Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[wierd]

Not so.

Human hands are not nearly as steady as you think they are. Tremor analysis can do wonders when coupled with orientation data.


(that is to say, a phone sitting in a cradle, or on a car seat, will have a smoother dataset than one held in human hands. Additionally, the added jitter from being held in the hands will be fairly regular, since it is predominantly from human heartbeat and respiration patterns, and thus can be picked out even with other sources of noise.)

Picking out those patterns is why you need a simple AI.

[Max™]

There is one and only one insurance provider which would be likely to pick this up initially: legal minimum providers for whom every potential payout is another sliver closer to bankruptcy.

Being able to prove that their driver wasn't texting or whatever is no doubt valuable to them, and only afterwards could negative publicity be used to get other providers to consider ideas like this.

"The General does this for me, why don't you, Allstate?"

[Trekkin]

Human hands are not nearly as steady as you think they are. Tremor analysis can do wonders when coupled with orientation data.

I know how tremulous human hands are. Unfortunately, people's phones vary, including in condition, and all the weird postures they can adopt in the car (which adds a suspension and a road's worth of variability, too). There's a lot of noise here, and any attempt to work back to quantify it once it matters has to contend with all the hardware involved having been through a car wreck in the intervening period. Then the system, with all the noise included, has to hold up in court -- and it has to be obvious enough that it will to justify implementing in the first place, no matter what the insurance company thinks bad actors might do.

Amateurs usually skip the parts of engineering that involve quantifying what happens if we're wrong and how likely that is, but I'd argue they're the most important ones.

[Craftsdwarf boi]

Or modify the black box so that it is even more orwellian.

[Max™]

Your phones already track where you are, listen to casual conversation, and in many cases know what you want before you know  you want it.

[Craftsdwarf boi]

Even if one removes the battery, or post-shutdown?

[Reelya]

Having your phone do something which involves updating the public ledger while in motion would do so, and best of all you can get some of that venture capital bullshit by blockchaining it up!

The critical bit would be having your insurance premium lessened by this, mining safe driver coins or some shit.

There are chatrooms you can only access when your battery is <5% and unplugged, this is a possible thing I'm talking about.

This doesn't many any sense, why don't you read about how crypto works?

First up, there's no need for a public ledger here. This scenario would be much more efficient with a database. Public ledgers are for when you don't have a trusted authority that can keep the data. The very fact that you're saying there's an insurance firm involved means it would be better kept as a database by the insurer. Additionally, volatile state is a poor choice for a blockchain.

Secondly, coins aren't just awarded for random activities, they're awarded for processing transactions for other people. The idea that coins are a purely random reward that can be for anything is a common misconception. Coins represent the processing fee. If you gained the coins for running the app, the everyone's phone would have to store the entire blockchain, gigabytes worth of data and growing ever longer the more everyone drove. That would be an n-squared level problem of data storage.

The way it actually works is that a very few percentage of people mine coins on systems which store the zillion-gigabyte blockchain public ledger. You can't do this on a phone, the ledger is just way too large, and taking in all the data to make new blocks would overload your phone's network connection - if 1 million people use the blockchain then every mining node is receiving constant communications from all 1 million users, all the time. The people willing to do that are the ones who get the coins, because they're giving up processing time and storage space to maintain the system. Then, you have the app users. App users do not have a copy of the blockchain whatsoever.

Plus, there's the fact that blockchains can't handle high volumes of data, and they're especially bad at fitting things into real-time. So, if you want accurate timestamps for like a million people, then blockchain is the worst possible option. Just have the insurance firm have a database with real-time reporting. Databases are a fucking million times faster than blockchains for retrieving an individual's records. For me, for example, to find out how many Dogecoin I own on a new computer, I have to scan the ledger for about 1 week just to get up to the current head, and that's not mining or storing the ledger, that's just skimming each block to see if my account has a transaction in it. For a maintained database, you can jump on a browser and ask "what's my account balance / state" and have an answer in seconds. For blockchain, you need to wait at least hours if not days to get the same query answered.

For the "I am driving now" idea the problem is that with a blockchain, to say whether you are driving or not, then the thing needs to scan the entire blockchain from the first block onwards, a series of "started driving, stopped driving, my velocity is X now" events from day 1 of the blockchain, and you have to read through *everyone else's data* at the same time. It's sequential data, there's no parallelism. This is a very poor fit for the needs of a system that stores someone's driving history: say you didn't drive for 1 week, then you still need to read past all the other blocks representing everyone else driving to get to the next time you drove.

[Reelya]

Great little 2 part video, how to design and build a working VGA monitor compatible video card from components

https://www.youtube.com/watch?v=l7rce6IQDWs
https://www.youtube.com/watch?v=uqY3FMuMuRo

[Kagus]

https://gizmodo.com/scientists-have-created-a-prosthetic-arm-that-lets-pati-1836669048/amp

Fuckin' augs...

[feelotraveller]

https://github.com/1995parham/github-do-not-ban-us

Fuckin' dis-augs...

[Reelya]

Similarly, the more parameters you have the algorithm consider, the more likely it is that someone will find some quirk of the logic that looks racist or classist or  sexist or something.

