The detailed analysis of the screenshot and the interpretation of chicken scratch to a working website were both extremely impressive. As a human I found the hand drawing hard to read. I do not know how it was able to determine this was a "discord UI window", what did they train it on?
To me I think what is interesting is not the delta in performance over GPT-3, but the reality that GPT-4 is almost capable enough to research more advanced AI systems. Being able to read error messages, write code, read docs, read visual input would seem to be the minimum capabilities for Recursive Self Improvement.
The detailed analysis of the screenshot and the interpretation of chicken scratch to a working website were both extremely impressive. As a human I found the hand drawing hard to read. I do not know how it was able to determine this was a "discord UI window", what did they train it on?
To me I think what is interesting is not the delta in performance over GPT-3, but the reality that GPT-4 is almost capable enough to research more advanced AI systems. Being able to read error messages, write code, read docs, read visual input would seem to be the minimum capabilities for Recursive Self Improvement.