GPT-3 based Signal bot
In experimenting with the OpenAI API, I developed a little bot for the Signal Messenger app.
To build the bot, I leveraged the semaphore framework, and set up the signald system using a spare phone number (see footnote 1).
After testing with the echo bot example, which replies to every message with the same message, I then developed my own test bot.
It reacts to every message with a 🌯 emoji.
You can see the code below.
import os
from semaphore import Bot, ChatContext
# Connect the bot# to number.
bot = Bot("+xxxxxx")
async def love(ctx: ChatContext) -> None:
await ctx.message.reply(body="🌯", reaction=True)
async def main() -> None:
"""Start the bot."""
# Connect the bot to number.
async with bot:
bot.register_handler("", love)
# Set profile name.
await bot.set_profile("PG GPT3 test bot")
# Run the bot until you press Ctrl-C.
await bot.start()
if __name__ == "__main__":
import anyio
anyio.run(main)
Next, I developed the GPT-3 bot, leveraging the OpenAI Python library.
My code is below, however it is fairly simple.
It checks if a given message starts with the string "!GPT", and if so, passes the text to the GPT-3 completion model.
I also used the quote=True flag to have the bot reply to the message that made the request.
I also added a $10 hard limit on API requests, since I am testing it out with friends and do not want to go broke.
I have the code running on an old Raspberry Pi 3.
#!/usr/bin/env python
import anyio
from semaphore import Bot, ChatContext
import openai
# Connect the bot# to number.
bot = Bot("+xxxxxxxxx")
@bot.handler("") #
async def echo(ctx: ChatContext) -> None:
msg = ctx.message.get_body()
# check if the message starts with "!GPT", return if not
if not msg.startswith("!GPT"):
return
# remove "!GPT" start from the message
msg = msg[5:]
# pass the message to the OpenAI API
response = openai.Completion.create(
engine="davinci",
prompt=msg,
max_tokens=50,
temperature=0.9,
top_p=1,
n=1,
stream=False,
logprobs=None,
stop=["\n"],
user="signal-bot-1",
)
new_msg = msg + response["choices"][0]["text"]
# reply with the text from the OpenAI API
await ctx.message.reply(new_msg, quote=True)
async def main():
async with bot:
# Set profile name.
await bot.set_profile("PG GPT3 test bot")
# Run the bot until you press Ctrl-C.
await bot.start()
anyio.run(main)
There is of course a lot more scope for these sorts of bots, but I’ll leave this code for anyone that wants to get more involved.
When I undertook my project for an AI-generated music video, one of the most important things I found in getting a good output was “prompt design”, i.e. passing the model text in a format that is likely to give a good response. Therefore, to get a better response, I made another bot that reads text from the user, and puts it in a format as if it is a chat log. For example, if you message with “Hello there”, the prompt could be:
Human: Hello there
AI:
Then the model has context clues that we are in an instant messaging situation, and it is to respond as if it is an AI talking to a human. Here’s some code to do this:
@bot.handler("") #
async def echo(ctx: ChatContext) -> None:
msg = ctx.message.get_body()
# TODO check if the message starts with "@GPT", return if not
if not msg.startswith("@GPT"):
return
await ctx.message.typing_started()
# remove "@GPT" start from the message
msg = msg[5:]
text = f"Human: {msg}\nAI: "
# pass the message to the OpenAI API
response = openai.Completion.create(
engine="davinci",
prompt=text,
max_tokens=50,
temperature=0.9,
top_p=1,
n=1,
stream=False,
logprobs=None,
stop=["\n"],
user="signal-bot-1",
)
new_msg = response["choices"][0]["text"]
# reply with the text from the OpenAI API
await ctx.message.reply(new_msg, quote=True)
await ctx.message.typing_stopped()
Here’s a sample chat:
Me: @GPT awright mate how ye daein'?
Bot: Very well, and you?
That’s all for now folks.