coherentpony

My employer is counting token usage, so explaining my project between tokens isn’t necessarily a bad thing. I am clearly a more productive engineer because of it \end{sarcasm}

Uber’s situation was different, though. The reason Uber were bleeding money is because they purposefully made all their rides cheap to undercut the taxi businesses. People used Uber because it was cheaper than renting a taxi.

Now you can’t really find taxis anywhere, even at airports it’s a lot more difficult than it used to be.

Once the taxi business was disrupted enough, Uber’s pricing skyrocketed and customers had basically no other options for competition on pricing.

OpenAI basically created a new market. There is no AI chatbot incumbent to disrupt and swallow.

If your comment is intended to convey sympathy on these workers, I think you're going to have a difficult time finding folks that align with you.

If your comment is intended to remind folks that these workers can simply resign of their own free will to find meaningful and dignified work at a different employer, I think you're going to have an easy time finding folks that align with you.

The task is to place four black queens and one black bishop on the chessboard so that there is no square not under their attack

In other words, after arranging the five black pieces, it must be impossible to place the white king anywhere without it being in checkmate.

These two sentences mean very different things in the normal rules of chess. And if you replace the word “checkmate” with the word “check” in the second sentence it still doesn’t mean the same thing as the first sentence.

The first sentence implies that all the pieces must be defended.

Edit: Eh, I guess it depends on how you view the word “attack” since all the pieces are the same colour.

You can saturate the entire power budget on pretty much all GPUs just by moving data in and out of HBM. There is no compute needed at all to do this, and bandwidth bound workloads are extremely common in the scientific computing space.

Is it? Nobody else can really build on their work.

I am somewhat cynically waiting for the AI community to rediscover the last half a century of linear algebra and optimisation techniques.

At some point someone will realise that backpropagation and adjoint solves are the same thing.

It doesn’t work per-song. Songs have multiple chords, some even with alterations. If you tune an E so that it is perfectly a major third above C, then that E won’t be a perfect fifth above an A note. The Am chord has the notes A, C and E, so Am has notes that all belong to C major.

Additionally, some songs even change keys, which makes “per-song” not enough of a constraint.

I’m trying to make sense of this question.

GEMMs are dense O(N^3) work operations that have roughly the same access pattern and data reuse properties across all matrices. Of course, I’m simplifying things a lot here; tall-skinny and short-fat patterns are much harder to get performance out of but the spirit of the approach is the same as big square matrices.

Sparse LU solves have a different character. There is nowhere near O(N^3) work. You typically expect something closer to O(N^2) but getting performance out of these operations is notoriously difficult because it depends a lot on the sparsity pattern of the linear system. Making matters worse is that you may commonly have a sparse A that factorises to dense L and/or U matrices.

Ah yes I forgot about that. Thanks for the correction.

Personally I have a bunch of machines dedicated to gaming in my house (https://lanparty.house)

Woah, that is extremely cool. Very nice work, sir.

The circumference of Earth at the equator is about 40,000 km and the speed of light is about 300,000 km/s. The appropriate division results in about 0.13 s.

That seems to track. The vast majority of requests won’t go half way around the Earth, so maybe halving that time at 0.06 seems like a reasonable target.

Ok that helps. Thank you.

I don’t understand what the visualisation screenshot in the README is trying to communicate to me.

It’s a natural observation, but it doesn’t address the floating point problem. I think the author should have said “fast or would accumulate floating point error” instead of “fast and would accumulate floating point error”.

You could compute in the reverse direction, starting from 1/n instead of starting from 1, this would produce a stable floating point sum but this method is slow.

Edit: Of course, for very large n, 1/n becomes unrepresentable in floating point.

What a wonderful comment. It was educational and friendly without holding my hand too much.

Thank you.

You originally called someone a redditor making a cringe joke for highlighting a serious historical problem. It wasn't clear to me that it was a joke at all, but my impression is that it seemed clear to you that it was a joke.

What if that person has also lived in Texas for 30 years? And what if they had a family member that died during that power grid failure in 2021? I personally would find it quite difficult to communicate to them the nuance of a local problem and a state-wide problem when the end result is the same: no power.

In the future, you might consider approaching an interaction online with more balanced judgement.

Edit: Actually, looking back at the original comment, it's not even clear they're talking about the Texas power outage in 2021. All they said was "Hope they have ample backup power." Seems like a reasonable thing to hope for what might be critical infrastructure.

