Comment by noduerme
4 years ago
I'm not convinced that it's not actually accessing the (real) internet. How would it "know" the latest version of Python? Or that 212.58.244.22 is a BBC IP address, unless it were at least capable of doing live searches?
I'm convinced that it's not accessing the real Internet. All of this information would be in its training data.
Do we have any idea how big of a file the model actually is? I assumed it worked like the txt2img models where precise data is not hardcoded or stored literally anywhere for reference. A model that could accurately render all the WHOIS data in the world, since addresses are often arbitrary, would by itself be almost as large as the WHOIS database of 400+ million domains. There's not that much about WHOIS data that can be compressed into a neural network as a set of features and spit back out losslessly. So if it had just produced a realistic looking IP address at random, sure... but that is an actual BBC IP address.