Last week, I had many interesting discussions about OpenAI and GPT on Laracon in Porto. Especially with Marcel Pociot.
I've learned much more in 2 days than on the Internet since December.
That feels great, and tips seem basic but effective. But as in any other fresh area, finding out about them takes a lot of work. I want to embrace sharing in the GPT community, so here is cherry-pick list of failures and tricks from people who were generous to share it with me.
"If you want to go fast, go alone.
If you want to go far, go together".
This is where I confused the people I spoke with. I thought GPT was the same as ChatGPT. But it's not. GPT is a model, ChatGPT is an online service with form that uses the GPT model.
You can use the ChatGPT from your browser here: https://chat.openai.com.
On the other hand, you can call GPT via REST API. The API is paid service, where you pay for tokens.
It's a premium service of the online form that will run faster. It should cost 42 $.
The go-to package in PHP is a openai-php/client created by Nuno Maduro.
composer require openai-php/client
It's a wrapper around the REST API.
$yourApiKey = getenv('YOUR_API_KEY');
$client = OpenAI::client($yourApiKey);
$result = $client->completions()->create([
'model' => 'text-davinci-003',
'prompt' => 'PHP is',
]);
// "an open-source, widely-used, server-side scripting language"
echo $result['choices'][0]['text'];
After you login in Openai.com, you can get the token here platform.openai.com/account/api-keys.
Exactly 1 267. Every model has different limits on input content. You cannot past a whole book and expect a summary.
You can check your prompt size with official Tokenizer page: platform.openai.com/tokenizer
It depends. The shorter the prompt, the faster the response. To give you an idea, the typical TestGen AI response time is 6-10 seconds.
Well, unless the GPT is down. Then it takes longer :)
Using real-time does not make sense because the response could be faster. It's better to send the request to a background queue, let the worker handle it, and show the response when it's done.
I want to refactor TestGen AI to Livewire to address this.
The day before the conference, TestGen AI stopped working. I didn't know if it was something on my side or if the GPT was down in general.
There is a website that tells you the answer: status.openai.com:
This is a bit advanced topic, but it could be helpful, so I put it here. The longer you use the GPT, the more you see that the context is everything.
E.g., let's say we want to ask GPT to generate a unit test for the ConferenceFactory
class. This class has a dependency in the constructor - a TalkFactory
and SpeakerRepository
. To make GPT works the best, you should provide these files too.
This is similar to the way GitHub Copilot works - it has a context of your project (not a whole, but some files), and it uses it to generate more tailored code. Here is how Copilot Internals.
In the next post, we'll look at the 2 different models and how to treat them right.
Let me know in the comments or share on Twitter with #GPTtips. I believe we can learn a lot from each other and reach our goals faster.
Happy coding!
Do you learn from my contents or use open-source packages like Rector every day?
Consider supporting it on GitHub Sponsors.
I'd really appreciate it!