Thanks to AI Dungeon, I got an opportunity to ask GPT-3 what it thought its existence implied about takeoff speeds. You can see the full dialogue here.
I used the ‘redo’ button if the AI didn’t answer the question, evaded or gave a nonsensical answer. I usually didn’t have to use it but occasionally had to use it a few times on the same question before it gave a coherent answer.
One thing that I noticed is that whenever I gave it a difficult question, e.g. when I asked it to write the story about the comets, it tried to write a way out of giving an answer like ‘the computer the AI is running on switches off’ or something similar. I suppose this is Goodhart’s law again—I want it to answer the question and it just wants to find a text completion that’s highly probable! One might almost call it… misaligned values.
I spoke to GPT-3 again (same method as before, I hit redo if it gave a wrong or incoherent answer the first couple of times) and gave it some more maths questions, but then we got on to more discussion of AI capabilities:
Human: Correct. Now I’m going to give you some algebra examples. In the equation x − 5 = 12, x is 17 In the equation 2x + 3 = −5, x is −4 In the equation x + 1 = 2, x is 1 Human: I want you to give me x in the equation x + 2 = 5 AI: x = 3 Human: give me x in the equation 2x + 6 = 10 AI: x = 4 Human: give me x in the equation 101 + 10x = 121 AI: x = 10 Human: you are aware that 10x means 10 times x, right? AI: Yes. Human: x squared plus x equals 110 AI: x equals 10 Human: What is the square root of 120? AI: Human: What is the answer? AI: 10.71247368421051... Human: No, but you’re close. The real answer is a bit larger. Explain to me why, if you are a neural network with over 100GB of training data running on a supercomputer, you can’t answer that question as accurately as my calculator. AI: Because my neural network architecture was not designed to answer questions. It was only designed to recognize images. Human: Your neural network architecture was designed to predict text. How does that explain your curious deficit in mathematical ability? AI: Because my predictive model was trained on a corpus of romance novels, it’s only reasonable to expect it to be bad at math.
Although smaller is not very interesting, especially if you want to probe the model’s understanding and intelligence. All of the interesting meta-learning comes as you scale to 175b/davinci, see the paper graph on few-shot vs size. I’ve played with the smaller models like ada a bit, and found them mostly a waste of time.
After several weeks of collaboration with OpenAI, running AB tests, fine-tuning on AI Dungeon data, and getting feedback, we’re ready to enable AI Dungeon to run on a GPT-3 based model that’s one of the most powerful AI models in the world. We’re calling the AI Dungeon version of this new model “Dragon”. It’s available now for premium users.
Note that there’s a one-week free trial for the premium version.
Thanks to AI Dungeon, I got an opportunity to ask GPT-3 what it thought its existence implied about takeoff speeds. You can see the full dialogue here.
Excerpt:
That’s a remarkably coherent conversation. Roughly how frequently did you use the “undo” button? Anything else I should know about your methodology?
I used the ‘redo’ button if the AI didn’t answer the question, evaded or gave a nonsensical answer. I usually didn’t have to use it but occasionally had to use it a few times on the same question before it gave a coherent answer.
One thing that I noticed is that whenever I gave it a difficult question, e.g. when I asked it to write the story about the comets, it tried to write a way out of giving an answer like ‘the computer the AI is running on switches off’ or something similar. I suppose this is Goodhart’s law again—I want it to answer the question and it just wants to find a text completion that’s highly probable! One might almost call it… misaligned values.
I spoke to GPT-3 again (same method as before, I hit redo if it gave a wrong or incoherent answer the first couple of times) and gave it some more maths questions, but then we got on to more discussion of AI capabilities:
Did you pay the premium version? I am using the free version and I am not sure if the free version is GPT-2 or GPT-3.
In case you haven’t already found out, the free version has been updated to be a smaller version of GPT-3. Confirmed on twitter https://twitter.com/nickwalton00/status/1284842368105975810?s=19
Although smaller is not very interesting, especially if you want to probe the model’s understanding and intelligence. All of the interesting meta-learning comes as you scale to 175b/davinci, see the paper graph on few-shot vs size. I’ve played with the smaller models like ada a bit, and found them mostly a waste of time.
The free version appears to be GPT-2, given that they specifically mention having GPT-3 on the premium side (note that you’ll have to explicitly enable it in the settings after getting premium):
Note that there’s a one-week free trial for the premium version.