We tell you how a robot engineer was interviewed at YuMoney and which of the tasks he did at the level of a middle developer. At the same time, II bypassed the real candidate Yuriy, who claimed to have three years of experience, but actually turned out to be a junior.

A typical interview at the YuMoney backend development department looks like this:

• We talk with the candidate for 5-10 minutes on abstract topics, we discuss the experience of the applicant, warm him up.

• It lasts 25 minutes practical section I — the candidate pokes around his screen and writes code in the editor in real time.

• It takes 35 minutes practical section II – We give the candidate a more difficult task.

• 15 minutes discuss the practical section.

• It lasts 30 minutes theoretical section.

As a result, an interview with one candidate lasts two hours without breaks If he successfully copes with all the tasks, we decide to hire him or not.

During the interview, we evaluate:

• How the candidate works with the requirements.

• How does he formulate the algorithm for solving the problem?

• Can he implement the algorithm in the time we have allocated for it.

Which candidate can pass an interview at YuMoney

• The one who correctly completed the task. Minor errors are acceptable.

• The one who invested in the timing.

As a rule, they do not cope with the first practical task 90% of candidates of any level. Even middles and seniors can make mistakes and not stick to the timing.

And now the dessert: we tell you how it coped with the test ChatGPT-4

There were three test tasks. In the first, we asked ChatGPT to calculate the amount of cashback as a fixed percentage of the total purchase amount.

Example:

• The total amount of the order is 100 rubles.

• The percentage of cashback is 2%.

• The amount of cashback will be 2 rubles (100 * 0.02).

It was a super simple task, not a combat one. We usually do not give such tests to real candidates.

The result of the first test: ShatGPT-4 did well

We can’t show you what the code that generated the AI ​​looked like, but what we learned when we got a test run of it:

• The neural network performed the test correctly.

• Used actual function calls.

• In general, we liked the result of the work.

The second task for ChatGPT was more difficult

It was necessary to write unit test: is a program that checks how correct the finished code is We asked the artificial intelligence to check its own code.

As a result, the neural network correctly formed the input data, but the result of the cashback calculation based on this data turned out to be incorrect.

The finals were decided as a real combat task for the candidates

We asked for ChatGPT calculate “progressive” cashback, which depends on the amount of the order.

I would like to note that we gave this task to a real candidate – a Java programmer (let’s call him Yuriy). According to Yuri, he had three years of work experience, two of which he was optional in Java development. So, the neural network coped with the task better than Yuriy, and 25 minutes faster. ChatGPT-4 completed the task at the level of a midshipman, and Yuriy, with three years of experience, at the level of a trainee.

But that’s not all: we fed ChatGPT a code censoring task

It was interesting to see if the neural network would suggest reworking anything in the benchmark solution. And she offered: she noted that you can use an already sorted structure to store cashback percentages, that is, do not do sorting during the cashback calculation.

But on this ChatGPT blew up – it made a mistake with specifying the API. To store sums, an object class was used, with which the neural network performed actions, such as a primitive type. Of course, the code didn’t compile because of this.

Conclusion

• ChatGPT-4 was interviewed as a minimum on June.

• Sometimes the neural network is wrong, but in general it can be a good helper.

• We did not accept ChatGPT-4 works because we can’t share bits of combat code with it. 😉

Neuronet can offer ideas for a code review – and even write the code itself. Therefore, we think that artificial intelligence is something that will change our lives in the future.

But it is still too early to fear that AI will replace programmers: that is still a long way off. It is better to be smarter than the neural network and learn to manage the processes that it can do instead of a person. So let’s sleep peacefully for now. 😉

---

Share if you also interviewed ChatGPT – what were the results? Would you hire such an employee?