166 karma · joined September 28, 2024
Submission numbers this year have been absolutely crazy. I honestly don't think it can be solved.
The way it appears to me, if there's multiple commits submitted as separate PRs, then maybe the PR wasn't so atomic to begin with.
If I want that granularity, I'd go read the original PR and the discussion that took place.
Oxbridge will always be tier 1. Undergraduate teaching quality + a rigorous selection process cannot be beat.
Microsoft is a huge company with a lot of momentum and so many products that you can't reasonably conclude anything about remote work from this stat alone. Worse, you don't even account for inflation (!).
What about revenue per employee before covid? What if I chose any arbitrary time period in the last 20 years? Lastly, maybe revenue per employee could've grown significantly more during 2020-2024 if everyone was in the office?
I'm in my 20s with good credentials and have quite a few friends in the startup world. I would never feel comfortable interviewing someone with 10+ years of industry experience.
If your strategies care so much about performance, can you really achieve the same or comparable performance with Cython? At that point, wouldn't those strategies be better suited for a custom "re-implementation"?
I suppose I'm not exactly sure how "high performance" and "low latency" this project is. Are there any latency stats on this project?
Speaking seriously, I agree there's definitely a lot of bloat in the new C++ standards. E.g. I'm not a fan of the C++26 linalg stuff. But most performance-focused trading firms still use the latest standard with the latest compiler. Just a small example of new C++ features that are used every day in those firms:
Smart pointers (C++11), Constexpr and consteval (all improvements since C++11), Concepts (C++20), Spans (C++20), Optional (C++17), String views (C++17)
Even with incremental builds, that surely does not sound plausible? I only mentioned C++ because that's my main working language, but this wouldn't sound reasonable for Rust either, no?
I'm also not sure how LLM could guess what the tests should be without having written all of the code, e.g. imagine writing code for a new data structure
I think long lecture playlist is a feature, not a bug. It's much harder to commit to such material when you're not full timing education.
In my experience, LLMs are great at small tasks (bash or python scripts); good at simple CRUD stuff (js, ts, html, css, python); good at prototyping; good at documentation; okay at writing unit tests; okay at adding simple features in more complex databases;
Anything more complex and I find it pretty much unusable, even with Claude 4. More complex C++ codebases; more niche libraries; ML, CV, more mathsy domains that require reasoning.
I only code for x86 with vectorclass library, so I never had to worry about portability. In practice, is it really possible to write generic SIMD code like the example using Highway? Or could you often find optimization opportunities if you targeted a particular architecture?
The limiting factor in practice, as with everything quantum, is noise. You are right -- we don't know for certain until it's implemented. I suppose it's part of a bigger question: whether quantum computing will work at all. My knowledge of quantum hardware is limited, so I can't really comment more on this.
The quantum papers on "factorization as optimization" are borderline scams though. I wouldn't put those papers in the same sentence as Shor.
To me, the interesting, fulfilling bits of building the next Google/ChatGPT/AI/climate change lie in the theory. Arguably with the exception of Kubernetes, this theory does not come from software engineering. As much as I enjoy software engineering, it's a trade. It's a tool to get the job done. And recently, I realized I like building things just as much as I like "the theory".
To me, that was a bitter pill to swallow. I'm not an ML engineer, but I suspect this is also the reason why you can find so many posts about ML engineers trying to pivot to ML scientist roles.
Although unique ptr is zero cost after make_unique(), I avoid polluting my heap unnecessarily. I've never benchmarked this though (keeping various objects on stack vs heap as unique ptr and how that impacts memory accesses)
I'm quite junior. Appreciate anyone pointing out if anything I said doesn't make sense.
I agree you have a valid point though. I'd be interested to know the committee's reasoning.