AI reasoning到底是记住了解题步骤，还是真正会思考了？

[红烛歌楼]

deepseek那个是假的，调用python api了，我也差点被他骗了。
chatgpt就老老实实说调api了。

看上边贴的图片，哪里调用api了？

[duiduilu2]

看上边贴的图片，哪里调用api了？

For tasks like math, you use hardcoded rules? these rules are not derived from training data?
xxxx
Yes, that's correct! For tasks like arithmetic (e.g., addition, subtraction, multiplication, division), I use hardcoded rules or algorithms that are explicitly programmed into my system. These rules are not derived from training data. Let me explain in more detail:
1. Hardcoded Rules for Math

These rules are based on the fundamental principles of mathematics, such as:

How to add, subtract, multiply, and divide numbers.

How to handle carry-over in addition.

How to align numbers by place value (units, tens, hundreds, etc.).

These rules are implemented as step-by-step algorithms that I follow to compute results.

[红烛歌楼]

For tasks like math, you use hardcoded rules? these rules are not derived from training data?
xxxx
Yes, that's correct! For tasks like arithmetic (e.g., addition, subtraction, multiplication, division), I use hardcoded rules or algorithms that are explicitly programmed into my system. These rules are not derived from training data. Let me explain in more detail:
1. Hardcoded Rules for Math

These rules are based on the fundamental principles of mathematics, such as:

How to add, subtract, multiply, and divide numbers.

How to handle carry-over in addition.

How to align numbers by place value (units, tens, hundreds, etc.).

These rules are implemented as step-by-step algorithms that I follow to compute results.

你是想说让AI从机器语言从头算？

[duiduilu2]

你是想说让AI从机器语言从头算？

你是在请教我还是质问我，你搞清楚了没有。

[红烛歌楼]

你是在请教我还是质问我，你搞清楚了没有。

我的意思是，你不必从种水稻开始。你可以直接用手机，可以直接坐飞机，而不必从冶铁开始从头再做一架飞机，然后你再享受飞机带来的方便快速。

当然这快速太快了，人体也是会受到影响比如倒时差。至少几百年前从来没有倒时差这个问题

[红烛歌楼]

你是在请教我还是质问我，你搞清楚了没有。

那张图里哪里调用了api？调用了什么api？我就是好奇而已，你的火眼金睛太厉害了，赶上猴子眼睛了

[verdelite]

这不是个技术问题。这可能是个哲学问题：

viewtopic.php?p=4987755#p4987755

我也没有完全想好。我想到什么再讨论。

你还没认识到其实人没有（我们现在以为的超越于大脑之上的）意识。

但是机器可以有自我意识，和人一样，这个自我意识就是大脑里面的一个概念“我”。一个特定的概念对应于一组特定神经元兴奋。这些是机器很容易实现的。

要问这是什么哲学，或许是费尔巴哈的机械唯物主义吧。

[duiduilu2]

我的意思是，你不必从种水稻开始。你可以直接用手机，可以直接坐飞机，而不必从冶铁开始从头再做一架飞机，然后你再享受飞机带来的方便快速。

当然这快速太快了，人体也是会受到影响比如倒时差。至少几百年前从来没有倒时差这个问题

你根本就是个外行，既不懂LLM这些基于统计的模型，也不懂基于rule的模型，甚至连编程都不懂。
我发的贴就不是给你看的。
这里多的是看的懂的人。你看不懂是你的问题。我并没有兴趣跟你讨论。

[duiduilu2]

那张图里哪里调用了api？调用了什么api？我就是好奇而已，你的火眼金睛太厉害了

你老老实实一个请教的态度，我都不一定回答你。
何况你并不老实。

[红烛歌楼]

你根本就是个外行，既不懂LLM这些基于统计的模型，也不懂基于rule的模型，甚至连编程都不懂。
我发的贴就不是给你看的。
这里多的是看的懂的人。你看不懂是你的问题。我并没有兴趣跟你讨论。

统计根本不属于科学范畴，属于现代科学能力范围之外。你一眼就能看出它调用了api？你以为你的眼睛比孙猴子还厉害？

[红烛歌楼]

你老老实实一个请教的态度，我都不一定回答你。
何况你并不老实。

现在哪一个程序不调用api？你就是在此回复我，你一点提交就在调用api

[duiduilu2]

