After LeetCode Exits the Stage, AI Interviews First Test How You Read an Unfamiliar Codebase

LeetCode is no longer tested, but AI coding interviews have not become easier. At least judging from one candidate's Bloomberg interview experience, the difficulty has simply moved elsewhere.

Baoyu mentioned in a tweet that although LeetCode is no longer tested, testing AI coding is not necessarily easy, and is even more mysterious. He cited a Reddit poster's Bloomberg interview experience, in London or Frankfurt, where the second round was an AI codebase interview.

The poster said the assessment covered Python's Django framework, but he knew nothing about that framework, and the entire codebase was very large. The first 30 minutes were all behavioral interview, leaving only 30 minutes for coding.

What made it more troublesome was that the boundaries for using AI tools were very vague. The interviewer did not explicitly say how much AI use was allowed, nor clarify their expectations. The poster used planning mode and ask mode, advancing one feature at a time, passed a few test cases, and was still rejected in the end.

He mentioned one detail: when using AI assistance, you have to read a large amount of generated content, and if you also have to communicate with the interviewer at the same time, it is almost impossible to finish. In the past, when grinding LeetCode, at least the problem was clear, and if you got stuck you could still brainstorm. Now, just understanding a huge codebase and figuring out where the problem is already consumes half the time.

The post also warned: do not think you can rely on a single prompt to have AI instantly kill the entire problem, otherwise you will still be rejected. The poster's friend encountered exactly this situation.

In the comments, someone put the problem very directly: if the interviewer themselves cannot clearly explain how much AI can be used, then what exactly is being tested: writing code, or guessing the interviewer's mind? Others said that in the past people memorized patterns by grinding problems, but now you have to understand an unfamiliar codebase on the spot, and the difficulty has shifted from algorithms to reading code.

Some netizens mentioned that the unwritten rule of 'you cannot one-shot it' is the funniest part. Others believed that if candidates are given only 30 minutes to write code and the boundaries of AI use are not defined, then what candidates are preparing for is really guessing the question, and the recruiting side should first write down the scoring criteria.

One netizen used his own experience of having Codex collect data and edit videos as an example. His rule was to receive only a one-page summary, and he himself only looked at the conclusions and the reasons for what was not obtained. A while ago, Codex reported for several days in a row that data was not obtained, and upon checking it turned out that browser bridge program protocol 1.7 and extension 1.8 did not match. Being able to understand why it stopped is harder than making it write more code.

Another netizen said that when asked to write a paper, Codex ran by itself for 11 hours and 3 minutes and did not write a single word of the main text. In an interview with only 30 minutes, there is simply no time to discover this kind of spinning in place, let alone while also talking to the interviewer. So it is not that you no longer need to grind problems; it is that acceptance and verification have become harder.

Some netizens also believe that most code is now written by AI, and the real work has become understanding and verifying what it writes. The difficulty lies in unclear boundaries: you simply do not know up to what point using it counts as your own ability, and at what point it will be treated as cheating. Someone summed it up even more sharply: in the past, they tested whether you could write code; now they test whether you dare click approve without reading the code AI wrote.

There is also a long comment in the replies that deserves separate attention. BG-VC believes that the quality of commercial-grade software depends not only on whether the code can run, but also on understanding business requirements, system boundaries, and long-term evolution, as well as trade-offs among architecture, security, reliability, and maintainability. Large models still have limitations in these areas. Anthropic has pushed front-end, 3D, and other visually intuitive coding capabilities in front of the public. These demos are impactful, but cannot fully reflect the engineering complexity behind commercial software. Programming languages, frameworks, and engineering practices form a huge body of knowledge, and back-end architecture, design patterns, and concrete implementation require judgment based on business scenarios and constraints. Anthropic's and OpenAI's demonstrations of Agent and Skill, intentionally or not, make the public think that calling a few scripts and chaining a few tools together is complete Agent development, thereby underestimating the real engineering effort of commercial-grade systems.

This comment explains why AI interviews easily go off track. Companies want to test code reading, verification of AI output, communication, and judgment. These abilities are indeed important in real work. But throwing a candidate into an unfamiliar large codebase, giving them only 30 minutes, and also requiring a behavioral interview at the same time makes the sample too narrow. If tool boundaries and scoring criteria are not written clearly, candidates can only guess.

Some people in the comments also think this is a good thing. Netizen Shinda said that he had been grinding problems since childhood, and that exams in China inherently mass-produce standardized parts: the problem is clear, the answer is unique, and you are not allowed to think too much. AI interviews make people panic precisely because real work inherently has no standard answer. Being able to understand a huge pile of code while also communicating with people as you work, that is the real skill. The era of the problem-grinding machine coming to an end is a good thing.

This view makes sense. But judging from this Bloomberg interview, AI interviews have not yet solved a basic problem: is it actually testing programming ability, tool-use ability, or the set of invisible rules in the interviewer's mind?

Original post link: Reddit discussion

Tweet link: Baoyu's tweet