Most code examples are written in Python, though the concepts can be applied in any language.
In the same way that a cookbook's recipes don't span all possible meals or techniques, these examples don't span all possible use cases or methods. Use them as starting points upon which to elaborate, discover, and invent.
[Large language models][Large language models Blog Post] are functions that map text to text. Given an input string of text, a large language model tries to predict the text that will come next.
The magic of large language models is that by being trained to minimize this prediction error over vast quantities of text, the models end up learning concepts useful for these predictions. For example, they learn concepts like:
* how to spell
* how grammar works
* how to paraphrase
* how to answer questions
* how to hold a conversation
* how to write in many languages
* how to code
* etc.
None of these capabilities are explicitly programmed in - they all emerge as a result of training.
GPT-3's capabilities now power [hundreds of different software products][GPT3 Apps Blog Post], including productivity apps, education apps, games, and more.
Instruction-following models (e.g., `text-davinci-003` or any model beginning with `text-`) are specially designed to follow instructions. Write your instruction at the top of the prompt (or at the bottom, or both), and the model will do its best to follow the instruction and then stop. Instructions can be detailed, so don't be afraid to write a paragraph explicitly detailing the output you want.
Extract the name of the author from the quotation below.
“Some humans theorize that intelligent species go extinct before they can expand into outer space. If they're correct, then the hush of the night sky is the silence of the graveyard.”
Completion-style prompts take advantage of how large language models try to write text they think is most likely to come next. To steer the model, try beginning a pattern or sentence that will be completed by the output you want to see. Relative to direct instructions, this mode of steering large language models can take more care and experimentation. In addition, the models won't necessarily know where to stop, so you will often need stop sequences or post-processing to cut off text generated beyond the desired output.
“Some humans theorize that intelligent species go extinct before they can expand into outer space. If they're correct, then the hush of the night sky is the silence of the graveyard.”
― Ted Chiang, Exhalation
The author of this quote is
```
Output:
```text
Ted Chiang
```
### Demonstration prompt example (few-shot learning)
Similar to completion-style prompts, demonstrations can show the model what you want it to do. This approach is sometimes called few-shot learning, as the model learns from a few examples provided in the prompt.
Example demonstration prompt:
```text
Quote:
“When the reasoning mind is forced to confront the impossible again and again, it has no choice but to adapt.”
― N.K. Jemisin, The Fifth Season
Author: N.K. Jemisin
Quote:
“Some humans theorize that intelligent species go extinct before they can expand into outer space. If they're correct, then the hush of the night sky is the silence of the graveyard.”
With enough training examples, you can [fine-tune][Fine Tuning Docs] a custom model. In this case, instructions become unnecessary, as the model can learn the task from the training data provided. However, it can be helpful to include separator sequences (e.g., `->` or `###` or any string that doesn't commonly appear in your inputs) to tell the model when the prompt has ended and the output should begin. Without separator sequences, there is a risk that the model continues elaborating on the input text rather than starting on the answer you want to see.
Example fine-tuned prompt (for a model that has been custom trained on similar prompt-completion pairs):
```text
“Some humans theorize that intelligent species go extinct before they can expand into outer space. If they're correct, then the hush of the night sky is the silence of the graveyard.”
* **Give more explicit instructions.** E.g., if you want the output to be a comma separated list, ask it to return a comma separated list. If you want it to say "I don't know" when it doesn't know the answer, tell it 'Say "I don't know" if you do not know the answer.'
* **Supply better examples.** If you're demonstrating examples in your prompt, make sure that your examples are diverse and high quality.
* **Ask the model to answer as if it was an expert.** Explicitly asking the model to produce high quality output or output as if it was written by an expert can induce the model to give higher quality answers that it thinks an expert would write. E.g., "The following answer is correct, high-quality, and written by an expert."
* **Prompt the model to write down the series of steps explaining its reasoning.** E.g., prepend your answer with something like "[Let's think step by step](https://arxiv.org/pdf/2205.11916v1.pdf)." Prompting the model to give an explanation of its reasoning before its final answer can increase the likelihood that its final answer is consistent and correct.
## Text Capabilities
### 1. Write text
Large language models are excellent at writing text. They can assist with:
* Blog posts
* Email copy
* Ad copy
* Website copy
* Product descriptions
* Memos
* Storytelling
* Brainstorming
* Question generation
* etc.
An example prompt for an instruction-following model:
```text
Write an email to a colleague named Jill congratulating her on her promotion. The tone should be warm yet professional. Mention how you admire the work she's been putting in. Include a joke about how her pet lizard Max enjoys eating grasshoppers. Mention how you're looking forward to the team off-site next week.
```
Output:
```text
Dear Jill,
Congratulations on your promotion! I've been admiring the great work you've been putting in and I'm really happy for your success. Max the lizard must be thrilled too - I bet he's looking forward to feasting on even more grasshoppers!
