Students have gone under academic investigation because their professor asked GPT "was this paper written by AI?" and pasted in their submitted assignment.
2,562 karma · joined January 11, 2017
I've seen life insurance referral fees at over $1000 per successful application.
Auto Insurance, when the market is soft (which it isn't right now), can sell at well over $100 per CLICK for premium consumers.
And because beliefs tend to be on a bit of a continuum, what starts with A believes Z <= 4 and B believes Z >=5 evolves to A believes Z <=1 and B believes Z >=9, and both believe anyone who believes 3<=Z<=7 is not part of their group.
As Christianity became dominant, it began to schism internally with much the same pattern, fractally defining parts of itself as things like "not Arian" or "not Catholic." Judaism, embedded within Christian society, still primary organizes around being "not Gentile" which ranges from simply Christian to full blown anti-semitism and pogroms.
Going back to the alternatives from Arc for browsing and Hey for email management is always jarring. They aren't massive UX changes, but they sure are well thought out and impactful.
Are the roads and text very different colors to you, so that you can easily read the text?
It would be easier to trace the shape of the roads, but that's not how I personally use a map. I'm much more focused on the name of a neighborhood or the name of a road, and its relative position.
I found out several years ago that I have some form of color blindness - amusingly enough during a design review for a map we were designing. I had the exact same problem, with everything bleeding together in an indistinguishable mass, and everyone looking at me like I had 3 heads when I asked how we could possibly ship something so unusable with no contrast.
I suspect there is something similar going on here. To me, the foreground text and background roads are nearly identical in hue and saturation. Is it high contrast to other people? The map I struggled with years ago was using two colors with similar saturation but different hues (a purple and a green/yellow), so I had to push to make sure that "high contrast" meant a difference in both hue AND saturation.
Shipping a couple map-related products has given me a real appreciation for how differently users can experience visual layouts. For one thing, technical people generally like maps a LOT more than the general population does, to the point that we take it for granted that it's the ideal way of presenting anything with a position component.
Quite a lot of people get overwhelmed by maps - they basically shut down and can't figure out the UI at all. Looking at the new version gave me that panicked feeling that I hear those users talk about.
> 1. The effectiveness of clever on other people is highly contingent on outside factors, over which you have no control and of which you may not have any knowledge; i.e., just because you intended to be clever doesn’t mean you will be perceived as clever, for all sorts of reasons.
> 2. The failure mode of clever is “asshole.”
[1] https://whatever.scalzi.com/2010/06/16/the-failure-state-of-...
Honestly, put all the hours spent on CV design towards trying to get coffee with useful people.
* The first versions of the CV are so, so bad. Almost parody-level fluff (the rainbow infinity sign is peak CV cringe) hiding what should be considered some pretty impressive outcome metrics. It would be significantly better if the entire CV was "CEO who scaled startup from $XM Valuation to $250M Valuation" in 72 point font.
* Applying through the front door doesn't work. Only networking is going to get you a job at this level of seniority (and lately almost any level of seniority). Tuning your CV is only helpful for organizing your thoughts for your elevator pitch.
* Being without a job near the top of the pyramid is very bad, and you need a much better story to explain the gap. No one is handing you the keys to their company if you seem like a risk. "I'm a Ukrainian refugee" is a great story and should probably be in there to explain the lack of current employment. Your journey of self discovery is a great story to tell after you land your next role, not while you're looking.
* Probably most importantly, we've been living in a 15-year tech bubble driven by low interest rates and quantitative easing. Valuations have been driven by the insane amount of money flying around, and those valuations have required zero-marginal-cost and infinite TAM to justify the numbers. SaaS, tech, AI/ML are pretty much the only stories that can claim to meet the "zero cost, infinite TAM" investment thesis. Zero interest rate meant you got a super long runway to keep the fantasy going. We have a glut of people who can launch and scale tech companies and there aren't as many opportunities for that as there were three years ago.
I do agree that the Guess label is a little off. That's a bit like saying a quarterback is only "guessing" that their receiver will break off their option route where they are expecting. It's really only Guess culture to an outsider who doesn't know the expectations (as you see when a new WR keeps getting the read wrong, leading to turnovers).
And as others have said, everything is on a continuum. There are very few "ask" cultures where you can just ask someone if you can sleep with their wife and expect no negative repercussions at all. And I doubt that you can get a "no" from someone 25 requests in a row and have neither party question the relationship a little bit.
And there's some unspoken aspects to every request; if you ask someone if you can grab some food from the fridge and they say yes, even in an ask culture they probably have some assumption of how much food you are reasonably going to take. If they come back to an empty fridge, you won't assuage their anger by saying "well I asked and you said yes."
In a new situation, I try to interpret requests like an Asker and make requests like a Guesser (without being offended if I get a no), until there's some shared understanding. That's taken a lot of work, since I'm naturally a Guesser through-and-through.