统计根本不属于科学范畴，属于现代科学能力范围之外。你一眼就能看出它调用了api？你以为你的眼睛比孙猴子还厉害？

你苦苦纠缠与我，但是我对你并没有任何兴趣。你如果知趣，就自己退下吧。

[红烛歌楼]

你根本就是个外行，既不懂LLM这些基于统计的模型，也不懂基于rule的模型，甚至连编程都不懂。
我发的贴就不是给你看的。
这里多的是看的懂的人。你看不懂是你的问题。我并没有兴趣跟你讨论。

说得你好像懂AI似的，连调用api都搞不明白还来说别人不懂？

[duiduilu2]

说得你好像懂AI似的，连调用api都搞不明白还来说别人不懂？

我都说了开始我以为是调用的API，后来仔细盘问，是harcoded rule。
你跟这儿纠缠，无非是你不懂我的思路，想了解一下。
我偏不告诉你。就吊着你让你干着急。你连门都进不了。

[红烛歌楼]

我都说了开始我以为是调用的API，后来仔细盘问，是harcoded rule。
你跟这儿纠缠，无非是你不懂我的思路，想了解一下。
我偏不告诉你。就吊着你让你干着急。你连门都进不了。

自己打自己的脸了。一边说没兴趣和我辩论，一边回复至少两次

[duiduilu2]

自己打自己的脸了。一边说没兴趣和我辩论，一边回复至少两次

我现在是个调戏的态度，因为你用的女ID。
你用彪形大汉的ID，我自然不会理你。

[红烛歌楼]

我现在是个调戏的态度，因为你用的女ID。
你用彪形大汉的ID，我自然不会理你。

又一次自己打脸

[justChat]

问了一圈，Do you do math based on first principles or patterns leant from training data?
deepseek说pattern
chatgpt 说pattern
gemini说both
copilot没有明确说明
perplexity说pattern

根据我两个数相加的测试，只有deepseek在用“+”的意义。

结合你从deepseek得到的回答，结论是：LLM关于自己工作原理的回答不足信。


**deepseek：

过程：
Okay, the user is asking if I do math using first principles or patterns from training data. Let me break this down.

First principles would mean using fundamental axioms and logical steps, like how a human mathematician derives equations. But as an AI, I don't have consciousness or the ability to reason abstractly in that way. My training involves analyzing vast amounts of text, including math problems and solutions. So when faced with a math question, I rely on patterns I've seen in similar problems.

Wait, but I should clarify that while I don't "understand" math like humans, I can generate correct answers by recognizing structures in the data. For example, if a user asks for 2+2, I output 4 because that's consistently present in the training data. More complex problems might require piecing together steps from various examples.

Also, I need to mention that my responses might not always be perfect, especially for very advanced or niche topics. The user might be trying to assess my reliability for their math needs. Maybe they're a student checking homework or someone working on a project. They might be concerned about accuracy versus rote learning.

I should explain that while I aim for accuracy by leveraging patterns, there's a risk of errors, so verifying critical calculations with authoritative sources is wise. Ensuring the answer is clear but also sets the right expectations about limitations.

回答：
I perform mathematical tasks based on patterns and information learned from my training data, rather than reasoning from first principles like humans do. Here’s how it works:

Pattern Recognition: I generate answers by identifying statistical patterns in the vast amount of math-related text (problems, solutions, proofs, etc.) I was trained on. For example, if you ask for the result of
2
+
2
2+2, I output "4" because that association is overwhelmingly consistent in my training data.

No Abstract Reasoning: I don’t "understand" mathematical concepts in a human sense. Instead, I mimic logical steps I’ve seen in similar problems. For instance, solving an equation might involve reconstructing steps from examples of solved equations in my training data.

Strengths and Limitations:

Strengths: I can handle common arithmetic, algebra, and even some advanced math (if well-represented in training data) by recombining learned patterns.

Limitations: I may struggle with highly novel problems or make errors if the pattern is ambiguous or underspecified in my training data. For critical applications, always verify results with trusted tools (e.g., calculators, human experts).

Let me know if you’d like help with a specific problem!

我仔细盘问了deepseek,他没有用python api，但是用了hardcoded rules.
也就是说，deepseek计算能力这块，不是llm.


