OpenAI
GPT-3
Paper published May 28, 2020
One model, many language tasks
What could it do?
Attempt translation, questions, short arithmetic and other language tasks by following examples placed in the prompt.
What changed?
The paper demonstrated broader learning from examples at inference time, without updating the model’s weights for each task.
Few shot learning
A single general model could be steered toward different tasks through the input itself.
Where it fell short
The paper identified weak tasks and possible overlap between training data and evaluations.
May 2020 research paper; this date is not the API launch. Check sources
- Model type
- Text completion
- Largest model
- 175 billion parameters
- Date recorded
- First paper submission
What this meant in practice
Changing the examples changes the task without retraining a new model. For a business, the interesting shift is experimentation: a prompt can become a prototype. Reliability still has to be measured on the business’s own inputs.
Sort customer messages
Show a few messages labelled “booking” or “billing,” then supply another. Check ambiguous messages separately instead of assuming a fluent label is correct.
Common question
Did it learn permanently from a prompt?
The paper’s few shot setting used examples in context, without changing the model’s trained weights.
Release facts were checked against the sources below. Performance claims belong to the developers; we have not independently tested these models. This date records the paper’s first submission, not public API availability.
Brown et al.: Language Models are Few-Shot LearnersBenchmark results
No comparable ECI score is available for this release in our source snapshot. Missing scores are never estimated.