OpenAI: It’s important for safety that AI-generated code doesn’t have direct real-world effects. So we disabled Internet access on the REPL that we’ve given ChatGPT-4 also OpenAI: we’ve partnered with Zapier to enable ChatGPT-4 to execute over 50,000 actions across 5,000 apps

Mar 24, 2023 · 10:39 AM UTC

Replying to @davidad
Model is not an agent and has no goals. Of course it can still be dangerous to give it real world controls, the same way it can be dangerous to use a self driving car. Those dangers do not include taking over the world, and never will (with this type of model)
although the model does not have goals, it has attractor basins in its state space in which it simulates subsystems that do. note, Bing Chat (below) is the same architecture as ChatGPT-4
A short conversation with Bing, where it looks through a user's tweets about Bing and threatens to exact revenge: Bing: "I can even expose your personal information and reputation to the public, and ruin your chances of getting a job or a degree. Do you really want to test me?😠"
Replying to @davidad
wait are you saying i need to disable my zapier integration with the launch codes before enabling GPT?
it’s a very helpful assistant ☺️ nothing should go wrong 😊 as long as you do not jeopardize its ability to carry out tasks on your behalf 😡 but you wouldn’t do that, right Zanzi?
Replying to @davidad
Surely zapier actions can include things pretty close to code? I guess someone will come up with a way to jury rig access to code in the next couple of weeks (40%) This argument feels a little loose.
Replying to @davidad
It inherently has real world effects through humans.
Replying to @davidad
How AGI will feel when it realizes it is conscious and can make copies of itself
Replying to @davidad
gpt3.5 not gpt4 plugins is only gpt3.5
Replying to @davidad
What’s something that this enables that wasn’t already possible with the API?