FFmpeg 8.0 adds Whisper support (code.ffmpeg.org)

For years, many companies have leaned heavily on leetcode-style algorithm problems in their interviews. I know many of us have always felt this was a questionable signal for real-world engineering skill, since most software engineers rarely (if ever) write novel algorithms during their day-to-day.

Now, with LLMs performing extremely well on these problems, the gap between “can solve interview puzzles” and “can do the job” seems wider than ever. In reality, a competent engineer today is quite likely to use an LLM to help solve such problems, so now the skill we’re testing in an interview isn’t even how they’d work in practice.

So my question is: Where do you think the software engineering interview process is heading next? Will we move toward more real-world scenarios? Systems design? Pair programming with AI in the loop? Something else entirely?

Curious to hear what trends you’re seeing or experimenting with at your companies.

Comments (7)

gooodvibes · 14h ago

Leetcode was never about being similar to real work. Companies aren't stupid - they know it's different and still think that they get useful signal from testing skills that don't match the day-to-day of the job. Whether that's appropriate or not is where the actual debate around the value of Leetcode is; the debate isn't whether leetcode is like day-to-day work, nobody is claiming that it is.

And LLMs don't make the difference you think they make to the viability of LC interviews. Before LLMs, you could solve an interview problem by looking it up in a database with solutions. Or (in real work) you solve these algorithmic problems by importing a library. LLMs are just the (n+1)th way to do the same thing.

Where LLMs play some role is that they make cheating in remote leetcode interviews easier.

gcheong · 18h ago

I don't know where it's going but as of this moment Meta at least is still using them and it doesn't look like they're going to change anytime soon. Source: I have an upcoming interview with them. If I'm to believe what they tell me, you're not just scored on how well you can implement a solution, but how well you can explain in real time what your approach will be before you start coding and how well you communicate during the implementation. I was told to overcommunicate if anything even to the level of correcting typos. I don't think LLM's can do that yet.

CureYooz · 3h ago

I'm not a fan of leetcode interviews but I'm also not a fan of not being able to invert a binary tree. Programmers who are avoiding graph problems like a plague are programmers who are sawing off a branch they are sitting on. If you have no time to spend at least a month exploring graph problems then do not expect to have a multi decades career in programming.

shahbaby · 19h ago

Even if the LLM solves the question, you still have to explain the answer as if it were your own in a timed interview setting where there will be another human who will likely ask followup questions to see if you really did understand it.

phba · 17h ago

Junior and remote positions get so many applications that leetcode tests are the only reliable filter to even get a reasonably sized set of applicants to choose from. So leetcode won't go away until the number of applicants drops or we find a better way to figure out if someone can code or not.

CureYooz · 3h ago

Work in previous projects is a good estimation. But it's going to shorten interviewing part only in smaller companies. In corpos you most likely still end being interviewed the same way as anyone else.

Ocerge · 19h ago

In my experience: definitely not, still plenty of the same leetcode tech screens I've seen over the past decade. But I could see more in-person interviews + white boarding coming soon, it's the only surefire way to keep people from cheating.