Fu Qiang said that an architect ‘s thinking and precipitation Menu

Why is the Khan Academy supported by ChatGPT more awesome?

Wang Fuqiang

2023-05-06

---

• 1 1V1
• 2 Context Matters (a lot)

---

Khan, the founder of Khan Academy, recently gave a speech on TED that attracted the attention of many people, and the effect of the live demonstration is also very good. If you are interested, you can go to station B to watch it yourself (the official channel of TED on Youtube is of course also available)

[The latest TED speech by Khan, the founder of Khan Academy: GPT-4, as AI learning private education, may bring the biggest change in the history of education (bilingual version, human-machine co-creation)] https://ift.tt/zuiDR56

Here, Mr. Fuqiang is simply saying two things about his own feelings…

1 1V1

In fact, as early as 1984, many studies have shown that the effect of private education or one-on-one real-person tutoring is better than that of “big pot rice”. The indicators are drawn and drawn, and someone gives them a hand to help them, and they fumble around with themselves. There are huge differences in effects and time costs, which eventually lead to huge gaps.

One day I had dinner with the cofounder of Tongyu Technology. At the dinner table, he expressed a similar point of view. To put it bluntly, he can grow rapidly and even his leaders in the workplace can grow rapidly. With someone who takes you one-on-one, you have a “personal trainer”.

We often complain that traditional education is rote memorization, but I find that education in other regions puts more emphasis on muscle memory. Why? Because only by emphasizing muscle memory first can you experience it personally, and if you don’t get started, you will never be able to get started!

The end of relying on rote memorization is that after ChatGPT comes out, it will be crushed directly, because its ability to achieve by rote memorization is far stronger than you alone.

• Counting the investment, you don’t have as much investment as it (hundreds of thousands of dollars and millions of dollars are burned every day);
• Computing power, you are not as powerful as it;

Therefore, rote memorization has no future. It is better to delegate this thing to ChatGPT, and then let it be your assistant or personal trainer. Rather than relying on personal ability and energy to do better than ChatGPT, it is better to be different of myself 😉

2 Context Matters (a lot)

I think this picture is the most essential part of the whole speech. No matter how good the whole speech is, it is just a superficial effect, but what is going on behind the scenes can only be understood by this picture.

In order to achieve the best effect of the speech, Khan Academy has adjusted the knowledge in the field of mathematics for 6 months on the basis of the existing model. At the same time, there is still a long period of fine-tuning of the model. Model fine-tuning? The key is one word: Context !

Context information (Context) is the key to the “superpower” of the GPT model . It is of course possible to simply give ChatGPT a simple command. In relatively simple and clear scenarios, such as word translation, it can indeed give you a good answer, but In a relatively complex scene, if you want to achieve good results, you need to give it more context information.

I talked to a deputy chief physician about AI very late yesterday, and then gave him a few ChatGPT entrances to experience. The feedback he gave was not satisfactory. The reason is that he did not give more and more detailed context to GPT. This is actually the same as when we write manuscripts and PPTs. We have to go through multiple revision iterations to get better results.上下文-结果-修正-反馈, after multiple cycles, we can get a relatively satisfactory result. the result of.

Therefore, ChatGPT is not as omniscient as many people imagine, and even becomes your “roundworm in the stomach”. In fact, it is more like a “device”, a mirror , what can give you, and more about yourself;)

If you want to use ChatGPT well, you still need to learn how to draft the prompt. It is of course better to have tool assistance, but the general idea and limitations must be clear. In fact, prompt is essentially a kind of model fine-tuning (fine tune), but it is more generalized and low-frequency fine-tuning, rather than focusing on specific fine-tuning.

---

Finally, I want to make an advertisement. If you want to know more about AI knowledge and tools, including but not limited to GPT, AI painting and other fields, welcome to join the “crazy for AI” knowledge planet. In addition to timely and fresh AI information, there are nearly 300 Minutes of paid courses, you can get it for free when you join the planet:

---

---

---

© Fuqiang Wang Personal Copyright, All Rights Reserved.