Too ambiguous and creator has given conflicting definitions, this will end up being disputed and resolved N/A.
For people who do not realize this, Isaac's definition of superintelligence includes "able to design and deploy a complicated website like a Facebook clone in under a minute."
This resolves NO, as do the 2040 and 2050 versions.
@DavidBolin I mean, I can write code for a Facebook clone (if you mean just basic features and not the entire bloated mess) in a week or so and deploy it (to a server I have up, unless it's built as some complex distributed thing) in very little time. It seems entirely reasonable to be able to do it faster. I don't think there are relevant hard limits here.
@DavidBolin I can run deployments in about five seconds (sync over code, restart service).
@DavidBolin Yes. My single 2020 consumer GPU running Mistral-7B in 6bpw with Exllama V2 gets 120 tokens a second. That is 6000 or so tokens in that time. I think you would need more than that to write the entire program, but that's easily doable with more GPUs (the task is partly parallelizable) or faster ones. Obviously to do it correctly you will need something smarter than a tiny quantized language model, but I think this suggests it's not fundamentally impossible.
@DavidBolin I usually know most of the libraries I will require before finishing the code.
@DavidBolin I don't agree with that and I also don't think you and people in general have good enough calibration to casually discuss probabilities like 10^-5.
"Conditional on no existential catastrophe"
That's true for every market, you don't need to state it 🤣