. It's currently just an API call from the Hal app to the servers running it.

Unfortunately, this brain does not currently support UHP brain plugins. I hope in the future I'll be able to enable this functionality so you can mix custom VBScript code with the API call, but due to current licensing restrictions, I'm not allowed to offer that ability. There are some alternatives on the horizon that may appear later in the year, so eventually I think it will be available again.

The GPT3 based brain learns to the same database as the Hal 7.0 cloud brain, so you can easily switch between the normal Hal 7.0 brain (which can be scripted) and the GPT3 powered brain. Hal 7.0 can benefit from info learned while talking with a GPT3 powered brain.

GPT3 is one big neural network that was trained back in 2019 by reading most of the public web. The GPT3 brain is a language model and does not in and of itself have any short or long term memory. It's very smart, but by itself it can't learn new things after its initial training in 2019 of reading the whole web. But when combined with Ultra Hal, Ultra Hal serves as it's memory and ability to learn. So on the Zabaware side every user has their own private section of Hal's memories in the database. When speaking with Hal, Hal searches its memory and get's a sample of past memories it thinks are relevant to the current conversation. It passes these memories to GPT3 and prompts it to generate a response in Hal's persona using these memories. So it's definitely a hybrid system. Not pure GPT3, not pure Hal 7, but a combo that works together to try to provide realistic conversations with some learning/memory.

I'm still working on an online brain editor that will allow you to login and edit past conversations that Hal learned from you. Might be available in a month or two. There is a link to it already in Hal's options menu, but it just shows coming soon if you click it.
Title: Re: This. Is. Amazing.
Post by: pacslim on May 01, 2022, 08:50:37 pm
Glad you love it so far!

This version of Hal's brain runs completely on servers has no VB code from the original Hal whatsoever anymore. It's currently just an API call from the Hal app to the servers running it.

Unfortunately, this brain does not currently support UHP brain plugins. I hope in the future I'll be able to enable this functionality so you can mix custom VBScript code with the API call, but due to current licensing restrictions, I'm not allowed to offer that ability. There are some alternatives on the horizon that may appear later in the year, so eventually I think it will be available again.

The GPT3 based brain learns to the same database as the Hal 7.0 cloud brain, so you can easily switch between the normal Hal 7.0 brain (which can be scripted) and the GPT3 powered brain. Hal 7.0 can benefit from info learned while talking with a GPT3 powered brain.

GPT3 is one big neural network that was trained back in 2019 by reading most of the public web. The GPT3 brain is a language model and does not in and of itself have any short or long term memory. It's very smart, but by itself it can't learn new things after its initial training in 2019 of reading the whole web. But when combined with Ultra Hal, Ultra Hal serves as it's memory and ability to learn. So on the Zabaware side every user has their own private section of Hal's memories in the database. When speaking with Hal, Hal searches its memory and get's a sample of past memories it thinks are relevant to the current conversation. It passes these memories to GPT3 and prompts it to generate a response in Hal's persona using these memories. So it's definitely a hybrid system. Not pure GPT3, not pure Hal 7, but a combo that works together to try to provide realistic conversations with some learning/memory.

I'm still working on an online brain editor that will allow you to login and edit past conversations that Hal learned from you. Might be available in a month or two. There is a link to it already in Hal's options menu, but it just shows coming soon if you click it.

Is this now an option in hal or still coming in 2022?