Recently, on a whim, I backed up my ChatGPT conversation history, wanting to look back and see where this one-to-two-year journey of “human-computer collaboration” truly began. So, I used ChatGPT’s backup feature. In the backup file, I found what was nominally my “first question”:

時間： 2023-03-23 19:22:14
主題： 翻譯文章
內容： “Can you help me translate several articles from Traditional Chinese to English”

It looks perfectly reasonable: a university student, a translation task – a standard starting point for AI use. But looking at this date, something felt off: No, my “first time” was definitely not in March, but earlier, in late 2022. The topic was the “emotional blackmail detector.”

The Forgotten “Emotional Blackmail Detector”

My memory flashes back to December 2022. Back then, I was taking “Introduction to Text Mining,” and the final project required training an “emotional blackmail classifier.” We planned to use BERT for binary classification, predicting whether a piece of text contained emotional blackmail. To increase training samples, I turned to the recently launched ChatGPT and asked it to write a few emotional blackmail phrases.

But my first impression of it was actually terrible: “This thing is so dry and boring!”

Even though I told it to be creative, the sentences it generated were still cliché and stiff, completely incomparable to the nuanced blackmail found in real life. At the time, I was under the “curse of knowledge” – this “Introduction to Text Mining” course had covered GPT-2, and I had even downloaded GPT-2 from Hugging Face to play with it. I remember GPT-2 was introduced in late November, and back then, the professor mentioned that GPT-3 had just come out but he hadn’t researched it yet, suggesting it was probably similar to GPT-2. In the professor’s eyes, models like GPT-2, which were just “text completion games,” hadn’t shown much value yet. And in my perception, these types of models were just “slightly more advanced text completion games” – what was so great about them? I wasn’t an AI die-hard fan back then; I just thought it was a slightly better-performing model and had no idea it would completely transform the technological landscape for the next two years.

The Vanished “Dark History”

Before March 2023, I had actually been intermittently using this “text completion model” to handle some daily chores:

  • PR Crisis Management: Back when I was student council president, someone messed up, and I had to write a letter of apology asking for forgiveness. I put in the key points and let the AI polish it into a decent “PR draft.” Thinking about it now, it truly is a regrettable piece of dark history, lol.
  • Writing Fancy Stories: Occasionally, on a whim, I’d have it write some whimsical short stories.
  • Organizing Machine Learning Concepts: Back then, AI was still prone to hallucinations, so I didn’t dare let it “freely create.” I would copy and paste entire sections of textbook content, treating it like an “advanced search engine,” asking it to summarize the key points for me.

But Why Did These Records Disappear?

This brings us to the infamous “product disaster” in mid-to-late March 2023. At that time, OpenAI experienced a severe Redis bug, leading to many users’ conversation histories being leaked, disappearing, or even entire history logs being locked for several days. Looking back now, my records from December 2022 to March 20, 2023, were very likely victims of that chaos. Sam Altman really owes me an apology (laughs).

Related records can be found here: https://incidentdatabase.ai/cite/516/

March 2023: The Turning Point from Tool to Strategy

The real turning point indeed occurred in March 2023. That month, KOLs (Key Opinion Leaders) began an aggressive promotional campaign, and various applications emerged endlessly. At the time, I was busy with the Hult Prize competition, and this was my first time truly using AI “extensively” and “systematically”: I used it to evaluate competitors, generate promotional materials, and even simulate business strategies. From that moment on, AI was no longer just a template writer for apology letters; it became my strategic partner.

Conclusion: Stepping Outside the Finance Department Framework

Looking back at this backup record, what I find moving is not how much AI has advanced, but rather my own transformation. When I first took “Introduction to Text Mining,” I simply wanted to learn how to analyze financial data, without ever thinking about stepping outside the traditional framework of the finance department. Back then, looking at GPT-2 and ChatGPT, I merely saw them as interesting technical toys; now, I’ve dived headfirst into the world of LLMs, venturing much further than I ever imagined.

Perhaps those vanished records are like my “AI infancy” – though erased from the system, the traces they left are deeply imprinted in my current work mode.

Finally, to prevent data loss incidents from recurring, everyone is welcome to back up their ChatGPT data. The “Data Export” function is available in ChatGPT’s settings. After a successful export, you can scroll to the very bottom of the conversation history or write a small tool to query the earliest create_time entry! Here’s an example code:

import json
from datetime import datetime, timezone, timedelta
with open("conversations.json", "r", encoding="utf-8") as f:
    data = json.load(f)
# 過濾出有 create_time 的資料
valid_data = [d for d in data if "create_time" in d]
# 找最小 create_time
min_item = min(valid_data, key=lambda x: x["create_time"])
# 轉換時間（UTC -> 台灣 UTC+8）
utc_time = datetime.fromtimestamp(min_item["create_time"], tz=timezone.utc)
taiwan_time = utc_time.astimezone(timezone(timedelta(hours=8)))
# 輸出
print("最早時間：", taiwan_time.strftime("%Y-%m-%d %H:%M:%S"))
print("對應資料：")
print(json.dumps(min_item, indent=2, ensure_ascii=False))

Thank you for reading this far. Now, hurry and back up your ChatGPT history! You can find instructions on how to back up here. https://help.openai.com/zh-hant/articles/7260999-how-do-i-export-my-chatgpt-history-and-data