There’s not much need for turnover, plenty of companies publish quite a bit on how their systems work. You can look for recsys papers from your favorite company if you want. The Netflix recommendation systems workshop is also good and has many industry talks.
The reality is that there’s a lot of (what is essentially) matrix factorization, and not much of anything nefarious or very interesting to most people who are not recommendation system engineers.
Another thing is that these systems are intentionally hard to game even if you know the weights. The systems optimize for things like P(comment|personalized features), so even if you know there’s a high weight in the scoring function for comments, just commenting a lot on your content isn’t sufficient. You need the system to predict that other people will also comment.
It's pretty easy to do almost no coding at larger companies as a more senior engineer and become quite rusty. It sounds like he was interviewing for a position which he's actually expected to ship things and code, so it's pretty reasonable to ask him to code in the interview.
400 GB/s is insane memory bandwidth. I think a m5.24xlarge for instance has something around 250 GB/s (hard to find exact number). Curious if anyone knows more details about how this compares.
SF is actually getting really fast internet now. All the new developments we looked at had gigabit internet at ~$80/month with multiple competing providers. It's getting to older apartments now as well.
At these speeds, the WiFi adapters are more of a bottleneck than the ISP. I can only get close to full speed on ethernet.
Your thinking rate might be fixed (or more likely, decreasing with time), but one thing you can do, which is not mentioned here, is become more efficient.
This means looking for ways to improve the things you do and finding shortcuts so you have to think less. Most commonly, I do this by looking back at how I did a task and writing down things to help me complete it faster next time. Examples:
- write down the common mistakes I've made in the past and use it as a checklist to check before I push code
- create run-books for debugging issues. This is both for specific areas and general process things such as "Make sure you check all the relevant grafana dashes."
- add questions and answers on StackExchange/Quora when I run into problems/questions that take me awhile to resolve.
Obviously high supply of animators is the main reason, but I wonder if part of it is that pirating/illegal streaming sites are extremely common for anime. All of the people I know that watch anime do so without paying for it. It's trivially easy to find anime for free online, so much so that it's easier to find it for free than pay for it.
It's not free, but I think the functionality is a bit better than this. It handles fast passages and can change temperament, show how off you were, etc.
What is going on here? Why have none of the commenters read the article? Perhaps because it's phrased as a question and people didn't realize it's a link?
Anyways, this matches my expectations--people tend to be overly negative and only remember the good part. The mobile web as a whole has gotten faster due to network speeds+cpu improvements.
It is worth noting that pages are doing more after loading now than they used to be though. This won't show up in onload or first meaningful paint, etc. So the first paint is fast, but then if you try to scroll immediately afterwards you'll probably hit some jankyness while the rest of the page loads asynchronously (but only kind of asynchronously since there's a single main thread).
Some other things that could cause the regression are that more people own a budget Android phone now than before. People may not realize how slow these phones are. The single core performance of the top budget phone, the Samsung A50, is comparable to an iPhone 6 which came out in 2015.
This is not actually happening, and I'm sad to see how long this misinformation persists on HackerNews... Not only not technically feasible, it'd also get out in the numerous lawsuits that Facebook is involved in, which get access to internal information like this.
I like One Medical a lot, I hope they find some way to be profitable. Their doctors are far less rushed than any other GP doctor office I've been to, and their app is well done too.
I've been confused how they're able to do this and not charge more than they do (only $200 a year).
The train/test data being imbalanced in the same way does give the model an advantage, but I don't think that making the test set 50% would solve the issue completely either. Doctors have been "trained" on the true distribution, while which is not 50% (I'd guess that the true distribution is actually extremely unbalanced).
The model isn't simply learning to predict no 80% of the time, it is learning the distribution of the data with respect to the input features. For example, let's say that we have a simple model with only 3 binary features. It may learn that when features X_0, X_1 and X_2 are 1, the probability of cancer is 70%. This isn't a simple multiplication of the true probability by the upscaling factor though--it depends on the percent of negative samples with this feature vector and the percent of positive samples with this feature vector.
If we are to change the test set to be 50% positive and keep the same train distribution, the model no longer has the correct information about cancer rates with respect to feature distributions, but neither does the dermatologist. The specificity and sensitivity continue to not be interpretable as predicted specificity and sensitivity in the real world.
There is no issue with reporting specificity/sensitivity if they had used the true distribution of cases. Yes, the curves/AUCs will look better than the precision/recall rates, but they do not mis-represent what the doctors are interested in (what percent of people will be missed, and what percent of healthy people will be subjected to unnecessary procedures).
Anyways, the classifier doesn't actually seem to be that good, there's actually doctors that were better than the classifiers if you check the paper.
I recently restarted classical music education, and have found it makes it much harder to listen to music and concentrate on my software engineering. It used to be okay if the music doesn’t have vocals, now it has to be something simple I’ve heard many times before.
This is really impressive, I didn't expect starcraft to be played this well by a machine learning based AI. I'm excited to read the paper when it comes out!
That said, I'm not sure I agree that it was winning mainly due to better decision making. For context, I've been ranked in the top 0.1% of players and beaten pros in Starcraft 2, and also work as a machine learning engineer.
The stalker micro in particular looked to be above what's physically possible, especially in the game against Mana where they were fighting in many places at once on the map. Human players have attempted the mass stalker strategy against immortals before, but haven't been able to make it work. The decisions in these fights aren't "interesting"--human players know what they're supposed to do, but can't physically make the actions to do it.
While they have similar APM to SC2 pros, it's probably far more efficient and accurate so I don't think that alone is enough. For example, human players have difficulty macroing while they attack because it takes valuable time to switch context, but the AI didn't appear to suffer from that and was extremely aggressive in many games.
I don’t think that’s a viable strategy in today’s world. The kids I knew growing up that were pushed hard by their parents are doing much better (and happier) than those that were not.
Kids don’t reliably make good long term decisions for themselves. It’s more fun to play fortnite than to practice for a math/piano competition. They can’t see the long term impact of not working out and eating poorly.
I got pushed pretty hard as a kid. It wasn’t fun at the time, but now I appreciate it. There’s a widening gap between the rich and poor in America, and it’s a lot nicer to be on the rich side.
I’m pretty sure they’ll limit the APM (or do something more fine grained since not all APM is equivalent). It’s already well known that bots can outmicro humans in certain aspects like splitting and focusing the correct units.
The fact that this dumb “Google/FB listens to you in the background and uses it to serve ads” myth is still around even on HackerNews is insane to me.
Do you all really think that if this contributed a measureable portion of ads that it wouldn’t leak? There’d also have to be a large amount of hardware dedicated to speech transcription—running constant speech to text on every google chrome browser would need so much compute.
Not to mention that it’d be possible to actually show that the company was lying by modifying microphone drivers and showing that they activate in the background. I’m pretty sure you could also just do it the same way they make game hacks—get access to chrome memory and show that there’s sections that match the recorded speech.
Indisputable proof that this happens would be a PR disaster for either company and cause legal action too.
I’m really in dismay at how long this myth has persisted.
The reality is that there’s a lot of (what is essentially) matrix factorization, and not much of anything nefarious or very interesting to most people who are not recommendation system engineers.
Another thing is that these systems are intentionally hard to game even if you know the weights. The systems optimize for things like P(comment|personalized features), so even if you know there’s a high weight in the scoring function for comments, just commenting a lot on your content isn’t sufficient. You need the system to predict that other people will also comment.