> Enabling allocation sampling profiling with a sampling period of 4 MB leads to a maximum time overhead of 25% in our benchmarks, over un-profiled regular execution
25% is not low overhead, but perhaps this example is the worse case and other tunings actually are low overhead. There's no exact definition I don't think, but anything much over 3% starts to feel like a lot of overhead to me.
Reading on:
> Our main technical insight is that the check whether an allocation should be sampled can be made free. This is done by folding it into the bump-pointer allocator check that PyPy’s GC uses to find out if it should start a minor collection. In this way the fast path with and without memory sampling are exactly the same.
That is cool, and means you only pay for the samples you produce = something you could leave enabled confidently, with a low-enough sample rate at least.
> He went to the Museum of Natural History to study sharks, and he had a model pose across a couple of stools for reference of what someone looks like swimming.
That explains why the swimmer, at least, looks a bit fake.
It's also unclear how the first graph, whose y-axis is labelled "req/s", is showing a latency spike. Latency is not visible on that graph, AFAICT, only throughput.
ah, yes, the "memory is no object" way of obtaining a weighted distribution. If you need that sweet sweet O(1) selection time, maybe check out the Alias Method :)
Do you have any examples of defamation lawsuits? I've sold two companies, and have plenty of rich and upper-level exec friends, and I have never heard of a defamation lawsuit amongst tech execs.
There is definitely a stupid game where people don't tell it like it is, but in my experience it's because people don't want to deal with HR at all, and so don't do or say anything that might bring a tut-tut from HR. Because people in HR are beyond annoying...
And HR fears a wrongful-termination lawsuit, not a defamation lawsuit.
This is all in the US, though. The UK has quite different libel laws and so maybe lawsuits are more common there.
> I'm increasingly convinced this is a generational thing
You are taking about gas cars, though. Nobody, of any generation, has "range anxiety" in a gas car. People might be afraid they will hit empty, or not trust their fuel gauge, but that's not "range anxiety" that's "running out of gas is a PITA". Those things are different, one is a commonplace that has always been true and the other is new and EV-specific for good reason.
An EV is much harder to recover from an empty battery than an ICE car is from an empty tank. There is no red jerry can you can fill that will give an EV the 10-30 miles you get from a gallon of gas, and there is no guarantee that a refueling station is close by and conveniently located. And ofc if your EV is empty, it will take hours to refill. Hence range anxiety.
I have an EV, I love it and have done 7 hour road trips with it no problem, but planning the battery charge level is something you have to think about in a way that is qualitatively different than with gas cars.
> Dismissing all LLM assistance because of some purity dance you want to enact is silly
That's not what I was saying. I was expressing surprise at the lack of spell- or grammar-check in a blog post about detecting slop. I think an AI would generate better text than that of the post, and I'm wondering if the errors are purposeful signifiers of "hey this was a human writing. this busted sentence"
> To be clear, I fault no one for augmenting their writing with LLMs. I do it. A lot now. It’s a great breaker of writers block. But I really do judge those who copy/paste directly from an LLM into a human-space text arena. Sure, take sentences – even proto-paragraphs – if they AI came up with something great.
I guess the sloppy writing ("I do it. A lot now" and "if they AI came up with") made me stop reading early, but: is this part of some big reveal? Sloppy grammar as a sign of not-AI? But it's still slop.
25% is not low overhead, but perhaps this example is the worse case and other tunings actually are low overhead. There's no exact definition I don't think, but anything much over 3% starts to feel like a lot of overhead to me.
Reading on:
> Our main technical insight is that the check whether an allocation should be sampled can be made free. This is done by folding it into the bump-pointer allocator check that PyPy’s GC uses to find out if it should start a minor collection. In this way the fast path with and without memory sampling are exactly the same.
That is cool, and means you only pay for the samples you produce = something you could leave enabled confidently, with a low-enough sample rate at least.