I'm really looking forward to next week's team off-site. It's going to be great to catch up with everyone and brainstorm some new ideas.
Best,
[Your Name]
```
In general, writing can work with any style of prompt. Experiment to see what works best for your use case.
| Base models<br>(e.g., `davinci`) | More creative | More expensive (as including examples demonstrations in prompt will cost tokens) |
| Fine-tuned models | Can train off of many examples; cheaper than including examples in the prompt | Hard to gather training data; training makes iteration slower and more expensive |
### 2. Explain text
One capability of large language models is distilling information from a piece of text. This can include:
* Answering questions about a piece of text, e.g.:
* Querying an unfamiliar document to understand what it contains
* Querying a document with structured questions in order to extract tags, classes, entities, etc.
* Summarizing text, e.g.:
* Summarizing long documents
* Summarizing back-and-forth emails or message threads
* Summarizing detailed meeting notes with key points and next steps
* Classifying text, e.g.:
* Classifying customer feedback messages by topic or type
* Classifying documents by topic or type
* Classifying the tone or sentiment of text
* Extracting entities, e.g.:
* Extracting contact information from a customer message
* Extracting names of people or companies or products from a document
* Extracting things mentioned in customer reviews or feedback
#### Answering questions about a piece of text
Example prompt for answering questions about a piece of text:
```text
Using the following text, answer the following question. If the answer is not contained within the text, say "I don't know."
Text:
"""
Oklo Mine (sometimes Oklo Reactor or Oklo Mines), located in Oklo, Gabon on the west coast of Central Africa, is believed to be the only natural nuclear fission reactor. Oklo consists of 16 sites at which self-sustaining nuclear fission reactions are thought to have taken place approximately 1.7 billion years ago, and ran for hundreds of thousands of years. It is estimated to have averaged under 100 kW of thermal power during that time.
"""
Question: How many natural fission reactors have ever been discovered?
If the text you wish to ask about is longer than the token limit (~4,000 tokens for `text-davinci-003` and ~2,000 tokens for earlier models), we recommend splitting the text into smaller pieces, ranking them by relevance, and then asking the most-relevant-looking pieces.
Two independent experiments reported their results this morning at CERN, Europe's high-energy physics laboratory near Geneva in Switzerland. Both show convincing evidence of a new boson particle weighing around 125 gigaelectronvolts, which so far fits predictions of the Higgs previously made by theoretical physicists.
"As a layman I would say: 'I think we have it'. Would you agree?" Rolf-Dieter Heuer, CERN's director-general, asked the packed auditorium. The physicists assembled there burst into applause.
"""
Summary:
```
Output:
```text
CERN has announced the discovery of a new particle, the Higgs boson. This particle has been predicted by theoretical physicists and is a major step forward in our understanding of the universe.
```
#### Classification
The best approach for classifying text depends on whether the classes are known in advance or not.
If your classes are known in advance, classification is best done with a fine-tuned model, as demonstrated in [Fine-tuned_classification.ipynb](examples/Fine-tuned_classification.ipynb).
If your classes are not known in advance (e.g., they are set by a user or generated on the fly), you can try zero-shot classification by either giving an instruction containing the classes or even by using embeddings to see which class label (or other classified texts) are most similar to the text ([Zero-shot_classification.ipynb](examples/Zero-shot_classification_with_embeddings.ipynb)).
#### Entity extraction
An example prompt for entity extraction:
```text
From the text below, extract the following entities in the following format:
People & titles: <comma-separatedlistofpeoplementioned(withtheirtitlesorrolesappendedinparentheses)>
Text:
"""
In March 1981, United States v. AT&T came to trial under Assistant Attorney General William Baxter. AT&T chairman Charles L. Brown thought the company would be gutted. He realized that AT&T would lose and, in December 1981, resumed negotiations with the Justice Department. Reaching an agreement less than a month later, Brown agreed to divestiture—the best and only realistic alternative. AT&T's decision allowed it to retain its research and manufacturing arms. The decree, titled the Modification of Final Judgment, was an adjustment of the Consent Decree of 14 January 1956. Judge Harold H. Greene was given the authority over the modified decree....
In 1982, the U.S. government announced that AT&T would cease to exist as a monopolistic entity. On 1 January 1984, it was split into seven smaller regional companies, Bell South, Bell Atlantic, NYNEX, American Information Technologies, Southwestern Bell, US West, and Pacific Telesis, to handle regional phone services in the U.S. AT&T retains control of its long distance services, but was no longer protected from competition.
"""
```
Output:
```text
Companies: United States v. AT&T, AT&T, Justice Department, Bell South, Bell Atlantic, NYNEX, American Information Technologies, Southwestern Bell, US West, Pacific Telesis
People & titles: William Baxter (Assistant Attorney General), Charles L. Brown (AT&T chairman), Harold H. Greene (Judge)
In addition to the [completion API endpoint][Completion API Docs], OpenAI now offers an [edit API endpoint][Edit API Docs] ([blog post][GPT3 Edit Blog Post]). In contrast to completions, which only take a single text input, edits take two text inputs: the instruction and the text to be modified.
