2,337 karma · joined July 7, 2008
The greatest superpower we have as humans is that we have the ability to introspect, and second guess our immediate responses in light of what we've learned. That allows us to jump to conclusions much quicker on shoddy evidence, because we can course correct later as we gather more, and it's super effective as a learning strategy. Learning how to better correct our initial hypotheses helps us perform even better, but it requires us to actually study the things that humans tend to get wrong. Most education, professions, and life experiences do not support that goal, so it's an easy win if you go outside the system to do so.
Do you really believe this in all situations? I don't. In a standoff between 5 cops and 15 criminals I'd rather see the cops kill them all than vice versa, since the cops didn't create the bad situation in the first place. We all weight lives based on their moral value and culpability in the situation.
Maybe that's an extreme example, but the way it happens more often is that a suspect pulls a gun, points it at a single cop, and the cop has to decide whether to kill the suspect or die him/herself. I'd always argue that in such a situation the cop should shoot.
Are you suggesting that shoot to wound is a better policy?
If we want to talk about the serious issue in America, let's talk about the number of unjustified homicides committed by police versus by others. I and most people don't care one shred about situations where a criminal is pointing a gun at another person and gets shot, that's exactly what should happen, and in many or most police shootings that's the way it goes down. Those make up a big part of the number in this article.
FWIW I'm not disagreeing that abuses by police are a serious problem in America, I think they are, and we need to do something about that, and it goes much deeper than just when they kill. Using shoddy numbers to promote awareness is counterproductive, though, because anyone who doesn't already agree with the conclusion will immediately see what's wrong with the argument, and dig their heels in deeper. There are enough real statistics to talk about (BLM has done a great job promoting those, so I won't dig them up here) that we don't need to resort to misleading linkbait ones.
Most distributed systems don't really scale ass you add nodes beyond a factor of 10 to 100. If they did, you wouldn't need to hire super expensive people to handle those systems.
The thing is, it's 100% correct to evaluate new claims in light of your priors. That's absolutely necessary if you want to make good decisions - my expectation that special relativity will not be shown to be inconsistent is so strong that no matter how interesting or new an argument is against it, I will shred it to bits way more viciously than I'd pick apart a simple replication of special relativity's claims.
But.
Some priors don't come from reliable places, and we have to be way more careful about applying them. In general, the more politically or socially relevant an issue is (gender, race, money, morals) the less any of us should trust both a) our own priors, and b) the priors of researchers in the area, who have many ways to unwittingly and without bad intention put fingers on the scale. In those situations we really have to throw our hands up and wait for meta-analyses to tell us what's up.
In this case, it's super worth noting that even if the IAT is a garbage test, that doesn't mean that implicit bias doesn't exist, or that racism is not a problem. It just means the test doesn't measure it, and we shouldn't be funding further research that leans on it as a tool. Which is why I really don't understand why criticisms of the test draw such intense opposition.
For instance, if you look at an idealized quantum violin string, each mode of vibration has a minimum energy (zero point energy) proportional to the frequency of vibration - classically they can all be 0, but not quantum mechanically. When you try to ask, then, what the minimum energy of the string is, you end up with a term that is literally 1+2+3+... and no particular reason to intepret those as complex or anything. But in a lot of ways, if you just barrel through and treat it as if you can do the sum, you get real results - the Casimir effect is an example where a real force can be predicted and measured based on calculations that are zeta regularized.
It's also worth noting that the dimensionalities in various string theories tend to hinge on the exact values of these infinite sums. Bosonic string theory being 26 dimensional comes out of, IIRC, a consistency equation that ends up including -1/12 because of a 1+2+3... sum. If memory serves, that result can be rigorously established in other ways, as well.
Yet it seems to work...
From experience I know that when it comes to computers, at least in San Francisco, even if you know the exact address where 20+ stolen computers are at, the police will not engage. Are most police similarly hesitant to track down stolen cars, or is that just a weird SFPD quirk about ignoring theft?
Am I missing something?
But if we go that far, then almost everything about you is just random chance. Congratulations, we've realized that we're not special, but products of evolution, and some of us are more fit than others in certain environments. What practical difference should that make?
Gamers will hate me for this, but I actually very much prefer some of the simple "Skinner box" puzzle and collection games that are out now to the shitty, twitchy, time-sink "hardcore" ones that came before that everyone loved so much.
Give me Tsum Tsum, Yokai Watch Wibble Wobble, and Puzzle and Dragons any day over Counterstrike, WoW (though I spent plenty of time on both of those), or any of the FPS crapshows on the market right now that are basically just Counterstrike with better graphics. I'll happily spend my $60 over time while enjoying casual bite-sized gameplay on my schedule, rather than the intensity and practice that "real gamer" games require.
I also fail to see WTF games have to do with the tech industry at all, it's a minuscule portion of the sector.
As a guy I feel like it's a "meh" issue, and most of the women I know agree and use the phrase themselves, but if it bugs someone it costs me zero to avoid it.
The middle ground is where thickness of skin comes in, and while it's not on you to buy in to your employer's standard, it is on the employer to set a standard.
In that respect, I don't think Tesla has been very clear here. I have no idea from this article where the actual line is.
Usually from the point of view of the problematic person, everyone else is the asshole. At some point the evaluation has to boil down to a judgement call, rather than an objective measurement. This process is always fraught, for instance when the bosses are the assholes themselves or support them (ahem, Travis), but I don't love the idea of just not considering interpersonal factors when evaluating performance. Bad behavior shouldn't have to rise to the level of an HR complaints before consequences are faced, and good behavior should be rewarded.
Congrats for being on the side of the undereducated, underskilled, and xenophobic.
It's possible to create rules that operate on an already-trained network and push it in this direction without totally destroying what it's learned, by "fuzzing" the original network to generate a bunch of input/output pairs, and then using that dataset to retrain smaller sub-networks. For instance, if you have a 5 layer network that you've trained on a classification task, you can often use that network as a teacher to train a smaller network to do pretty damn well on the same classification task, even in some cases where training the smaller network directly would have been very difficult. There are several reasons that this trick can work, not the least of which is that in a sense it is a way to expand the training set dramatically.
NB: the above approach is probably not how you'd implement this, there are less crude methods to incentivize shallower levels to have more activation than deeper ones that would probably work better
I can easily imagine a phased training strategy that oscillates between a) learning new things by making the deeper layers more malleable and the shallower ones fairly rigid, and b) compressing all the data by opening up the shallow layers to change and replaying input/output into itself. I have no idea if there are any benchmarks around this sort of thing, though, typically benchmarks have fixed goals so ability to retrain for additional tasks is not really measured.