人工智能入门

原来大预言模型，每输出一个 Token 就要做一次计算，是根据过去所有的文字，来预测下一个字。

原来图像生成Diffusion，是先把好的图片一步一步加噪声到纯噪声图片，然后再让网络一步一步预测回去训练的。生成图片就是用噪声生成图片。

Zero-shot Prompting：只让 LLM 做事情

K-shot Prompting：给几个例子，适合特定领域的 API、生成特定风格的代码。不要用于知名库和通用编码任务，以及过度的限制和例子。

1
Write a for-loop iterating over a list of strings using the naming convention in our repo.
1
2
3
4
5
6
7
8
9
Write a for-loop iterating over a list of strings using the naming convention in our repo. Here are some examples of how we typically format variable names.

<example>
var StRaRrAy = [‘cat’, ‘dog’, ‘wombat’]
</example>

<example>
def func CaPiTaLiZeStR = () => {}
</example>

chain of thought prompting.Show reasoning steps for a given task.COT Chain of Thought，思维链。

最适合需要包含多个逻辑步骤的任务

1
Write a function to check if a number is a perfect cube and a perfect square.
1
2
3
4
5
6
7
8
9
10
11
12
I want to write a function to check if a number is a perfect cube and a perfect square. Make sure to provide your reasoning first. Here are some examples of how to  provide reasoning for a coding task.

<example>
Write a function that finds the maximum element in a list.
Steps: Initialize a variable with the first element. Traverse the list, comparing…
</example>

<example>
Write a function that checks is a number is a palindrome
Steps: Take the number. Reverse the elements in the numbers. Check if …
</example>

zero-shot cot, multishot cot.

self-consistency prompting：让LLM输出多次，取个最好的

让 LLM 能够 defer to 一个外部系统，遵从外部系统，使用工具，也是降低幻觉的好办法。

1
After you have fixed this IndexError can you ensure that the CI tests still pass?
1
2
3
4
5
6
7
Fix the IndexError. Ensure the CI tests still pass once you have made the fix. Here are the available tools. 

<tools>
pytest -s /path/to/unit_tests

pytest -v /path/to/integration_tests
</tools>

Retrieval Augmented Generation，RAG，检索增强生成，向 LLM Infuse（注入）上下文数据，让LLM在没重新训练的情况下保持最新状态。

1
Extend the UserAuthService class to check that the client provides a valid OAuth token.
1
2
3
4
5
6
7
8
9
10
11
12
I want to extend the UserAuthService class to check that the client provides a valid OAuth token. 

Here is how the UserAuthService works now:
<code_snippet>
def issue_oauth_token():
	….
</code_snippet>

Here is the path to the requests-oauthlib documentation:
<url>
https://requests-oauthlib.readthedocs.io/en/latest/
</url>

Reflexion 就是让 LLM 能够根据它自己的输出来进行反映，这是目前全代理行为在IDE中的核心行为

系统级提示，可以做一些整体的风格。

向上下文信息有限的人提供提示，如果他们感到困惑，LLM（语言学习导师）也会感到困惑。

Prompt的最佳实践，Prompt应该用Structure来结构化：

1
2
3
4
5
6
7
8
Here are the logs:
<log>
LOG MESSAGE
<log> and the stack trace:
<error>
STACK TRACE
<error>

确切地明白自己要做什么，分解任务。