Texas loses power one time for a week and the redditors will never let it go. Wild how this is still a cringe joke so many years later.

Texas had the most number of power outages between 2019 and 2023 [1].

It wasn't one time. And it's not a joke. Infrastructure weatherization is a very real overlooked (and expensive) investment that still has not taken place.

[1] https://www.congress.gov/118/meeting/house/116952/documents/...

Not totally orthogonal, and quite interesting: https://arxiv.org/abs/1410.6723

For example, ChatGPT is even more aligned with Japan than with the US

I have no expertise in this field.

Is it actually even more aligned? Or is it simply aligned with the elements of Japanese culture and/or media that are exported to the West?

Honestly, I didn’t realise it was an advert until you told me it was.

The concept exists now. You can "reverse offload" work to the CPU.

While true, it also produces a topology that makes you want to throw your PC out the window.

Can you elaborate?

You are the person that made the initial claim. The burden of proof is on you, not someone else.

I have tried the first two things in that list. I haven't tried the third. I would need to think about how to do that in a way that is general enough that it can be re-used across candidates with different skill sets. I like the idea, though. Thanks for sharing it.

It just doesn't feel like this is right.

I know the feeling.

The author says this is one of their favourite interview questions. I stop to wonder what the others are.

When I'm interviewing a candidate, I'm trying to assess really a few things: 1) the capability of the person I'm interviewing to have a technical conversation with another human being; 2) how this person thinks when they are presented with a problem they have to solve; and 3) can this person be trusted with important work?

For 1) and 2), coding interviews and the types of artificially constructed and unrealistic scenarios really aren't helpful, in my experience. I care a lot less about the person being able to solve one specific problem I hand them and I care a lot more about the person being able to handle a much more realistic scenario of being hand handed an ill-defined thing and picking it apart. Those conversations are typically much more open-ended; the goal is to hear how the person approaches the problem, what assumptions they make about the problem, and what follow-ups are needed once they realise at least one of their assumptions is wrong.

This is a really hard thing to do. For example, I imagine (but do not know) that when a medical practice hires a doctor for a certain role, there is an expectation that they already know how the human body works. For an ER doctor, you might care more about how well that person can prioritise and triage patients based on their initial symptoms. And you might also care about how that person handles an escalation when a patient presents not too awfully but is in fact seriously ill. For a GP, it's probably more important for a practice to care more about healthcare philosophy and patient care approaches rather than the prioritisation thing I mentioned above. I'm spit-balling here, but the point is these two situations are both hiring doctors. You care less about what the person knows because there is a tacit assumption that they know what they need to know; you're not giving the candidate a trial surgery or differential diagnosis (maybe... again I'm not a doctor so I don't actually know what I'm talking about here).

If I'm hiring a software engineer or performance engineer, I am trying to figure out how you approach a software design problem or a performance problem. I am not trying to figure out if you can design an async queue in a single-threaded client. This problem doesn't even generalise well to a real situation. It would be like asking a doctor to assume that a patient has no allergies.

Item number 3) is "Can this person be trusted with important work?" and this is basically impossible to determine from an interview. It's also impossible to determine from a CV. The only way to find out is to hire them and give them important work. CVs will say that a candidate was responsible for X, Y and Z. They never say what their contribution was, or whether or not that contribution was a group effort or solo. The only way to find out, is to hire. And I've hired candidates that I thought could be trusted and I was wrong. It sucks. You course-correct them and figure out how to get them to a place where they can be trusted.

Hiring is hard. It's a massive risk. Interviews only give you a partial picture. When you get a good hire it's such a blessing and reduces my anxiety. When you hire a candidate you thought would be good and turns out to be an absolute pain to manage it causes me many sleepless nights.

Converting complex HTML pages with navigation, ads, and JavaScript into LLM-friendly plain text is both difficult and imprecise.

Oh my god. It just occurred to me that LLMs may have a better experience “browsing the internet” than humans do.

That is so tragically depressing.

I did an in-person class. I highly recommend it.

Real-time accent feedback is something language learners have never had throughout all of human history, until now.

Do you have a source for this? It doesn't seem plausible to me, but I'm not an expert.

Ah yes. Ok.

Yes, that makes perfect sense. Thanks.