Chat GPT: Temperature as a factor for creativity

Anyone who has worked with Chat GPT more than once will notice that the deeper you want to go into the subject, the Chat GPT will eventually always return the same answer. It takes a while, but at some point it's pointless to keep talking to the AI. One reason for this is the temperature parameter, which provides some creativity or better the necessary fuzziness in the response behavior. Since this parameter does not have the value 1 (no blur), it takes some time to get to this point.

Stephen Wolfram describes it this way: if we always pick the highest-ranked word, we’ll typically get a very flat essay, that never seems to show any creativity. We need some randomness, that is, a parameter that allows the AI to also use a less likely term as the selected token.

What does the temperature mean in the AI

Temperature is a parameter in artificial intelligence that affects the variety and quality of generated texts and other content. Thus, the temperature determines how often a lower valued term should be used in the response. A temperature of 0.8 is a frequently used value. Exciting, because the value 0.8 according to Wolfram is not scientifically justified, but is based on empirical values from practice. In other words, a subjective value that is still determined by people. It is, I think, but also no coincidence that the temperature also follows the 80/20 rule, the Pareto principle. Stephen Wolfram continues the the concept of temperature is there because exponential distributions familiar from statistical physics happen to be being used, but there’s no physical connection—at least so far as we know..

However, using a temperature of 0.8 also has ethical implications. AI systems may tend to generate misinformation or racist, sexist, or other discriminatory content by using a higher temperature. This is because the higher variety of texts generated may include inappropriate or problematic content. For this reason, it is important that AI developers and researchers handle the temperature responsibly and ensure that the texts generated are ethical and fair.

Creativity, verification and research

At c-rex.net, we have been using AI in various places for quite some time. Of course, we use the Co Pilot in software development. It generates code for us and recently we started using Chat GPT in our blog posts. But neither Co Pilot develops the c-rex.net software for us nor Chat GPT writes the blog and web presence for us. We use both systems more and more as support in our daily work. They save us time in the tedious yet mindless work of writing source code or writing texts grammatically clean and also understandable.

The rapid development of automation technologies is particularly evident in the field of translation. An area in which creativity has not really been desired for a long time anyway, and in the meantime is no longer lived. Instead of translating texts manually into different languages with the support of translation memory systems, today we can simply use any MT engine with the help of c-rex.net's automation scenarios. We simply define which languages we want to support and automation takes care of the rest. In various projects, we have noticed that machine translations are being relied on more and more frequently and the AI's suggestions are simply "waved through" in the post-editing process. This is not the only reason why there are fewer and fewer reasons for us to still rely on human support in the translation process every day. Particularly when considered in the triangle of conflict between cost, quality and time, the cost argument is becoming increasingly important.

As in the translation process, in the future we will no longer need developers who can only write code or know a particular programming language. In the future, we will also no longer need technical writers who only know how to write and write grammar. At the very least, these requirements are no longer required as core qualifications in job descriptions.

What we will increasingly need, however, are experts who have a deep knowledge of technologies and concepts, who can evaluate whether a system's response is correct or at least plausible. This requires the ability to do research. Last but not least, we need creativity, or perhaps better a warm temperature, which provides us with new ideas and moves us forward innovatively.

Conclusion

So let's use systems like Chat GPT, DeepL and Co Pilot for what they are: better search engines and writing aids that can and will take a lot of mindless and less creative work off our hands. Let's keep automating that to stay competitive and move into the future.

But this also requires that we are willing to give ourselves more breadth and depth. That's hard work to stay competitive with machine supporters. Whether we will succeed in doing so is open to doubt, and it is all the more important that we finally begin to create a social basis for mitigating the negative effects: Digitalization and the unconditional basic income!?.