Disclaimer: This nice introduction wasn’t written by us, or by any human being for that matter. We simply asked an AI to write an introduction for a blog post about OpenAI. It doesn’t become more meta than this, right? Flabbergasted? Confused? Mind-blown? So were we ...
OpenAI is an Artificial Intelligence research laboratory, which conducts research in the field of AI. The goal of this laboratory, as the official site states, is to "promote and develop friendly AI in a way that benefits humanity as a whole". The company has been founded by none other than Elon Musk, Sam Altman, and others.
In 2020, the laboratory developed a natural language processing model, called GPT-3 (the third-generation Generative Pre-trained Transformer). This model was trained on trillions of words from the internet. At its core, GPT-3 has the ability to 'complete' an input. So, if we’re providing the input 'one small step for man', the response will most likely be 'one giant leap for mankind'. But what makes it really powerful, is that the same language model can do a diverse set of tasks: answer questions, create ad descriptions, generate code, come up with ideas for a blog post, and much, much more.
Lucky for us, OpenAI decided to expose this model via an API. They don’t release the actual model because of possible abuse or misuse. Their most important concern is that the model is biased. After all, its intelligence is based on actual written text by humans. That’s why they’ve taken these precautions. Using an API, they are able to monitor responses and contain misuse. Let’s say news sites use this model to automatically generate headlines and summaries for articles. That would mean that the output of GPT-3 could be published directly to their homepage without any human intervention. Since we know that the GPT-3 model has a negative sentiment towards blackness, these generated headlines will have the potential to feed more Black Lives Matter protests or extreme right thinking.
When using GPT-3, it seems like pure magic. Or like a human being is on the other end generating a response. But nothing is further from the truth: it’s a neural network that has been trained on 175 billion parameters. And that’s a lot if you compare this to other natural language models.
In the Playground interface, you can select a trained model and experiment with different settings. The Frequency Penalty setting, for example, can decrease or increase the model’s likelihood to repeat the same words. Once defined, the model will take them into account, trying to understand the context and giving you the most probable answer. And just like that, it feels like the output was given by a human.
What’s the real power of OpenAI? We don’t think it’s going to replace jobs in the near future. In some cases, you can definitely spot mistakes. But it can help you out a lot.
Let’s illustrate with some examples!
Our Creative Strategists could have used OpenAI to spark their creativity and make some noise for Netflix’ Sky Rojo.
In this example, we asked the AI to create an ad description. We provide context by setting the target audience and the social media platform. This way, you could e.g. automate ads or generate meta descriptions for thousands of products in your webshop.
You might have heard the phrase 'Anyone can code' before. One could argue about this statement, but with OpenAI that may very well be true ...
This wouldn’t be a tech blog post without focussing on actual code. Did you know there’s currently a GPT-3 'codex model' available in private beta? This is a special derivative of the GPT-3 model, focusing on programming languages. The model was trained against the biggest public code repository in the world, GitHub. Not only does it understand dozens of code languages, but it also writes code. Here’s an example:
Leveraging this model, GitHub Copilot was born. It’s available as a code editor extension that is basically able to help a programmer to do his or her job. It’s like code completion on steroids. Write what the code should be doing, and Copilot generates the code. Or let Copilot generate unit tests based on the code you’ve written.
It’s pretty interesting to see that programmers are actually creating a tool that could one day replace them. The faith of programmers, it seems, is not without a sense of irony.
Slowly but surely, AI is sneaking into our lives. And just like Augmented Reality, AI has the power to enhance our lives in ways we can’t even imagine. But prudence is in order, as it can go wrong really quickly: a biased AI could result in all sorts of catastrophic events. It could be used to generate fake news on a giant scale. And many jobs may disappear because of it, including ours.
Nevertheless, the potential is there. GPT-3 is a giant step forward in creating meaningful AI interactions. And we are very curious to see what will be possible in the near future.
Ready to play? You can sign up for the beta and test the API via the playground. Or you can integrate it in your own application by simply creating an API key and sending an http request (which also works in Dutch).
Do you want to see what we can do with AI? Then you should definitely read our blog post about Wattoo, an app that makes it easier for everyone to make the best out of their leftovers.