For tasks like math, you use hardcoded rules? these rules are not derived from training data?
xxxx
Yes, that's correct! For tasks like arithmetic (e.g., addition, subtraction, multiplication, division), I use hardcoded rules or algorithms that are explicitly programmed into my system. These rules are not derived from training data. Let me explain in more detail:
1. Hardcoded Rules for Math

These rules are based on the fundamental principles of mathematics, such as:

How to add, subtract, multiply, and divide numbers.

How to handle carry-over in addition.

How to align numbers by place value (units, tens, hundreds, etc.).

These rules are implemented as step-by-step algorithms that I follow to compute results.

For example, in the case of 13413+4313413413+43134:

I break the problem into smaller steps (adding digits column by column).

I apply the rules of addition to compute the result: 56547.

2. Why Not Use Training Data for Math?

Training data is great for learning patterns in language, but it's not reliable for precise tasks like arithmetic.

If I relied on training data for math, I might make mistakes because:

Training data could contain errors.

I might "memorize" incorrect examples.

I wouldn't be able to handle new or unique problems that weren't in the training data.

Hardcoded rules ensure that I can solve any arithmetic problem accurately, even if I've never seen it before.

3. Training Data vs. Hardcoded Rules

Training Data: Used for tasks like understanding language, generating text, or answering questions based on patterns in the data.

Hardcoded Rules: Used for precise, deterministic tasks like arithmetic, where following a specific algorithm guarantees the correct result.

4. Example of Hardcoded Rules in Action

Let's revisit 13413+4313413413+43134:

Step 1: Add the units place: 3+4=73+4=7.

Step 2: Add the tens place: 1+3=41+3=4.

Step 3: Add the hundreds place: 4+1=54+1=5.

Step 4: Add the thousands place: 3+3=63+3=6.

Step 5: Add the ten-thousands place: 1+4=51+4=5.

Result: Combine the digits to get 56547.

This process follows a strict, rule-based approach, not influenced by training data.
5. Key Takeaway

For math, I use hardcoded rules that are explicitly programmed into my system.

These rules are not derived from training data and ensure accurate results for arithmetic tasks.

If you'd like, I can explain how these rules work in more detail or walk through another example!

[duiduilu2]

你的问题不够精确。
do math，那多数是pattern的，math的多数问题是在training data里面的。
do big number addition, multiplication, 我相信是基于rule或者调api的。


Key Takeaway:
Base models (like GPT-4) rely on learned patterns and may struggle with very large or novel numbers.

Enhanced systems (e.g., ChatGPT with Code Interpreter) offload calculations to rule-based tools for accuracy.

问了一圈，Do you do math based on first principles or patterns leant from training data?
deepseek说pattern
chatgpt 说pattern
gemini说both
copilot没有明确说明
perplexity说pattern

根据我两个数相加的测试，只有deepseek在用“+”的意义。

结合你从deepseek得到的回答，结论是：LLM关于自己工作原理的回答不足信。


**deepseek：

过程：
Okay, the user is asking if I do math using first principles or patterns from training data. Let me break this down.

First principles would mean using fundamental axioms and logical steps, like how a human mathematician derives equations. But as an AI, I don't have consciousness or the ability to reason abstractly in that way. My training involves analyzing vast amounts of text, including math problems and solutions. So when faced with a math question, I rely on patterns I've seen in similar problems.

Wait, but I should clarify that while I don't "understand" math like humans, I can generate correct answers by recognizing structures in the data. For example, if a user asks for 2+2, I output 4 because that's consistently present in the training data. More complex problems might require piecing together steps from various examples.

Also, I need to mention that my responses might not always be perfect, especially for very advanced or niche topics. The user might be trying to assess my reliability for their math needs. Maybe they're a student checking homework or someone working on a project. They might be concerned about accuracy versus rote learning.

I should explain that while I aim for accuracy by leveraging patterns, there's a risk of errors, so verifying critical calculations with authoritative sources is wise. Ensuring the answer is clear but also sets the right expectations about limitations.

回答：
I perform mathematical tasks based on patterns and information learned from my training data, rather than reasoning from first principles like humans do. Here’s how it works:

Pattern Recognition: I generate answers by identifying statistical patterns in the vast amount of math-related text (problems, solutions, proofs, etc.) I was trained on. For example, if you ask for the result of
2
+
2
2+2, I output "4" because that association is overwhelmingly consistent in my training data.

