LinkedIn Software Engineer Interview Experience — Two of Four AI-Coding Tasks Completed

I just received a rejection. The feedback was reportedly that the AI-coding round had four tasks and I had only completed two.

I'd like to ask everyone, especially people with experience inside LinkedIn: what exactly is expected in this kind of AI-coding interview? Was I led down the wrong path?

First round: System design. Design a system to collect user activity, with queries for the last minute, hour, day, and similar periods. We mainly discussed the overall architecture, from API to middleware to backend storage and queries, along with how to store and query the data.

Second round: Coding. LeetCode 366, with a follow-up asking how to handle it if it were a graph.

Third round: AI coding. This is the part I want to focus on. It felt a little different from the other interview reports I had read on the forum, which is why I especially want to ask about it.

Before the interview, the recruiter told me that I couldn't rely entirely on AI; it was only an assistant.

At the start, the interviewer also told me clearly that I had to come up with the algorithm myself. I couldn't have AI produce the algorithm; AI was mainly for helping with implementation.

That seemed different from what other people on the forum had described, so I'm not sure whether this was a special requirement for my interview.

The problem was to implement an LRU cache.

I first explained the overall approach: a doubly linked list plus a hash map.

The interviewer repeatedly checked some implementation details. Once we had settled the algorithm and data structures, I started using AI to write the code.

CoderPad offered various models, including GPT, Gemini, and Claude. I chose Claude Opus 5.

But in practice, I found this AI incredibly hard to use, worlds apart from using AI for coding in my everyday work.

It often made very basic mistakes and wasn't consistent from one step to the next. For example:

• It updated the map when adding an entry.
• During eviction, it forgot to remove the corresponding entry from the map.
• Sometimes it invented variables that the problem had never defined.
• Code generated earlier and code generated later often didn't match up.

So it basically became: AI generates a section -> I review it line by line -> find bugs -> tell AI to fix them -> review again.

Throughout the process, AI didn't seem to solve many problems for me. Instead, I spent a lot of time checking whether it had written something wrong.

There were reportedly four coding tasks in this round:

1. Standard LRU.
2. LRU plus TTL.
3. I don't remember the details clearly.
4. Make it thread-safe.

I only finished the first two before time ran out. I wasn't getting stuck, either. I could have finished these problems by typing the code myself; using AI made it slower.

What I especially want to ask is: what is this kind of AI-coding interview actually testing?

Both the recruiter and the interviewer emphasized that I couldn't rely on AI to come up with the algorithm. That seems different from some AI-coding interview reports I had previously read on the forum.

I'd like to ask anyone at LinkedIn, or anyone who has recently had a similar round, to share what they think. What does this interview officially aim to assess?

And if the feedback is "only two of four tasks completed", does that usually mean progress was too slow, or could it mean they were unhappy with how I used AI or reviewed the code?

If anyone has had a similar experience, I'd welcome a discussion. 🙏

I'd like to ask everyone, or people inside LinkedIn, what this AI-coding round really expects. Was I led down the wrong path?

The fourth and fifth rounds were a project deep dive and a Host Manager interview, respectively. There isn't anything especially worth sharing about those.