Therewassomehostilityntheenergybehindthe researchreportedinPerceptrons....Part of ourdrivecame,aswequiteplainlyacknoweldgednourbook,fromhe facthatfundingndresearchnergywerebeingdissipatedon. . .misleadingttemptsouseconnectionistmethodsnpracticalappli-cations.
```
Output:
```text
There was some hostility in the energy behind the research reported in Perceptrons....Part of our drive came, as we quite plainly acknowledged in our book, from the fact that funding and research energy were being dissipated on...misleading attempts to use connectionist methods in practical applications.
```
#### Translation
Translation is another emergent capability of large language models. In 2021, [GPT-3 was used](https://arxiv.org/abs/2110.05448) to set a new state-of-the-art record in unsupervised translation on the WMT14 English-French benchmark.
Example translation prompt using the edits endpoint:
Instruction input:
```text
translation into French
```
Text input:
```text
That's life.
```
Output:
```text
C'est la vie.
```
Example translation prompt using the completions endpoint:
```text
Translate the following text from English to French.
* We've seen better performance when the instruction is given in the final language (so if translating into French, give the instruction `Traduire le texte de l'anglais au français.` rather than `Translate the following text from English to French.`)
* Backtranslation (as described [here](https://arxiv.org/abs/2110.05448)) can also increase performance
* Text with colons and heavy punctuation can trip up the instruction-following models, especially if the instruction is using colons (e.g., `English: {english text} French:`)
* The edits endpoint has been seen to sometimes repeat the text input alongside the translation
When it comes to translation, large language models particularly shine at combining other instructions alongside translation. For example, you can ask GPT-3 to translate Slovenian to English but keep all LaTeX typesetting commands unchanged. The following notebook details how we translated a Slovenian math book into English:
The [OpenAI API embeddings endpoint][Embeddings Docs] can be used to measure similarity between pieces of text ([blog post][Embeddings Blog Post]). By leveraging GPT-3's understanding of text, these embeddings [achieved state-of-the-art results](https://arxiv.org/abs/2201.10005) on benchmarks in both unsupervised learning and transfer learning settings.
* Store those embeddings in your own database or in a vector search provider like [Pinecone](https://www.pinecone.io) or [Weaviate](https://weaviate.io)
An example of how to use embeddings for search is shown in [Semantic_text_search_using_embeddings.ipynb](examples/Semantic_text_search_using_embeddings.ipynb).
An example of how to use embeddings for recommendations is shown in [Recommendation_using_embeddings.ipynb](examples/Recommendation_using_embeddings.ipynb).
In the following notebook, we provide an example method for customizing your embeddings using training data. The idea of the method is to train a custom matrix to multiply embedding vectors by in order to get new customized embeddings. With good training data, this custom matrix will highlight the features relevant to your training labels and suppress the rest. You can equivalently consider the matrix multiplication as (a) a modification of the embeddings or (b) a modification of the distance function used to measure the distances between embeddings.
Note that unlike instruction-following text models (e.g., `text-davinci-003`), Codex is *not* trained to follow instructions. As a result, designing good prompts can take more care.
A well-written SQL query that lists customers who signed up during March 2020 and watched more than 50 hours of video in their first 30 days:
```
````
Output:
```sql
SELECT c.customer_id
FROM Customers c
JOIN Streaming s
ON c.customer_id = s.customer_id
WHERE c.signup_date BETWEEN '2020-03-01' AND '2020-03-31'
AND s.watch_date BETWEEN c.signup_date AND DATE_ADD(c.signup_date, INTERVAL 30 DAY)
GROUP BY c.customer_id
HAVING SUM(s.watch_minutes) > 50 * 60
```
`code-davinci-002` is able to make inferences from variable names; for example, it infers that `watch_minutes` has units of minutes and therefore needs to be converted by a factor of 60 before being compared with 50 hours.
### 2. Explain code
Code explanation can be applied to many use cases:
Translate to JavaScript (or Rust or Lisp or any language you like)
```
Example output after improving the runtime and translating to JavaScript:
```JavaScript
function tribonacci(n) {
let a = 0;
let b = 1;
let c = 1;
for (let i = 0; i <n;i++){
[a, b, c] = [b, c, a + b + c];
}
return a;
}
```
As you can see, `code-davinci-edit-001` was able to successfully reduce the function's runtime from exponential down to linear, as well as convert from Python to JavaScript.
### 4. Compare code
The OpenAI API also features code search embeddings, which can measure the relevance of a section of code to a text query, or the similarity between two sections of code.
OpenAI code search embeddings significantly improved the state-of-the-art on the [CodeSearchNet] evaluation suite, scoring 93.5% versus the previous record of 77.4%.