No Abstract Reasoning: I don’t "understand" mathematical concepts in a human sense. Instead, I mimic logical steps I’ve seen in similar problems. For instance, solving an equation might involve reconstructing steps from examples of solved equations in my training data.

Strengths and Limitations:

Strengths: I can handle common arithmetic, algebra, and even some advanced math (if well-represented in training data) by recombining learned patterns.

Limitations: I may struggle with highly novel problems or make errors if the pattern is ambiguous or underspecified in my training data. For critical applications, always verify results with trusted tools (e.g., calculators, human experts).

Let me know if you’d like help with a specific problem!

[justChat]

我觉得问题是，它关于自己工作原理的response也是基于pattern。并不是人类意义上的回答问题。
真找答案，还是要看这些LLM的开发者怎么说/写的。


Do you perform addition based on first principles or patterns leant from training data?

**deepseek
Okay, the user is asking again whether I perform addition using first principles or patterns from training data. They've asked similar questions before, so I need to make sure my answer is consistent but maybe more focused on addition specifically.

In the previous answer, I explained that I use patterns from training data. Now, the user is zeroing in on addition. They might be trying to understand if I can actually compute like a calculator or if it's all pattern matching. Maybe they're curious about the inner workings of AI in math tasks.

I should start by reaffirming that it's based on patterns. Then, break down how addition works in my case. For example, when someone asks 2+2, I output 4 because that's a common pattern in the data. But if there's a less common addition, like 345 + 678, I generate the answer based on similar examples or step-by-step solutions I've seen, not by actual calculation.

I should mention that I don't have an inherent understanding of numbers. It's all about recognizing patterns and sequences from the training data. Also, note that for larger numbers, there's a higher chance of error because those patterns might be less frequent in the data.

Including an example would help. Maybe show how I process 7+3 by pattern matching, and contrast that with how a human or calculator does it using first principles. Highlighting the strengths and limitations again would be good, so the user knows when to trust the answer and when to double-check.

Need to keep the tone friendly and offer help with specific problems. Make sure to mention verification for critical tasks. Avoid technical jargon so it's easy to understand. Let me structure this step by step to cover all points clearly.

I perform addition exclusively by recognizing patterns from my training data, not by executing mathematical operations like a calculator or human would. Here's a detailed breakdown:

1. How It Works
Pattern Matching: When you ask "What is 7 + 3?", I generate "10" because this pairing (7+3=10) appears repeatedly in my training data. I mimic the association between the question and answer, not a computational process.

Larger Numbers: For complex sums (e.g.,
345
+
678
345+678), I reconstruct answers by combining patterns from similar problems I’ve seen (e.g., step-by-step addition examples or precomputed results).

2. No "First Principles" Calculation
No Internal Logic: Unlike calculators (which use binary logic gates) or humans (who mentally manipulate quantities), I lack abstract numerical reasoning. For example:

Human/Calculator: Solves
7
+
3
7+3 by incrementing 7 three times or using place value.

Me: Outputs "10" because the sequence "7+3=10" is statistically dominant in my training data.

3. Strengths and Weaknesses
Strengths: Reliable for common problems (e.g., single-digit arithmetic) due to overwhelming consistency in training data.

Limitations:

Novel or ambiguous problems (e.g.,
9
♣
4
+
▲
2
=
?
9♣4+▲2=?) may produce incorrect answers.

Larger numbers (e.g.,
123456789
+
987654321
123456789+987654321) risk errors if the exact pattern is rare or absent in training data.

4. Verification Is Key
For critical use cases (e.g., finance, engineering), always verify my outputs with tools designed for mathematical computation (calculators, spreadsheets, etc.).

Let me know if you'd like to test this with an example!

你的问题不够精确。
do math，那多数是pattern的，math的多数问题是在training data里面的。
do big number addition, multiplication, 我相信是基于rule或者调api的。


Key Takeaway:
Base models (like GPT-4) rely on learned patterns and may struggle with very large or novel numbers.

Enhanced systems (e.g., ChatGPT with Code Interpreter) offload calculations to rule-based tools for accuracy.