Exploring the OpenAI GPT-3 API

0. Why This Guide Exists

Everybody on campus keeps asking the same thing:
"Show me GPT-3 without the playground."
The completions endpoint is live, the docs are thin, and the possibilities feel endless. I built GPT-3 Academy—a single-file Streamlit app—to demo the engine and took notes along the way. This page is a distilled, hands-on guide so you can recreate (or remix) the magic yourself.

1. Sign Up and Grab Your Key

Create an OpenAI account.
Open View API Keys in the dashboard.
Click Create secret key and copy it once.
Store it safely—environment variables keep the key out of source control.

export OPENAI_API_KEY="sk-…your-long-string…"

2. Install the SDK

pip install openai==0.27.0      # current latest
pip install streamlit

The openai client hides HTTP plumbing so you speak plain Python.

3. Hello Completion

import openai, os
openai.api_key = os.getenv("OPENAI_API_KEY")

resp = openai.Completion.create(
    model="text-davinci-002",
    prompt="Write a haiku about autumn",
    max_tokens=32,
    temperature=0.8
)
print(resp["choices"][0]["text"].strip())

Key Parameters

Field

Plain-English Effect

model

Which brain (text-davinci-002 tops the list)

prompt

Everything the model sees before it starts typing

max_tokens

Hard ceiling on output length

temperature

0 = laser-focused, 1 = let it riff

top_p

Alternative to temperature for nucleus sampling

frequency_penalty

Penalize token repetition

presence_penalty

Encourage the model to introduce new topics

4. Prompt Recipes to Try

Summaries

Summarize the following for a second-grade reader:
<your text>

Outline generator

Make a bullet outline covering the key ideas:
<your text>

Fact check

List factual claims in the text, rate their plausibility 1-5, and cite sources if possible.

Quiz builder

Create five multiple-choice questions that test understanding of:
<topic>

5. Keep the Bill Down

@st.cache  # Streamlit memoizes identical calls
def call_openai(prompt):
    return openai.Completion.create(
        model="text-davinci-002",
        prompt=prompt,
        max_tokens=256,
        temperature=0.7
    )["choices"][0]["text"]

Caching saves credits while you tweak the UI (and impress friends).

6. Wrap It in Streamlit

import streamlit as st, openai, os
openai.api_key = os.getenv("OPENAI_API_KEY")

st.set_page_config(layout="centered", page_title="GPT-3 Academy")
st.header("GPT-3 Academy")

text = st.text_area("Paste or write text")
task = st.selectbox(
    "Choose capability",
    ("Summarize for a 2nd grader", "Make an outline",
     "Fact Check", "Make a quiz")
)

if st.button("Run AI"):
    if not text:
        st.warning("Enter some text")
    else:
        prompt = f"{task}:\n{text[:1500]}"  # 1500-char guard
        st.write(call_openai(prompt))

Why Streamlit? One file, instant widgets, no front-end hassle—perfect for quick demos.

7. What GPT-3 Does Well (October 2022 Snapshot)

Capability

Real-world Note

Summarization

Excellent on articles under ~2k tokens

Light reasoning

Good if you guide with clear chain-of-thought cues

Creative writing

Poems, jokes, short fiction feel human-ish

Translation

Strong for high-resource languages

Code snippets

Short Python functions okay; long blocks need careful prompts

Factual recall

Mixed accuracy—add "If unsure, say 'I don't know'" to reduce hallucinations

8. Common Pitfalls

Token limit: prompt + completion must stay under 4096 tokens.
High temperature: entertaining but can spiral into nonsense.
Missing newlines: dense prompts sometimes yield run-on gibberish.
Rate limits: starter accounts sit at 60 requests per minute.

9. Next Steps

Expose temperature and max_tokens as UI sliders.
Log every prompt–response pair to SQLite for later study.
Benchmark cheaper models (curie, babbage) against Davinci.
Watch the changelog—rumor says a chat-style endpoint is brewing.