mgraczyk

54 ms·

841.
by mgraczyk 4y ago
I'm specifically talking about cases where you don't have the whole program, like SDKs
842.
by mgraczyk 4y ago
Even if virtual call branches are well predicted, they still have high overhead. Yeah GCC and clang will do what you described in some cases, they will insert a check and try to make a non-virtual call when possible. But also as you mention
843.
by mgraczyk 4y ago
You can't do that in general. For example you can't statically devirtualize a call to a virtual function on a non-final class because it could be inherited from in a different dynamically loaded library. This pattern is pretty co
844.
by mgraczyk 4y ago
Yeah we did lots of things like this at Instagram. Can be very brittle and dangerous though to share any caching amongst multiple users. If you work at Facebook you can search for some SEVs related to this lol
845.
by mgraczyk 4y ago
This depends a lot on what you're doing. If you are ranking 1M qps in a recommender system, then training cost will be tiny compared to inference.
846.
by mgraczyk 4y ago
I do a lot of personal development and everything for my job on cloud VMs. I've been doing more development on a cloud VM than locally for around 4 years now. You can spin up whatever hardware you need more easily, backups are easier,
847.
by mgraczyk 4y ago
No, it's that you haven't worked on ranking systems so you probably don't know that it's also easy to check and prove by anybody who "knows what to look for".
848.
by mgraczyk 4y ago
I implemented heuristics like this one at Facebook. It's not bullshit, I even know how this particular one was implemented. In this case there is some non-determinism but the false positive rate is very low and the false negative rate
849.
by mgraczyk 4y ago
Sure I agree with that if we want to define "algorithm" in such a way that "if (url.contains('sex')) ban()" is "censorship by the algorithm" I just don't think that's a useful definition. Yo
850.
by mgraczyk 4y ago
You can do this on Reddit. People do it all the time https://www.reddit.com/r/modhelp/comments/1px939/hiding_post...
851.
by mgraczyk 4y ago
What is the difference between this and a moderator on Reddit manually deleting a post?
852.
by mgraczyk 4y ago
What I'm saying is that this is a heuristic, not an algorithm, and that you should blame Facebook, not "the algorithm"
853.
by mgraczyk 4y ago
"The algorithm" here refers to Facebook's backend. They are using that word carelessly. I know how it works, I worked on the actual "algorithms" at Facebook. This censorship wasn't done by an "algorithm&qu
854.
by mgraczyk 4y ago
This isn't "censorship by algorithm" though. This was a human being intentionally banning a specific thing. There was no algorithm in the loop. They didn't "tweak their algorithm" anymore than an antivirus twea
855.
by mgraczyk 4y ago
It wasn't true though. The lives of enslaved people were obviously, measurably worse after enslavement and transportation. They had shorter lifespans, less freedom, literally enslavement, etc. The lives of people groups in African capt
856.
by mgraczyk 4y ago
A lot of comments here are criticizing Apple for chasing cheap labor. Ignoring for a second the fact that this has a lot more to do with geopolitics and hedging geopolitical risk, I think it's important to think more carefully about th
857.
by mgraczyk 4y ago
Facebook does not optimize directly for time spent, it's too hard. I know because I worked on these algorithms. Time spent is a metric Facebook monitors and works to increase over time, but basically no work is done to increase time sp
858.
by mgraczyk 4y ago
The fact that it's video first was necessary but not sufficient
859.
by mgraczyk 4y ago
TikTok is winning because it is video first. The more immersive the media, the more profitable the ads. Facebook is having trouble pivoting it's products to video. Facebook is betting big that the next more-immersive medium after video
860.
by mgraczyk 4y ago
This has always been my idea for how I would try to extract a large amount of money if I ever managed to hack a big Defi protocol. Distract exchanges and regulators by splitting the stolen proceeds into tons of tiny amounts, send most of it
861.
by mgraczyk 4y ago
Lots of EU companies have the same language in their policies. For example https://us.ovhcloud.com/legal/privacy-policy
862.
by mgraczyk 4y ago
Right, that's why I included my last paragraph. EU companies can also be forced by the US government to spy on European citizens. It happens all the time.
863.
by mgraczyk 4y ago
Right, and the goal of this sort of ruling is to make sure that those EU-based engineers have work and get paid. It has nothing to do with consumer protection or privacy. If it did, EU companies that have servers in the US would be banned f
864.
by mgraczyk 4y ago
Sounds like blatant protectionism to me. If I'm reading the ruling correctly, the relevant legal standard applied here is completely bogus. They find that it is a violation of GDPR because the parent company could access the data, in p
865.
by mgraczyk 4y ago
There's a bit more to the analysis that you have to consider. Two points 1. The amount of money you can mix depends on the scale of the mixing operation 2. It's easy to build something that looks like a mixer but at some poi
866.
by mgraczyk 4y ago
Code that complies with the standard will be rejected by your compiler. The effect would probably be that few people would use your compiler.
867.
by mgraczyk 4y ago
Even without a plan you can get snapshots for about $0.01 each that I believe are less than 15s old
868.
by mgraczyk 4y ago
See you have to insert the word "spying" to make your case, because by itself there's clearly just nothing wrong with personalized ads. Online ads are good because they make tons of business and economic activity possible and
869.
by mgraczyk 4y ago
Sure, and internal combustion engines were critical in allowing the Nazis to overwhelm their European neighbors. But like ads and product recommendations, internal combustion engines are overwhelmingly used for good so we accept that bad ac
870.
by mgraczyk 4y ago
This is a slippery slope argument, and by itself is just a restatement of your belief that ads are bad. I don't think there's any plausible path from building ads and product recommendations to gulags. Maybe if we go another 20 ye

More ›