For the first time in years, I took the time to go to a developer conference. The schedule looked like it would give me at least some sessions of value or interest both days, and the conference is in Stockholm so at least there’s no travel involved.
It was better than expected, even. The sessions that sounded good, were – and those that sounded possibly slightly promising, were better than that.
Among the most interesting sessions were two about ChatGPT – one about how it works under the hood, and one about how to get the most out of it, using prompt engineering.
The first one should be required listening for all the bloggers and journalists out there who keep saying that if we just keep making ChatGPT stronger and better, it will reach actual intelligence and truly understand the answers that it gives. ChatGPT is entirely and only about word prediction. Not even words, but character sequences, shorter than full words. Given these words, what is likely to come next? We saw actual live demos of an earlier version of ChatGPT running locally on the speaker’s computer, and he demonstrated how tweaking specific parameters will make ChatGPT more “adventurous”, i.e. more likely to vary its word sequence from the most likely, or more “conservative”, i.e. more likely to use the most common next word. Given “The cat sat on the…”, pulling the controls in one direction would make ChatGPT always continue with “floor”, whereas pulling them in the other increases the likelihood of getting “bed”, “table”, “shoulder” etc instead.
The second one taught me, for example, that instructing ChatGPT to provide step-by-step reasoning for whatever its conclusion is, makes it more likely that it will reach the correct conclusion. And sometimes you can get better, more detailed answers if you offer it money. In the source material that it has processed and is regurgitating, money leads to better work, so it behaves the same.