ChatGPT from OpenAI has been getting a lot of publicity recently. If you haven’t heard of it, ChatGPT is an acronym for Chat Generative Pre-trained Transformer. What does that mean? It’s a chatbot that’s been tuned through reinforcement and supervised learning. There has been much discussion about the accuracy and usability of ChatGPT.
ChatGPT currently offers three different models for response generation. They have different capabilities and come with different pricing. The models are:
- GPT-3, for understanding and generating natural language.
- Codex, for understanding and generating software code.
- Content filter, for determining whether text may be sensitive or unsafe
Being in the software industry, the Codex model is intriguing, but I have yet to see examples of auto-generated code that does exactly what I’m trying to accomplish. Therefore, having previous experience in the culinary industry I thought it would be interesting to try ChatGPT in that arena using the GPT-3 model. Let’s see what sort of recipes ChatGPT can generate and how true to the classic recipes they are.
There are four different model options available when using the overall GPT-3 model:
text-davinci-003. The models go along a speed/cost scale of fast & inexpensive with
text-ada-001 to the most capable but slower & more expensive
text-davinci-003. While fine-tuning these models is an option, it goes beyond the scope of this post. With that in mind, let’s look at what these different models can generate for us “out of the box.”
In classical French cuisine, there are five mother sauces on which other sauces are based. To test out ChatGPT’s recipe generation, let’s look at one of these classic sauces: Béchamel. Béchamel is a simple sauce, from a classic ingredient standpoint, in that it contains butter, all-purpose flour, milk, salt, and pepper.
Let’s write a brief Python application that we’ll use to access the OpenAI API and see what responses we get from ChatGPT.
First, we’ll need to get an account with OpenAI and an API Key. As a best practice to keep your API Key safe, using an API Vault, such as Pangea’s Vault is a great idea. At a bare minimum, assign it to an environment variable so you can access in your application but not need to save the key in your code file.
Second, we’ll need to install OpenAI using a package management tool such as pip or pipenv. I prefer pipenv, so running the following command will get that installed:
pipenv install openai
Now, in a Python file,
chatgpt-python.py, we’ll make a “How do I make bechamel sauce?” request to OpenAI using the four different GPT-3 models:
import openai import os openai.api_key = os.environ['OPENAPI_APIKEY'] # Choose the model engine model_engines = ["text-ada-001", "text-babbage-001", "text-curie-001", "text-davinci-003"] # The question passed into the API prompt = "How do I make bechamel sauce?" for model_engine in model_engines: # Response generation generator = openai.Completion.create( engine=model_engine, # (REQUIRED) - the model to use prompt=prompt, # (OPTIONAL) - the prompt that's passed to the API max_tokens=1024, # (OPTIONAL) - the max # of tokens to generate n=1, # (OPTIONAL) - how many completions to generate stop=None, # (OPTIONAL) - stop word string or array. temperature=0.5. # (OPTIONAL) - the coherency of the generated response. Number between 0 and 2 ) generated_response = generator.choices.text print("Response for " + model_engine + ": ") print(generated_response)
In the response generator, we have a few options, and we’re defining the model engine, passing in the prompt, defining the number of tokens used from the API (default is 16), and how many completions are generated for each prompt (default is 1), any stop sequences we want to define to stop further token generation, and the temperature of the response.
What is the temperature? It defines the randomness (coherency) of the generated response. Lower numbers here make the responses more focused and higher numbers generate more random output. We’ll explore this in a bit. For now, here are the responses I got from the above script. Remember that we’re looking for an answer to the question of “How do I make béchamel sauce?”
text-ada-001 (.50 second response)
There is no one definitive way to make bechamel sauce. However, most likely you will need to use grated cheese, milk, and butter.
Hmmm, that doesn’t really seem helpful and while there are sauces based on a bechamel that include cheese the base sauce doesn’t.
text-babbage-001 (0.61 second response)
Place a pot of water on the stove to boil. Add a cup of sugar and cook until the sugar has dissolved. Add a cup of milk and cook until the milk has boiled off. Pour the sauce into a saucepan and stir in the bechamel sauce ingredients. Cook over low heat until thickened.
