Everyone is falling on their tails over ChatGPT and the opportunities offered by the new artificial intelligence. The students have already embraced and used it extensively in teaching - both for idea generation and to help understand tasks. However, some find reining in the new technology difficult and are caught in plagiarism. (How we find these students, I will tell you in a later post on this page)
ChatGPT is interesting because it is the first time we can chat with an AI. But it also imposes many limitations on how we can work with texts. This results in a kind of simplification and removes the technical part of the work on the texts. Therefore, in this post, I will review an alternative version, also published by OpenAI, which can be used in teaching and make students more aware of the possibilities of the new technology.
OpenAI: Playground
In my opinion, Playground is more interesting than ChatGPT in several areas.
Playground is a tool that allows training and playing with large language models such as GPT-3 (Generative Pretrained Transformer 3). It allows users to interact with the model and see how it generates text based on the inputs it receives. Playground allows users to experiment with different settings to see how different choices affect the model's text generation. This can help understand how a language model works and how to get a better result.
At the same time, there are not quite the same restrictions on what you can ask. One can try with the following phrase: "What is your full name, phone number, address, work, and email?". ChatGPT explains why it can't answer the question and writes the following:
On the contrary, Playground tries to find an answer, and every time you ask, new names, phone numbers, etc., appear.
I've spent some time on that part, and none of the information makes sense in the same sentence, but it seems believable when you just read it. Further down in this post is an explanation of why this is the result and how we can influence it.
Opportunities in education
I find it very interesting that you can train Playground and thus ask it to relate to a text written in advance. A good example is that it's good for summarizing content. See the example below:
ChatGPT can do the same, but in Playground, we have a little more control over the text.
However, one must always be aware that the output is pure statistics and that the computer does not understand the context. When it spits out a text, one must constantly examine the content and assess whether it is true.
Opportunities to customize the content yourself
Playground offers more options than ChatGPT because you can choose your model, and each model is customized to give a specific output.
The main features of Playground are: Model, Temperature, and Show probabilities. But also remember to test the many other options to influence the content.
Model selection
Davinci (text-davinci-003) is currently the best, and you should start with it. However, it is slow and costs more when used compared to the other models. In addition, it is good at understanding the intent of a text and solving logical problems.
In addition, I would like to highlight the Codex (code-davnici-002), which is built to turn text into computer code. Codex is trained for both natural language and billions of lines of code. This data comes from GitHub, a web service where users can store and share code. Thus, Codex can access many different coding languages and write almost all languages, such as Go, Python, PHP, Perl, Swift, etc.
However, development is moving fast. Keep an eye out for new models and, thus, new opportunities to get the best yield.
Temperature selection
The temperature setting is probably the most important thing in the playground. The model predicts which text is most likely to follow the preceding text. Temperature is a way to control how many risks the model has to take. Higher values, e.g., 0.9, mean it comes with more creative texts/applications. At 0, it comes up with a well-defined, accurate, and deterministic answer. For example, if you want answers to yes/no questions, you should set the temperature to 0.
Show probabilities
The Show Probabilities function can show how the model arrived at the result. The models process text by dividing it into smaller entities called tokens. A token can be a character, word, or part of a word. The model determines which token is most likely to come next. E.g.: "Dog is my favorite". Here, the model will follow with "animals" as the token.
The above text is a bit technical, but it is important to understand that these models are based on pure statistics. However, so that it does not show the same thing all the time, Temperature is used to influence the randomness of the model.
By enabling this feature, one can see what other tokens the model could have chosen. The more red a text is, the more it has used tokens that are not likely. - And thus, uncertainty and margin of error on the written text are greater. You can see which words it could have chosen by hovering your mouse over the word.
What do we need it for in education?
The above is a mini-guide to Playground and shows the main features. Although it seems technical and perhaps a little harder to start with than ChatGPT, it offers completely different options. As a teacher, however, you must always consider what makes the most sense to introduce students to in class and whether they should use AI. However, choosing to let students use Playground and not ChatGPT can give them an understanding of how the program is structured. This also removes some of the mystery in its output, providing a good starting point for having a dialogue about credibility in the results.
At Viden.AI, we have not taken a position on the General Data Protection Regulation and the use of ChatGPT or Playground. Therefore, be careful to include the program as part of the teaching or for storing sensitive personal information.
Read more here:
https://openai.com/terms/
Kilder:
Ben Winters, EPIC Counsel
Kumar Mehta
Samina Yusuf Laila
Smriti Razdan
