1分鐘了解ChatGPT為什麼可以如此強大

Описание: 1分鐘了解 ChatGPT為什麼可以如此的強大？

我們以2022年初OpenAI發表的相關論文，以及網路上收集到的資料，來猜測最少會有4個步驟來練成

第一步：以175B的模型參數大小，訓練大量文字內容資料，讓模型學會文字接龍
第二步：在不同的自然語言處理任務上，以監督式的學習來Fine Tuning model，俗稱SFT
第三步：以人類來train一個Reward Model，可以幫ChatGPT的雛型生成的內容打分數
第四步：利用PPO強化學習演算法，來調整第二步SFT完的模型，最後就練成ChatGPT

1 minute to understand why ChatGPT can be so powerful?

Based on the relevant papers published by OpenAI in early 2022 and the information collected on the Internet, we guess that there will be at least 4 steps to practice

Step 1: With a model parameter size of 175B, train a large amount of text content data, so that the model can learn word solitaire
Step 2: On different natural language processing tasks, use supervised learning to Fine Tuning model, commonly known as SFT
Step 3: Train a Reward Model with humans, which can score the content generated by the prototype of ChatGPT
Step 4: Use the PPO reinforcement learning algorithm to adjust the model after SFT in the second step, and finally practice ChatGPT