Alibaba Dharma Institute: GPT-4 replaces an annual salary of 600000 yuan for data analysts with only a few thousand yuan, and the paper has been published

Dream morning comes from Aofei Temple
Quantum bit | official account QbitAI

The cost of replacing junior data analysts with GPT-4 is only 0.71%, while replacing it with senior data analysts is 0.45%

You're not mistaken, it's 0.71%, not 71%.

According to the Singapore market, annual salary86000 to 90000 US dollars (600000-630000 RMB)As a senior data analyst, switching to GPT-4 only requiresThree to four hundred US dollars (over 2000 RMB)Yes.

This conclusion comes fromAli Damo AcademyAndNanyang Polytechnic University of SingaporeThe new paper has been evaluated by netizens as a required reading paper interested in the fields of AI and data analysis.

Specifically, senior analysts in the conclusion refer to data analysts with years of work experience in the financial industry.

And the performance of GPT-4 can be achieved on most indicatorsAnd6The accuracy is lower than that of humans, but the complexity and consistency indicators are higher than those of humans.

And5GPT-4

And2GPT-4

But GPT-4 completes all types of tasks much faster than humans.

The final conclusion is reached on the premise that there are 21 working days per month, 8 hours of working hours per day, and the salary is paid at the market value.

What can GPT-4 do as a data analyst

The paper focuses on the following abilities of GPT-4 as a data analyst:

• Generate SQL and Python code
• Execute code to obtain data and charts
• Analyze data and external knowledge sources to draw conclusions

The experiment with 200 samples showed that forDraw Chart TaskGPT-4 is able to understand the meaning of instructions and has some background knowledge of chart types to draw correct charts.

Most of the charts are clearly visible without any formatting errors. The aesthetic indicators of the icons have a maximum score of 3, with an average GPT-4 score of 2.73.

However, manual inspection can still detect some minor errors, with a maximum score of 1 for chart accuracy indicators and an average GPT-4 score of 0.78.

The paper specifically states that their evaluation criteria are very strict, and any data or labels on the x-axis or y-axis that have errors will result in points being deducted.

aboutData analysis tasksGPT-4 achieved an average of full marks in consistency and fluency, verifying that generating fluent and grammatically correct sentences is definitely not a problem for GPT-4.

Interestingly, the accuracy of the data analysis step is much higher than the accuracy of the chart information, indicating that although GPT-4 drew the wrong chart, the analysis yielded the correct conclusion.

GPT-4And

• Human analysts can express themselves through personal thoughts and emotionsFor example, when analyzing, write 'Surprisingly...'; Human readers can easily understand whether data meets expectations or is abnormal from such expressions.
• Human analysts tend to draw conclusions based on background knowledgeAs written, "... is common in..."; GPT-4 usually only focuses on the extracted data itself, allowing GPT-4 to search for real-time online information online can improve this.
• When providing insights or suggestions,Human analysts tend to be conservativeIf there are no issues with the data; GPT-4 will provide advice directly in a confident tone without mentioning assumptions.

AndGPT-4

The final conclusion is:

GPT-4And

Paper:
https://arxiv.org/abs/2305.15038