jlebar

24 ms·

211.
by jlebar 7y ago
> There is a bit of a market for compiler/ML cross overs ATM (at MS, Facebook, and Google at least). But you need at least a PhD or 10+ years of experience. Until a few months ago I lead part of the XLA team at Google, working on ML
212.
by jlebar 7y ago
The article discusses at length a systemic cover-up of Epstein's involvement with the Media Lab. Multiple high-level people knew about it, knew that MIT had deemed that donations from him were not welcome, and tried to keep his involv
213.
by jlebar 7y ago
> What does "tainted" mean? What does "seriously" mean? Did you read TFA? I think it's quite clear what "tainted" and "seriously" mean.
214.
by jlebar 7y ago
Any content that can be sped up [by 8x] without loss is not worth listening to in the first place. "That's just, like, your opinion, man." Obviously the people who are getting something out of it, otherwise they wouldn&#x
215.
by jlebar 7y ago
Came here thinking this was about the demise of caviar the app.
216.
by jlebar 7y ago
> Would you think that was reasonable The thing I object to is judging them at all. They're not my coworker, someone I have to rely on and therefore might be annoyed if they're unproductive or whatever. They're a random p
217.
by jlebar 7y ago
> Was "asking someone who knows how git works" on your list of things to try? If I can make an analogy. There's little love lost on HN for C++. Anytime people mention C++ (possibly even here in this thread) people come ou
218.
by jlebar 7y ago
I hadn't heard of this! It definitely helps with startup time. $ time hg help > /dev/null real 0m0.232s user 0m0.196s sys 0m0.033s $ time chg help > /dev/null real 0m0.081s use
219.
by jlebar 7y ago
I'll see what I can do. Because it's Google code, it may not be feasible. But we have a whole team of hg developers ourselves, maybe I should bug them before putting it on you. :)
220.
by jlebar 7y ago
I don't believe hg is any better in this respect.
221.
by jlebar 7y ago
> It is superior to Git in every way except performance and adoption. Performance is such an important advantage , though. I type git diff hundreds of times a day, and when I switched from hg to git at Mozilla, this made such a big diff
222.
by jlebar 7y ago
> Git will let you blow away a month's work if you mistype a command. It's sad to hear someone say that, presumably you've been bitten by it. What you describe is not true. While it's easy to blow away work you haven
223.
by jlebar 7y ago
I suspect the whole post was a bit tongue-in-cheek, not just the line about "There's dozens of us. Dozens!"
224.
by jlebar 7y ago
Have a look at the numbers e.g. on drowdown.org to put the 8% in perspective.
225.
by jlebar 7y ago
> My inside-FAAMNG experience has been that most employees seem to be not interested in the politics ...or perhaps they're not interested in talking about politics with you.
226.
by jlebar 7y ago
String search algorithms can cleverly skip forward when they don't find a match. They can skip forward more for longer "needle" strings. That said grep is pretty fast period, so probably doesn't make a huge difference i
227.
by jlebar 7y ago
My project, XLA, will get you quite close to the nominal 900GB/s. :)
228.
by jlebar 7y ago
Bfloat is specifically designed for ML. It is the native type in Google's TPUs. It is quite good at ML; most models that work with fp32 work with bfloat with no adjustments; that's in contrast to IEEE fp16. You're right tha
229.
by jlebar 7y ago
A V100 GPU gets 900GB/s of memory bandwidth. I am less of a CPU expert but afaict you'll be lucky to get much more than 10% of that out of a CPU. This is going to make a huge difference that Intel can't make up with bigger
230.
by jlebar 7y ago
Cutlass doesn't match the performance of cublass, even on nvidia's benchmarks. Hand assembly is still alive and well!
231.
by jlebar 7y ago
Someone needs to write this book / make this movie.
232.
by jlebar 7y ago
A practical example is the Alive theorem prover, which found many bugs in LLVM's instcombine pass (i.e. peephole optimizations) and which we now use to prove the correctness of new transformations. https://rise4fun.com/
233.
by jlebar 7y ago
V100s are often memory bound when using tensor cores, yes. But I guess my point is broader than that. There is a "right shape" for hardware that wants to excel at a particular workload, depending on the arithmetic intensity, deg
234.
by jlebar 7y ago
> Can you imagine how much faster it would get if they released a chip with say 10k tensor cores? I can, actually. :) Adding 10k tensor cores to a GPU would not make it run much faster, and would be prohibitive in terms of die space.
235.
by jlebar 7y ago
(I work at Google on compilers for ml, including compilers for Nvidia gpus.) Devices like the v100 and t4 are ml-specific chips. You can do graphics stuff on them, but that doesn't mean that Nvidia if leaving a ton of ml performance
236.
by jlebar 7y ago
profit sharing is a terrible financial decision [...] it's basically a bet against yourself. It's a hedge, which is different than a true bet against yourself (i.e. a short). Semantics aside, surely hedging against yourself is
237.
by jlebar 8y ago
Not necessarily -- perhaps you're in debt day-to-day. This is the reality for many Americans.
238.
by jlebar 8y ago
In general, if I need an rvalue and it's legal to convert the lvalue I have into an rvalue, the compiler should do it automatically. Having to make it explicit should be the exception, not the rule. The compiler does many other kinds
239.
by jlebar 8y ago
Coal is literally 1000x more dangerous than nuclear power. https://en.m.wikipedia.org/wiki/Energy_accidents#Fatalities Humanity's inability to understand basic math will be our downfall -- perhaps already has sea
240.
by jlebar 8y ago
What about this situation makes you think that white and Asian men are privileged as compared to women? What fraction of companies would you estimate have policies similar to yours?

More ›