Let’s Break This Down…
One thing that I appreciate after playing with the release of GPT4 is that it breaks down problems and shows reasoning. The tone, at least through the limited testing any individual can conduct, seems more modest. This could just be my biased, but it’s also been programmed to apologize and correct itself when confronted with new information.
Here’s what happened when I asked what would happen if you tried to fold a dollar bill seven times. First, the “old” GPT3.5:
GPT4 then responds as follows:
It breaks down the calculation to “show its work”. In going through this process, it helps highlight where there might be errors so they can be corrected. In other calculations, I had seen where it had assumed a square kilometer was 10 hectares instead of 100. When I brought this up, it recalculated and corrected itself.
I have had sessions with GPT4 that have lasted over an hour, working through intricate problems and testing different assumptions and paths. It’s getting good enough for this. It’s now on those using to use these insights to make meaningul advances.