[infurl]

https://blog.seekwell.io/gpt3

The SQL (Structured Query) Language was originally developed in the hope that non-programmers would be able to use it to get their reports and answer their questions without the help of trained and experienced computer programmers. That didn't happen but it remains a valuable goal.

The linked article outlines another very promising attempt which uses GPT3 to write working SQL queries in response to plain english questions. What is interesting is that it can do this even without knowing the database schema in some cases because it can guess sensible names for the fields and tables that it would need. What is even more interesting is how easily it can learn from a few examples.

Input: how much revenue have we received from users with an email ending in 'seekwell.io' in the last 3 months?

GPT-3 Response: SELECT SUM(charges.amount) FROM users INNER JOIN charges ON users.id = charges.user_id WHERE users.email LIKE '%seekwell.io%' AND users.signup_dt>= now() - interval '3 months'

[Don Patrick]

Considering that GPT does not actually accommodate logical structure and takes example from random people's suggestions on stackoverflow, it seems to me that this changes the job from programming to debugging. It is impressive that it can associate words to, well, capitalised words, but on the whole I don't see the improvement: One still has to type the request and one still needs an expert to debug the code. Even more so because the machine makes inhuman mistakes.

[infurl]

GPT3 reminds me of the adage about a million monkeys. Even if one of them did type out the works of Shakespeare, none of them would know which one it was. GPT3 only produces results that might seem impressive only when they have been heavily curated. It only saves time compared to something that generates random strings.