ChatGPT降智求問

如題,ChatGPT free tier,各種檢測方式都測過了,沒發現降智,但實際提問讓它思考的時候它老是偷懶,這是對的嗎? :tieba_087: 各位佬快來幫忙看看 :lark_046:

Tool summary,穩定 6 個:

chat-requirements,61a80,感覺也不低:

登入狀態成功 trigger search:

不登入(無痕模式)的話則失敗:

IP是家寬,上面測試應該也能體現 IP 質量如何了,這裡不多分析 IP 了

來看提問:

我總共問了三題,每題的 prompt 都大致相似,都明確告訴了它「Think longer before giving the answer.」,且 request 的時候開啟了 Think 的 tool。

第一題,thought for 4 seconds:

展開只有一段思考,然後就 Done 了:

第二題,儘管我明確讓它思考,但並未成功 trigger 思考,可能是題目較為簡單:

第三題,這次我直接傳圖,而不是我自己 OCR 後再問,成功 trigger 思考,thought for 4 seconds,思考內容有兩段,可以看到第二段貌似被截斷:

是我對 ChatGPT free 的期望太高,還是我被降智了而不自知?

這個 Think 我用得真的不舒服

Prompt 1

This is a problem about database design and concepts. Think longer before giving the answer. You should ensure the accuracy and correctness of the solution. You need to give the steps of solving the problem. Here the problem starts: :rocket:

Let relations r_1 (A, B, C) and r_2 (C, D, E) have the following properties: r_1 has 20,000 tuples, r_2 has 45,000 tuples, 25 tuples of r_1 fit on one block, and 30 tuples of r_2 fit on one block. Estimate the number of block transfers and seeks required using each of the following join strategies for r_1 \join r_2:
a. Nested-loop join.
b. Block nested-loop join.
c. Merge join.
d. Hash join.

Prompt 2

This is a problem about database design and concepts. Think longer before giving the answer. You should ensure the accuracy and correctness of the solution. You need to give the steps for solving the problem. Here the problem starts: :rocket:

Consider the bank database of Figure 15.14 (see below), where the primary keys are underlined. Suppose that a B±tree index on branch_city is available on relation branch, and that no other index is available. List different ways to handle the following selections that involve negation:

a. \sigma_{\neg(branch_city<“Brooklyn”)} (branch)
b. \sigma_{\neg(branch_city=“Brooklyn”)} (branch)
c. \sigma_{\neg(branch_city<“Brooklyn”\vee assets<5000)} (branch)

branch (\underline{branch_name}, branch_city, assets)
customer (\underline{customer_name}, customer_street, customer_city)
loan (\underline{loan_number}, branch_name, amount)
borrower (\underline{customer_name}, \underline{loan_number})
account (\underline{account_number}, branch_name, balance)
depositor (\underline{customer_name}, \underline{account_number})
Figure 15.14 Bank database.

Prompt 3


In the image, there are several problems with database design and concepts (to be accurate, 3). You should first analyze and understand Exercise 15.20 and then solve it. Think longer before giving the answer. You should ensure the accuracy and correctness of the solution. You need to give the steps for solving the problem. Note: I didn’t ask you to solve Exercise 15.3, it is provided as background information along with Exercise 15.19 of the Exercise 15.20.

转Claude吧,都快2年没把GPT当主力了,降智降的忒恶心了

首先第一个图,回复里的DALLE说明画图被降智了,DALLE现在已经被4o画图替代了
第一个问题在我这里思考了16秒,不确定是不是降智

可以找一下mozi的题库测一下,免费用户应该是o4-mini,看看题单上o4-mini能做出来的题在自己这里是否能做出来

感觉plus的还不错(不降智的情况下),不过gpt40mini-high跟这个o3差得太多了,o3确实很智能

可以把题放出来帮你试试

我记得站内始皇有讲过关于降智的话题,主要是ip干净就一般不会

已編輯 ㅤ

第一题


第二题

第三题

好棒!

但是免費用戶用的好像還是 GPT-4o

剛才 linux[.]do 好像崩了,我 ERROR 522

如果开think就是o4 mini了

而且降智已经算恶心了,抗拒执行就更过分了,以前的3.5至少勤快,现在的没有一次输出不偷懒的,除了比较考验逻辑的任务,它其他的啥也干不了,而且有些活我明知道他能干就是不干,就更不爽了,人家隔壁Claude就很爽快