| Interesting. I pasted in the outage report, and then asked ChatGPT to give me a single sentence summary and it gave me an answer way longer than a sentence. I pointed out that it wasn't a single sentence and it gave me: > ChatGPT experienced an outage due to a bug in an open-source library that allowed some users to see titles from another active user's chat history and payment-related information of 1.2% of ChatGPT Plus subscribers was exposed; the bug has since been patched and actions have been taken to prevent future incidents. I then asked it why there was the outage and it spit out: > The exact cause of the Facebook outage is unknown, but it is believed to be related to a configuration change in the company's routers. It's likely because I ran out of tokens because the OpenAI outage report is long. Pasting in the text of the outage report, and then re-asking about why, it was able to give a much better answer: > There was an outage due to a bug in an open-source library that allowed some users to see titles from another active user's chat history and also unintentionally exposed payment-related information of 1.2% of ChatGPT Plus subscribers who were active during a specific nine-hour window. Querying it further, again having to repeat the whole OpenAI outage report, and asking it a few different ways I eventually managed to get this succinct answer: > The bug was caused by the redis-py library's shared pool of connections becoming corrupted and returning cached data belonging to another user when a request was cancelled before the corresponding response was received, due to a spike in Redis request cancellations caused by a server change on March 20. It did take me more than a few minutes to get to there, so just actually reading the report would have been faster, and I ended up having to read the report to verify that answer was correct and not a hallucination anyway, so our jobs are safe for now. |