Don't skip the playground

This is my second article of two on the use of Chat AI models, for normal people. In this article (written May 2024) I will try to share my insights from 2 months of exploration and the feedback others gave me. In this article I try to answer the question:

What is the best AI model to use to get a good result?

As usual this depends on your use case but let me try and explain:

So many choices..

When I started there was really only ChatGPT, but now there are also Googles Gemini and Microsoft’s Co-pilot, Perplexity and more. Then each of them has various ways to use the model depending on how serious you are.

So, “What is the best AI model to get a good result” is actually 2 questions in one, which service to use, and how?

First, lets discuss OpenAi’s ChatGPT and the current alternatives. Let me explain the differences, and why I would use the one over the other:

• ChatGPT: Now in its 4.5 iteration, this is still the best general-purpose model if you want free flowing discussion on a topic, create idea’s or make pictures. In its newest iteration it can combine pictures, files and your text, in order to create an answer. In short, it’s still the go-to for me. Find the model at: https://chatgpt.com/
• ?Co-pilot: This Microsoft’s version of the ChatGPT model, and it pops up everywhere, in Bing search, as a browser extension or as a coding assistant. The difference is that where you visit OpenAI, Co-pilot comes to you in your application. Microsoft tailors the model to the environment to make it more likely it will be useful to you. It even has a specific version to create recipes! There are various versions, but the browser plugin (which is standard on Edge for windows, just click the top right corner) is the better option. This way it is really different in its interaction then ChatGPT, as it can actively use the current webpage as a starting point. And in my workflow I often have follow-up questions or need a summary of a webpage and its excellent for this. See the image below on how to find it.

?

• Google Gemini: Google Gemini is Googles answer to ChatGPT, as a chat model is an existential thereat to the google business model. Why use Google and look at ads when you can get the information for the AI Chat? As a result, Gemini excels at summarizing multiple websites and it allows for larger size documents and images to create your results. It’s the runner up and if you have large files, web specific content or python code, it’s a good choice. You can find the model at: https://gemini.google.com/
• Perplexity: A replacement for Google search, using both OpenAI and other models together. It gives references with the results and is fast. If you are unhappy with Google, and research a topic, give it a try. I use it if I need a quick factual answer. You can find the model at: https://www.perplexity.ai/

There are of course many other options I could list, but I find that these are the tools I come back to, if you find I missed an essential one let me know!

Hidden feature, the playground

Besides the actual service, the AI models allow you different way to interact with them. And as an advanced user this can be very powerful. To differentiate between casual users and power-users the AI companies have made three different ways (environments) to use their models: Chat , Playground and API.

Let me try and explain the different environments to use the models, which to use when and how I use them. I’ll use OpenAI ChatGPT as an example, but the same applies to Google Gemini or Microsoft Co-pilot.

For casual use

As a casual user the most obvious interaction with ChatGPT is just trough the chat interface ( https://chatgpt.com/). This is quite straightforward. The main options are the model type (GPT 3.5 /GPT4 / GPT 4 -o) and if you want to continue a conversation or start a new one. Here you can ask a question, upload a file or picture or generate a picture. Pretty straightforward, and ideal for casual use. Note the data submitted here can be used by OpenAi to improve the model.

Recently normal users got more room to discover, and the option was added to the chat app to slightly modify the standard model by giving the system instructions to follow. Nice, but this is much better done in the next step, the developer playground.

?Power users look here

Because if you want more control over how the model behaves, log into the playground on the developer’s page. This environment is actually super useful in my view, to try different prompt styles or to develop your own custom GPT’s, as here you can do a lot more. For instance:

First there is more control over the model behavior, both in its creativity when answering questions (called temperature) and limit its output.

Secondly there is the option to give the model “Memory”. This is done by uploading files (chat history, documentation, pictures) that it will use while answering questions. This is very powerful and only 3 months ago this was only available to professionals that built specific RAG databases to create hybrid models. Now as a power user you can add a number of documents as an example of how to fill a template or to inform the model of you writing style, or a manual to consult when answering questions, it really is a gamechanger.

Third the system instruction describing tone, audience and output length carry more weight. For example, instructing the model to answer “I don’t know”, if the model is uncertain, to avoid misinformation works much better here.

I find that if you are serious about using models for actual projects this is the place to be, there are far more options here and everything that you can do with programming in Python can be done here as well. Maybe even more important, at the time of writing, data submitted here will NOT be used to further enhance the model.

You can find the OpenAI developer playground here (https://platform.openai.com/playground/chat?models=gpt-4o ?) and the Google playground for Gemini – called AI Studio – here: (https://aistudio.google.com/app/prompts/new_chat )

?Seamless integration

If you want to integrate AI chat systems into your product (like Perplexity) the way to interact with the model is access via the API, this allows fully automated access to the features of the model as part of more complex scripts or software. For example, I used this in Python to automatically convert a mp3 audio file to a transcript and then use GPT4 to generate an extensive summary. (The new 4-o? model can do this from the chat). This is only interesting if you want to bring something into production, so really oriented towards developers making a production version. I use google co-lab for this, (https://colab.research.google.com/ )

?Wrap-up

The landscape in AI models changes super quickly, but in the end, there are only so many tools you can use. Personally, I tend to use the OpenAI playground the most, as its more advanced controls are very useful in more business minded use cases. It’s easy to use because no programming is required and its ideal to try out proof of concepts before committing serious resources. I will be trying out Gemini this summer, but in the playground as well.

My advice would be if you are a serious user of ChatGPT and have a paid subscription, have a look at the playground today, its included in your subscription and supercharges the way that you can work with AI.

?My tips, if you wan to learn more

To better understand the options that are available in the playground, have a look at the documentation from OpenAI, I find that its surprisingly well written, for a software product and explains these advanced features very well. You can find it here: (https://platform.openai.com/docs/overview )

I would also like to mention that Google has made a number of comprehensive, free learning courses on Generative AI as part of their cloud platform: https://www.cloudskillsboost.google/paths/118 , have a look.

See you next time,

Gijs

Thank you Joost Wouters , Christopher Grock and Duco Hulscher for the inspiration.

#DiscoveringAI