I thought you had replied in that philosopher AI thread.
No. I replied only here.
What do you got on this philosopher AI
Philosopher AI is not a philosopher and not an AI. It is interface to simple neural network that is pre-teached to predict next word in a sentence. It is just a text syntesizer. It is trained on many texts and result of that training named GPT-3. OpenAI (not open and not AI) sell this closed pre-trained model for money using modern hype about AI.
What's this GPT-J 6B?
There are another Generative Pre-trained Transformer (GPT) models, open and free, f.e. GPT-2 and GPT-J you could easily play with. GPT-J 6B is larger than GPT-3. GPT-J 6B is a GPT-J transform model that has 6 billion trainable parameters. https://huggingface.co/EleutherAI/gpt-j-6B
I'd like to use something where it wouldn't say "nonsense" all the time.
You have to spend some time to get something not "nonsense" from any GPT. Examples that you saw about Philosopher AI (GPT-3) is just a little part selected from a huge pile of nonsense answers. Marketing, you know...
Coefficients in neural network. Hard to tell what they adjust exactly, but you are free to fine tune them and see what happens. There are tooltips if you hover cursor over the field, but they are not very informative, except recommended ranges for the value.
"seed" is a number used to "seed" neural network. If set to 0 it will use random seed for every pass. You will always get different completion texts. If you set it to some numeric value, it will use that value every time and you will always get same text for same phrase. If you want to play with other settings, set seed to some value, to view how settings will change output, otherwise you will get different texts each time and will not notice changes coming from changing other settings.
Where it seemed to have access to government 3 letter agency databases.
It does not have access to anything, really.
GPT just try to continue text in the way it trained. That's all. It's like somebody read all texts on Earth and figure out the most probable words sequences that should follow some already existing sequence. Nothing more, nothing less. There are no any texts at all, just, say, a list of all words and neural network coefficients that select next words sequence using previous sequence as input.
There are no any magic or even some complex things. It's just a simple pre-trained neural network that sets output according to input as trained.
You want a site that an idiot can type stuff in, and doesn't have to do code.
Here it is - https://bellard.org/textsynth/
Just type in anything you want and press "Complete text". That's all. If you want go further, you could play with parameters Fabrice made changeable. You will get different results with different settings.
is it going to be the same databases as that one philosopher AI was using?
It depends are OpenAI lied about source or not. If they tell a truth, both GPT-2 and GPT-J trained on the same data. It's hard to check, really.
The thing was really fishy because it would say "sensitive information"
Really it is hard to tell anything about that information. AFAIK, "The Pile" which was used to train GPTs is a 800Gb of different texts, including SciFi and other things. So, when GPT tries to complete text, it could easily generate a sequence that is trained from SciFi or Conspiracy books, or any other text that get into "The Pile", including possible confidential documents or their parts from books.
You can download "The Pile" and try to check it by yourself, but it is >800Gb, so it could take time. :)
I would do that and it halted, said: "quota exceeded for free access (max 90 requests per 30 minutes)"
Just wait 30 minutes, and continue. Be a nice user, don't overload Fabrice's server, he is not a millionare to buy a server farm for you.
This thing really doesn't want to tell you what folder and file names it sees in whatever database though.
It can't.
All that GPT, including PhilosopherAI is just a fancy toy, nothing more. It just try to continue what you write using the words that is most probable continuation of previous text. Simple like that. PhilosopherAI does not have access to any databases. It is not an AI. It just look for most probable words or sentencies to continue previous text using some kind of calculated weights for words it know.
GPTs is funny, but completely useless to get some info. Best use for them is to help some incompetent writer to create some senseless article. Say, you need to write an ad of 1000 words for some product, but you don't know what to write. Start with "<Product name> is outstanding new something" and click a button. And finally you will get some ad text you probably could sell to your dumb boss as your work. It will be posted on company site to make it look updated and smart and everybody will forget about it. That is the only possible purpose of all that simple neural networks trained on literature and internet content. OpenAI company try to make money selling it to companies that will create sensless ad articles for their sites without a need to hire a writer.
No. I replied only here.
Philosopher AI is not a philosopher and not an AI. It is interface to simple neural network that is pre-teached to predict next word in a sentence. It is just a text syntesizer. It is trained on many texts and result of that training named GPT-3. OpenAI (not open and not AI) sell this closed pre-trained model for money using modern hype about AI.
There are another Generative Pre-trained Transformer (GPT) models, open and free, f.e. GPT-2 and GPT-J you could easily play with. GPT-J 6B is larger than GPT-3. GPT-J 6B is a GPT-J transform model that has 6 billion trainable parameters. https://huggingface.co/EleutherAI/gpt-j-6B
You are not limited in length here - https://bellard.org/textsynth/
You have to spend some time to get something not "nonsense" from any GPT. Examples that you saw about Philosopher AI (GPT-3) is just a little part selected from a huge pile of nonsense answers. Marketing, you know...
Coefficients in neural network. Hard to tell what they adjust exactly, but you are free to fine tune them and see what happens. There are tooltips if you hover cursor over the field, but they are not very informative, except recommended ranges for the value.
"seed" is a number used to "seed" neural network. If set to 0 it will use random seed for every pass. You will always get different completion texts. If you set it to some numeric value, it will use that value every time and you will always get same text for same phrase. If you want to play with other settings, set seed to some value, to view how settings will change output, otherwise you will get different texts each time and will not notice changes coming from changing other settings.
It does not have access to anything, really. GPT just try to continue text in the way it trained. That's all. It's like somebody read all texts on Earth and figure out the most probable words sequences that should follow some already existing sequence. Nothing more, nothing less. There are no any texts at all, just, say, a list of all words and neural network coefficients that select next words sequence using previous sequence as input.
There are no any magic or even some complex things. It's just a simple pre-trained neural network that sets output according to input as trained.
Here it is - https://bellard.org/textsynth/ Just type in anything you want and press "Complete text". That's all. If you want go further, you could play with parameters Fabrice made changeable. You will get different results with different settings.
It depends are OpenAI lied about source or not. If they tell a truth, both GPT-2 and GPT-J trained on the same data. It's hard to check, really.
Really it is hard to tell anything about that information. AFAIK, "The Pile" which was used to train GPTs is a 800Gb of different texts, including SciFi and other things. So, when GPT tries to complete text, it could easily generate a sequence that is trained from SciFi or Conspiracy books, or any other text that get into "The Pile", including possible confidential documents or their parts from books. You can download "The Pile" and try to check it by yourself, but it is >800Gb, so it could take time. :)
Just wait 30 minutes, and continue. Be a nice user, don't overload Fabrice's server, he is not a millionare to buy a server farm for you.
It can't.
All that GPT, including PhilosopherAI is just a fancy toy, nothing more. It just try to continue what you write using the words that is most probable continuation of previous text. Simple like that. PhilosopherAI does not have access to any databases. It is not an AI. It just look for most probable words or sentencies to continue previous text using some kind of calculated weights for words it know.
GPTs is funny, but completely useless to get some info. Best use for them is to help some incompetent writer to create some senseless article. Say, you need to write an ad of 1000 words for some product, but you don't know what to write. Start with "<Product name> is outstanding new something" and click a button. And finally you will get some ad text you probably could sell to your dumb boss as your work. It will be posted on company site to make it look updated and smart and everybody will forget about it. That is the only possible purpose of all that simple neural networks trained on literature and internet content. OpenAI company try to make money selling it to companies that will create sensless ad articles for their sites without a need to hire a writer.