OpenAI last month launched the third iteration of its natural language processing software called Generative Pretrained Transformer, GPT-3. Quite an enhancement over GPT-2, released in a testing version last year, GPT-3 promises more functionality. It has a larger database than the previous iteration and offers more reliability and accuracy. GPT-3 is like any artificial intelligence programme, just with a more extensive database.
What is natural language processing?
Natural language processing is how a software or programme reads language. It first checks for grammar and then for semantics. The more a software reads, the better it understands. So, if the information fed into the programme increases, so does its understanding. GPT-3 has a database of 175 billion parameters.
How is GPT-3 different? What are the improvements over GPT-2?
When OpenAI released its first iteration of the software, the sole play was advanced machine learning. While GPT-3 has a better AI, no doubt, what differentiates it, is the vast database and high-level processing. It has 100-times the database that GPT-2 had. So, its accuracy is much better, but that does
not mean that it can match human intelligence. But it will certainly be good at specific tasks. Last year, a GPT-2 test showed that the programme had an accuracy of 89% as far as children’s book writing was concerned; humans did only a bit better at 92%. But in the case of Winograd Schema Challenge, it was only 70.7% accurate as against 92% human accuracy. For Lambada challenge, it fared worse.
What can be done with GPT-3?
It can undoubtedly spin stories for children stories books, but what it does best is complete codes, write headlines and tweets. More important, it can be used to query a database.
