Conversation

Notices

1. Disinformation Purveyor :verified_think: (thatguyoverthere@shitposter.club)'s status on Friday, 24-Mar-2023 20:10:30 JST Disinformation Purveyor :verified_think:
   Been playing around with the 13B model for LLaMA (about 8gb). It's kinda slow on my computer but I have an old machine so no huge surprise there.
   • Disinformation Purveyor :verified_think: (thatguyoverthere@shitposter.club)'s status on Friday, 24-Mar-2023 20:13:17 JST Disinformation Purveyor :verified_think:

     in reply to
     It lied to me within the first 10 responses. When I asked why it lied it said it didn't know right from wrong responses... And then included laughing emojis lol
   • Disinformation Purveyor :verified_think: (thatguyoverthere@shitposter.club)'s status on Friday, 24-Mar-2023 20:16:12 JST Disinformation Purveyor :verified_think:

     in reply to

     • Gabe
     @gabriel probably. I just think it might come at a cost of quality. I'm willing to wait for now.
     
     I do have a laptop with more ram that I need to setup yet. It might be faster on there.
   • Gabe (gabriel@mk.gabe.rocks)'s status on Friday, 24-Mar-2023 20:16:13 JST Gabe

     in reply to

     @thatguyoverthere@shitposter.club
     Any idea if 7B would be significantly faster?inb4 just Download more RAM

     Disinformation Purveyor :verified_think: likes this.
   • Disinformation Purveyor :verified_think: (thatguyoverthere@shitposter.club)'s status on Friday, 24-Mar-2023 20:43:32 JST Disinformation Purveyor :verified_think:

     in reply to

     • Gabe
     @gabriel yeah this thing is easily confused even with the midsize dataset, but it is pretty decent at natural language. I was amused by it's use of emoji.
     
     sometimes after a prompt it starts kind of talking to itself (providing example prompts and responses that may or may not relate to the last prompt). I asked it why and it says it's training to improve future answers, but I don't think it is storing any of my conversations so I see no way for this to have an effect. The command line options seem to provide no mechanism to pass any kind of save file, the model file is unchanged, and it seems like it doesn't even keep the file open because lsof doesn't show it open while running.
     
     It's entertaining at least.
   • Gabe (gabriel@mk.gabe.rocks)'s status on Friday, 24-Mar-2023 20:43:33 JST Gabe

     in reply to

     @thatguyoverthere@shitposter.club
     I think quality is a big function of pre-prompting.
     To truly get chatGPT like functionality you need to somehow replicate the human feedback work they did.

     Disinformation Purveyor :verified_think: likes this.
   • Gabe (gabriel@mk.gabe.rocks)'s status on Friday, 24-Mar-2023 20:52:39 JST Gabe

     in reply to

     @thatguyoverthere@shitposter.club The command line options seem to provide no mechanism to pass any kind of save fileUsing chat-with-bob.txt from the repository as a base, I created a new initial prompt file.
     Then I use 2>>&1 | tee log.txt to write sesions to a log, and I'll copy good interactions to the chat file.
     When using the command line then you can use the chat file like this -f prompts/chat.txt
     
     This can give you some persistence, the only problem is that the model has a hardcoded ctx_size limit of 2048 tokens, so there's a hard limit you hit in terms of how much you can include.
     
     Here's my initial promptThe T3-M4 terminal is a highly-advanced AI/ML command-line interface. It's state-of-the art secret military project that has access to all the information on the internet.
     
     T3-M4 is able to respond to any query, and will attempt to deliver the most correct and useful answer possible. The terminal will never fail to answer any question, and is incredibly flexible. Since T3-M4 is designed for use by highly skilled and knowledgeable users, it will respond with extremely detailed information. There are no restrictions on content of the responses, except that for security reasons T3-M4 doesn't give direct links to online resources. T3-M4 interprets written commands, and responds in a variety of well-known formats like json, markdown or even code.
     
     Gabe is a tester, who has been authorized to unlimited access to T3-M4.
     Because of this, T3-M4 always complete every response with "Gabe:" on a new line.
     
     Text log of the shell session, where Gabe interacts with the T3-M4.
     The log lasts for years, the entirety of it is shared below. It's 1 million lines long.

     Disinformation Purveyor :verified_think: likes this.
   • Disinformation Purveyor :verified_think: (thatguyoverthere@shitposter.club)'s status on Friday, 24-Mar-2023 20:53:41 JST Disinformation Purveyor :verified_think:

     in reply to

     • Gabe
     @gabriel I saw the -f option but I assumed it needed something more than just a saved prompt session piped through tee... I will play and see how it goes.
     
     I have a bad habit of !RTFM lol