GPT-3¶
Quick Facts¶
| Attribute | Value |
|---|---|
| Released | June 2020 |
| Organization | OpenAI |
| Size | 175 billion parameters |
| Training Data | 570GB (Common Crawl, WebText2, Books, Wikipedia) |
| Context Window | 2,048 tokens |
| Key Innovation | Few-shot learning without fine-tuning |
| License | API only (proprietary) |
Historical Significance¶
GPT-3 was the breakthrough moment for LLMs:
- First model to show true few-shot learning
- Demonstrated emergent abilities at scale
- Made LLMs practically useful beyond research
- Sparked the current AI revolution
Architecture¶
Decoder-only Transformer
- 96 transformer layers
- 12,288 hidden dimensions
- 96 attention heads
- Trained on 300 billion tokens
- Estimated $4.6M training cost
Few-Shot Learning Example¶
import openai
# GPT-3's magic
prompt = """
Translate English to French:
English: "Hello, how are you?"
French: "Bonjour, comment allez-vous?"
English: "Good morning"
French: "Bon matin"
English: "What is your name?"
French:"""
response = openai.Completion.create(
engine="text-davinci-003", # GPT-3 variant
prompt=prompt,
temperature=0.7,
max_tokens=100
)
print(response.choices[0].text)
# Output
Performance¶
| Task | Accuracy | Notes |
|---|---|---|
| MMLU | 71.3% | Good for knowledge tasks |
| HumanEval | 48.1% | Weaker on code |
| Few-shot | Excellent | Core strength |
Why It Mattered¶
Before GPT-3:
- ML required task-specific training
- Fine-tuning on labeled data necessary
- Separate model per task
After GPT-3:
- Few-shot learning works
- One model for many tasks
- Prompt engineering replaces fine-tuning
Real-World Applications¶
- Text Summarization
prompt = """
Summarize this text in one sentence:
The quick brown fox jumps over the lazy dog.
This sentence contains all letters of the alphabet.
"""
# Result
- Question Answering
prompt = """
Q: What is the capital of France?
A: Paris
Q: What is the capital of Germany?
A:"""
# Result
- Code Generation
prompt = """
# Python function to calculate factorial
def factorial(n):"""
# Result
Strengths & Weaknesses¶
Strengths¶
- Revolutionary few-shot learning
- Excellent at diverse tasks
- Natural language understanding
- Good generalization
Weaknesses¶
- Can hallucinate facts
- Doesn't learn from instructions perfectly
- Slower than newer models
- Expensive for high volume
- 2,048 token limit (short context)
Current Status (2024)¶
- Deprecated (replaced by GPT-3.5, GPT-4)
- Historical significance remains
- Available via API for backwards compatibility
- Paved way for modern LLMs
Legacy¶
GPT-3 fundamentally changed AI:
- Proved scale leads to emergent abilities
- Showed practical applications beyond research
- Sparked multi-billion dollar AI industry
- Influenced architecture of all modern LLMs
-
Last Updated: 2026-08-09 Status: Superseded but historically crucial