That occurred with, I think Amazon, who fed a lot of employee performance data into an NN and then trained it to read resumes and estimate how viable hires were. Turns out that it was flagging women as worse hires, and people said the algorithm is biased. However, the algorithm was only doing what it was programmed to do: estimate how much performance different people would have. Women are much more likely to take time out of work (family commitments, parental leave) therefore they don't end up with the same performance metrics down the track, on average, And that's a particularly strong signal, stronger than for instance whether or not you went to Harvard.

The AI was in fact performing properly, it just uncovered uncomfortable truths about the structure of the modern workplace and family commitments, which no amount of affirmative action hiring can really address.

The end result will be that they remove any mention of gender from the resumes, then feed them into another AI and get an output, which will solve the problem by sweeping it under the carpet and obfuscating what the algorithm is actually doing. Such an algorithm could probably guess the gender of the applicant pretty accurately (that's what NNs are good at doing, estimating hidden variables) and mark down the applicant based on that. But by removing the explicit mention of gender, you merely obfuscate what the algorithm is doing.

[MorleyDev]

There have been cases where they just forgot to train the AI against non-"cisgendered white people", so sometimes the algorithm is biased Though that's because of the inherent unconscious bias of the people who fed information into it to train it. Likewise in the case above the algorithm isn't so-much biases as highlighting the inherent sexism in society which things like affirmative action are intended to address (e.g pressure businesses into offering equal available maternity leave for both partners, providing nursery areas etc). There's a whole wide conversation about how to go about solving that particular problem though.

https://github.com/1995parham/github-do-not-ban-us

Fuckin' dis-augs...

It's a tricky one for GitHub. They are an American-owned business based in America and doing business with Americans. So by American law, they can't do business with Iranian businesses. Or, as it turns out, even give advanced warning (seriously, they checked). GitHub are allowed to host non-profit open source projects for free though it seems, but not support Iranian businesses.

The best they can really do is find ways to tell the spirit of the law to go fuck itself, whilst following the letter of it. Still allowing them some access (which is harder to do than a blanket IP ban, which would be the lowest effort approach) is probably the closest thing to that they could even find.

This does highlight another way in which current business laws and politicians still really don't know how to deal with this whole "global interconnected network" thing, since it just makes everything worse for everyone on all sides.

(Okay, some of them know how they want to handle it: Break it and make it go away so their imaginary lines and the physical goods that cross them matter again. And the people and their dangerous thoughts and ideas and knowledge about what the rest of the world is doing or what their state is doing with or to the rest of the world can stop crossing their imaginary lines and getting their good little sheep all riled up. *soviet national anthem starts playing* Rise up! Rise up and cast of the shackles of the oppressors! We have nothing to lose but our restraining fields!).

[Reelya]

There have been cases where they just forgot to train the AI against non-"cisgendered white people", so sometimes the algorithm is biased Though that's because of the inherent unconscious bias of the people who fed information into it to train it. Likewise in the case above the algorithm isn't so-much biases as highlighting the inherent sexism in society which things like affirmative action are intended to address (e.g pressure businesses into offering equal available maternity leave for both partners, providing nursery areas etc). There's a whole wide conversation about how to go about solving that particular problem though.

That also doesn't necessarily solve the 'problem'. Parental leave, if optional, will almost certainly be about 95% taken by women (that's what happens even in places like Sweden, for the optional component of leave), and there is in fact a positive correlation between generous parental leave and the gender wage gap: more generous leave = higher wage gap, because more generous baby-making allowances encourages women to have more babies, and have them younger, none of which are positively correlated with career advancement. Offering flexi-time and paid leave isn't going to get women into higher levels: they're competing against male workaholics who don't take flexi-time or leave. It's like saying you want to train 'part-time' to be in the olympics.

Sure, better parental leave options are 100% a good thing, but it's foolish to assume that offering it will improve all metrics for female employment, such as the gender wage gap. That's kind of a sacred myth now: that any change that benefits women will magically make all metrics for all women improve. Choices are always about trade-offs, and offering more choices in one direction will probably make the 'stats' worse in the other direction.

[MorleyDev]

Sorry, I meant intended to help address. Didn't mean to imply it by itself solved the problem. Affirmative action is primarily a counterweight against societies inherent biases.

For example, you'd also need to work to fix the current social pressures that make it so women are the ones who are more likely to be the primary caregiver for the children when such care giving is a task that can be performed perfectly well by any gender. Which means removing the current bias, conscious or not, towards the idea of a man being judged by his ability to provide for his family, since again: It needn't be the job of a male and to leave the idea that it should be woven into the fabric of society only damages all parties involved.

Affirmative action needs to be a part of a huge long term path to address these problems with society, not a solution in itself.

[wierd]

Sounds to me like they need to mandate leave for fathers, if the aim is to correct the wage gap, and the performance divide.  The AI is just crunching raw numbers; dads who keep working are truly being more "Productive" than the moms who stay home.  Dad needs to bond with baby too, whether he knows it or not.

(of  course, that would still have a wage disparity, just not a gendered one. It would cause a "Productivity difference" between people with kids, and those without.)