NL2SQL: why memory-less ChatGPT strongly ouptperforms text-davinci-003?

General API discussion
1

Hi, I’m using the same prompt, in 2 environments - ChatGPT (at the start of the session) and Completion (via Playground: text-davinci-003, Temperature 0.7). The result from ChatGPT is correct, the result of text-davinci-003 is patently wrong.
Prompt:

Schema for table Reports:
	UserID VARCHAR
	ReportDate DATETIME

As a senior analyst, given the above schema, write a detailed and correct Postgres SQL query to answer the analytical question:
How is distributed maximal time gap between consequent reports sent by the given user?

ChatGPT results (correct):

WITH cte AS (
  SELECT 
    UserID, 
    ReportDate, 
    lag(ReportDate) OVER (PARTITION BY UserID ORDER BY ReportDate) AS PrevReportDate
  FROM Reports
)
SELECT 
  UserID, 
  max(ReportDate - PrevReportDate) AS MaxTimeGap
FROM cte
GROUP BY UserID;

Playground result (wrong):

SELECT 
    MAX(TIME_TO_SEC(TIMEDIFF(t2.ReportDate, t1.ReportDate))) AS Max_Time_Gap
FROM
    Reports t1
INNER JOIN 
    Reports t2
ON 
    t1.UserID = t2.UserID AND 
    t2.ReportDate > t1.ReportDate 
GROUP BY
    t1.UserID

How could I reproduce those memory-less ChatGPT results with OpenAI API or Playground?