Zabaware Forums > Ultra Hal 7.5 Beta - Now powered by OpenAI GPT-3
Ultra Hal 7.5.129 Beta - Now free to use
Medeksza:
--- Quote from: Carl2 on February 26, 2021, 07:51:36 am --- Just a thought that might prove useful, I remember at one time we were trying to get hal to go to lets say wiki, would it be possible to have Hal decide
to go to the cloud for a reply. Probably not as easy as I made it sounds but so much time and effort went into plugins which I had used and I hate to see them all going to waste. Maybe the user could decide and use a trigger word to use the cloud.
Carl2
--- End quote ---
I hope to re-enable the ability to combine plugins with the GPT-3 version of the brain later in the year. Right now a licensing restrictions prevents that ability but things are in motion that I believe it will be possible again later.
Medeksza:
--- Quote from: Dave1152 on February 26, 2021, 01:12:03 pm ---I am curoius as to how exactly you are attempting to impliment gpt 3 with hal, or is it more lile an addon that uses gpt 3 kinda turning off the hal brain while gpt 3 is in use? In my opinion gpt 3 is an impressive thing but in my view its really a sort of Rube Goldberg version of autocomplete. Yes these models have great promise however they are not releasing anything everyone can look at yet and probably wont until gpt 3 is obsolete. If you can see a bigger picture gpt 3 is a small part of a greater set of ideas (algorythims). I am really not a fan of anything that needs to be connected to the internet. I stopped using the newer version of hal when the activation and perodic checkin was requiring me to connect to the internet. I whent back to a version i had on windows xp. When we can see an implimentation of gpt 3 that runs on the individual pc without the need of the internet will greatly improve the convience and utility to the individual user. If they ever release anything usefull like the source code to the api or for that matter allowing us to see the api in general. As to the programming of systems like Hal i do understand the use of high level languages however as an assembly programmer i like the facility and flexability of the newer cisc based processors. When programming for arm i kinda get the 80s esk feeling of programming a commodore 64 :) In the future the ideas presented in Hal and gpt 3 may become baked into operating system distributions as we move into the future it seems that the os may not be just for operating the machine but as an assistive technology for everyday human life. Think about it. How are you using Hal now? As an experiment? An interesting toy for your desktop? Or mabe as a practical tool to aide you in your day to day life like a voice controlled day planner offering theraputic conversation? Software like Hal just scratches the surface of what it could do. In my opinion in order to realise the true potental of newer computers we need to better focus on translating human language into machine code or assembly language much better than we do today. Think about the limits of what you can actually do with a language like C, Python or any of the other flow chart like languages in popular use today and how the computing resources you have are used on your pc. As we see improvements here we will realise improvements in our daily lives. Look at how you use your not so smart phone. If you look at how people are actually using the technology you can see where the operating system of the future mat not just be for operating the machine but for operating your life.
--- End quote ---
Currently GPT-3 is only available as an API call over the internet to servers owned by Microsoft and OpenAI. Hal's GPT-3 powered brain is a hybrid brain that run's a script for simple responses and to establish Hal's personality. It passes off to GPT-3 responses the scripted side doesn't handle. The Hal side of the hybrid system also does simple learning and functions as both the short term and long term memory for GPT-3. GPT-3 by itself has no learning ability and no memory, so Hal fills in that role for it. Hal also tries to guide GPT-3's personality trying to keep it friendly and not get overly opinionated on some taboo subjects.
What's interesting is the source code to GPT-3 is actually posted and available. But the training material and resulting neural network that was generated from that material is under control of OpenAI and Microsoft. There are other groups that are creating their own versions of GPT-3 that should become available later this year. It takes about $2 million dollars of computer hardware (GPU heavy) to train a GPT-3 like brain and then takes about a $200,000 computer (also GPU heavy) to run the resulting neural network once its trained. So still a bit too expensive for people to run the brain on their home computers, so it will be cloud based for a while I think.
Dave1152:
Where is it posted at? Every time i goto open ai web site or git hub theres a waiting list. From my understanding of it gpt2 is the same as gpt 3 with just a larger model used. Gpt 2 can run on a pc.
Medeksza:
--- Quote from: Dave1152 on February 27, 2021, 09:05:42 pm ---Where is it posted at? Every time i goto open ai web site or git hub theres a waiting list. From my understanding of it gpt2 is the same as gpt 3 with just a larger model used. Gpt 2 can run on a pc.
--- End quote ---
Correct, the source code of GPT-2 and GPT-3 are exactly the same. The training data of GPT-3 is much larger and the resulting trained neural net is much larger. OpenAI published the source code, but not the training data or trained neural net.
There is a group at https://eleuther.ai who took the source code and created their own training dataset a little larger than GPT-3 with some improvements. They are currently training it and maybe by August will complete it and will publish both the training data and trained model.
Carl2:
I'll just put this in which may be helpful, while looking at different things GP3 can do I came across GPT3 can copy script to a file that has been given a name so that info has not been used and lost.
Carl2
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version