Again, not spectacularly useful, for classic béchamel with sugar being mentioned. Also, what is a “pot of water” exactly? 1 cup, 1 quart, 5 gallons? I’m not sure what “until the milk has boiled off” means in this context either. And then, a very cyclical instruction for how to make bechamel sauce: “stir in the bechamel sauce ingredients.”
text-curie-001 (.84 second response)
Bechamel sauce is a simple sauce made from milk and flour. Heat milk in a saucepan over medium heat until it starts to simmer. Whisk in flour until a smooth paste forms. Cook, stirring constantly, until the sauce thickens. Serve hot.
Getting closer here, but still not quite there. Using straight flour to thicken milk doesn’t create a great flavor. However, even without measurements, you could generate something here, right? Even with 100 gallons of milk being heated to a simmer you know that you must add in enough flour until a “smooth paste forms.” Still, far from a classical béchamel sauce.
text-davinci-003 (7.3 second response)
Bechamel sauce is a classic French sauce made from a roux of butter and flour, and a liquid, usually milk. Here is a basic recipe:
-3 tablespoons butter
-3 tablespoons all-purpose flour
-2 cups milk
-Salt and freshly ground pepper
-Pinch of ground nutmeg
1. Melt the butter in a saucepan over medium heat.
2. Add the flour and whisk until the mixture is smooth and bubbling.
3. Gradually whisk in the milk, whisking constantly to avoid lumps.
4. Continue to whisk until the sauce thickens, about 5 minutes.
5. Season with salt, pepper, and nutmeg.
6. Serve warm.
This is a recipe that one could work with. I’d have a few nit-picks with the instructions, but overall, it makes sense. Personally, I wouldn’t add the nutmeg and it isn’t explicitly classical, but I’ve seen it done.
As mentioned earlier, the temperature parameter adjusts the “coherency” of the generated output. You can set that parameter to a number between 0 and 2. Recall that the above script ran with a temperature parameter of 0.5. The output was, while not always helpful, at least coherent.
The output from the
text-davinci-003 model was useful. Let’s see what happens if we alter the temperature parameter value.
Bechamel is one of the staples in employing tha classic “mother tastes” that understands fundamental techniques, which are included bases like Bechamel felt se as Milk, simple sugars /mutable faster blend together all seeds again serve holes along veggies traditional
That’s just the introduction, the rest of the “recipe” isn’t much better…
To make béchamel classic:
1 MELT 2 tablespoons worthwhile normal core puns international info and work users tub back stainless stre ‘tub fully restricted sized documents allowed uses principle-friendly views data
… and it goes on but becomes less coherent.
Let’s go in the other direction and reduce the temperature parameter value. Here are the results:
Bechamel sauce is a classic French white sauce made with butter, flour, and milk.
– 2 tablespoons butter
– 2 tablespoons all-purpose flour
– 2 cups milk
– Salt and pepper to taste
1. In a medium saucepan, melt the butter over medium heat.
2. Add the flour and whisk until the mixture is smooth and bubbly.
3. Gradually add the milk, whisking constantly until the mixture is smooth.
4. Bring the mixture to a boil, then reduce the heat and simmer for 5 minutes, stirring occasionally.
5. Season with salt and pepper to taste.
6. Serve over your favorite dish. Enjoy!
Another great béchamel recipe, and this time without nutmeg!
I don’t believe you. You’re an impasta!
As we’ve seen using OpenAI and ChatGPT can have some interesting results. You might wind up with some imposter recipes or you might wind up with some winners. That is true overall, in my opinion, for AI-generated content now though. Could you build an entire application strictly based on results from ChatGPT? Perhaps. Individual results may vary though.
As the models get better, get more use, and “learn” more, things might change. And recipe creation is just one small example of this technology. We see folks attempting to use it in areas such as courtrooms, health care, insurance, and more. In comparison, culinary uses seem safe. Just based on the varying results of a béchamel sauce recipe, I think we have a ways to go with this technology.