1,983 karma · joined October 29, 2021
[ my public key: https://keybase.io/pxeger; my proof: https://keybase.io/pxeger/sigs/pWFo0gTe2jOWx8j82aJ_hhdt2wU97tpY-l0AzxVfZFw ]
No, it doesn't matter that much what task the candidate is doing in the interview. It matters what the interviewer is looking at. I'm sure plenty of interviewers don't understand this, and I think this is often missed when people debate about Leetcode interviews - including in this article.
> most interview questions have very little to do with day-to-day responsibilities
You're missing what the interview questions are. Yes, one part of the interview is "here is a puzzle, can you solve it?", but many of the other questions should be things like "explain why this doesn't work", "why did you start with this approach?" and "are you sure that is the best name for that function?"
Leetcode interviews are perfectly "applicable", as long as the interviewer is using them as a convenient frame to see how you think, communicate, and write/read/edit code and isn't trying purely to assess your skill at quickly solving leetcode puzzles.
> cannot distinguish a senior programmer from a marketer using chatGPT
This is empirically false, because companies haven't suddenly lost all their hiring signal since ChatGPT came out. But if a company has this problem, they suck at interviewing.
(I admit the style of interviewing I'm describing has its own problems, and in particular doesn't address what I think is the biggest issue: the fact you're partly testing people's ability to (appear to) perform under acute pressure.)
Knife sales are close to zero-sum, but Palantir is innovating with its lethal technology for the sole purpose of being a more effective weapon. If you were stabbed to death by a knife that is somehow designed specifically to be deadlier or harder to defend against, would you call its designer evil?
[numbers made up, I don't know the exact values, but the general point still stands, and you can't change the speed!!]
E.g.:
https://www.theguardian.com/technology/2023/aug/15/digital-g...
https://www.telegraph.co.uk/news/2023/09/19/online-safety-ac...
It's how the (previous) government was able to look like it was doing something about the mental health crisis caused by social media, even though this bill won't fix that at all.
If companies think the paperwork is too onerous for acquihires (convincing regulators the exclusivity inherent with acquisitions isn't creating an illegal monopoly), and they can get a satsifactory result with a non-exclusive HALO (which regulators therefore don't need to check), why shouldn't they do a HALO instead?
Anyway, row polymorphism can technically be used with nominal typing, it's just that it usually makes sense to use structural typing instead.
The key benefit of row polymorphism is a bit of an implementation detail - it lets you get something resembling (a limited form of) subtyping in your language, without needing as complicated a type inference algorithm as fully-general subtyping requires.
Row polymorphism can be (IMO should usually be) made opt-in, so you can avoid problems like the scenario you describe.
I think one reason they can be as good as or better than dedicated silicon is that they can be adjusted on the fly. If a hardware bug is found in your network chip, too bad. If one is found in your software emulation of a network chip, you can update it easily. What if a new network protocol comes along?
Don't forget the design, verification, mask production, and other one-time costs of making a new type of chip are immense ($millions at least).
> Its why I advocate for a separate ML/AI card instead of using GPU's. Sure their is hardware architecture overlap but your sacrificing so much because your AI cards are founded on GPU hardware.
I think you may have the wrong impression of what modern GPUs are like. They may be descended from graphics cards (as in graphics ), but today they are designed fully with the AI market in mind. And they are design to strike an optional balance between fixed functionality for super-efficient calculations that we believe AI will always need, and programmability to allow innovation in algorithms. Anything more fixed would be unviable immediately because AI would have moved on by the time it could hit the market (and anything less fixed would be too slow).
If the link goes to something that should open in another app (e.g. goes to instagram.com when I have the Instagram app installed), unless I satisfy its demands to install Chrome, it takes like 3 extra clicks to open in that other app.
And anyway why is there any reason to believe they really do use the system they say they use?