My wife had an emergency department refuse to treat her because they disagreed with an unrelated diagnosis that was in her medical record. Until she admitted that she was "making it up" and didn't have it, so they could "correct" the record, they would not treat her. She ended up having to leave and go to another hospital.
False Positive Identification Rate grows as the enrolled database gets larger, but True Positive Identification Rate does not (it actually goes down a tiny bit, as some true matches get lost due to lower-accuracy initial search passes). Both FPIR and TPIR assume that the entire universe of possible people is enrolled - i.e. that you are in a "closed set." So what they're really telling you is what the error and success rates are given how many people are in the universe that you are trying to distinguish from each other.
But FPIR is a closed-set metric, that tells you how many false positives you will generate per search. If not everyone is enrolled, you have an open-set problem. And the rate users actually care about is Precision (True Matches / (True Matches + False Matches)). They want to know if the matches that come out of the system are correct.
So looking at rates, that's simply TPIR / (TPIR + FPIR), right? And TPIR is largely constant as N grows, right? So FPIR is the effective error rate, right? Pretty much the whole industry agreed, so Aadhaar dynamically changed its threshold during initial enrollment so that FPIR would stay constant and therefore Precision should stay constant.
** Pause here to see if you can spot the issue **
The intuition that should have screamed out at everyone is that comparisons are independent events and it would be really, really weird if the precision of any given returned match depended on how many of these events were independently happening in other computing threads. It shouldn't matter at all how many independent comparisons are happening. Because they're independent.
TPIR doesn't grow because it assumes everyone is enrolled. Which is definitionally not the case during ramp up. If you want to know how many True Matches the system will spit out, you need to multiply TPIR by probability of enrollment! The Precision is defined by how big of a population you are trying to distinguish, not how many of them are enrolled in the database. The database size is basically used to give you a prior that any given comparison is true; the simplest prior is 1/pop_size, which is the same a 1/db_size for a closed set. But not for an open one.
If the database grows by 10x, you get 10x the false matches because FPIR goes up and you get 10x the true matches because they are 10x more likely to be enrolled. In other words, 10x the matches with the same ratio of true to false. Instead, they lowered the threshold early in the lifecycle, making it more error-prone. This system cost more than $1B in a country where that's a big deal, and had a lot of eyes on it. Trust me, the "experts" were involved and a bunch of them had PhDs. Didn't help.
The first is that the face matching algorithms are inherently less accurate on people of color given the sensors and algorithms in use. This has been an issue going back to color film being developed for Caucasian consumers - the range on a lot of sensors simply doesn't provide enough contrasting tones for brown and black people. The face detection, extraction, and matching algorithms are often not developed using datasets that match the demographics they will be used on, and at least 7-8 years ago there wasn't much focus on providing ethnicity-specific error rates. The biometrics community largely knows about this and understands it.
The second issue is that even the flagship programs don't measure Precision and Recall. They operate purely off metrics like False Match Rate and False Positive Identification Rate. Those are closed-set (i.e. "everyone is enrolled"), tech-only metrics that can be measured by NIST, that don't consider system usage or CONOPs. Program management asks for Precision and Recall metrics (not by name, because they don't know what they are), and tech responds with True Match Rate and True Positive Identification Rate. This is an entire worldview where whole-system accuracy has been placed fully on algorithmic error rates. Leadership generally DOES NOT KNOW that error rates are just as dependent on usage - i.e. that if you search for someone who isn't in the database, 100% of the matches will be false. They do not understand which part of the confusion matrix they are being told about.
The third issue is that, while these massive federal programs are run by ethical and mission-oriented professionals who have hill-climbed their way to a pretty accurate system with stable usage, the talent and professionalism isn't there across the state and local level (there are a couple bright spots where at least the tech is well understood). Face matching systems are generally not lights-out like fingerprint or iris systems (fingerprint systems do generally require a manual verification step, but that's driven more by policy and history than the accuracy metrics require); FR requires a second step where a trained face matching expert looks at the proposed matches in the consideration set and uses several techniques to assess if they are correct - it's a notable difference compared to fingerprint that candidate faces are not supposed to be presumed matches when they come out of the automated system, and even after the human expert they are generally supposed to be evidence/leads that need to be corroborated through other means. As far as I know (again, several years out of date now), no one has checked whether the accuracy of the human experts varies by ethnicity of the subject.
Even the FBI's professionalism broke down with the Madrid Train Bombing, resulting in the arrest of a random Brown lawyer in Seattle (again, no one considered that looking for an international political terrorist in a domestic criminal database meant a much lower prior probability than IAFIS generally operates with, before you even start talking about stitching together multiple partial fingerprints, and you shouldn't be considering anything in the person's file while making an objective matching decision). A local police department is going to be much worse in terms of policy and system usage.
And finally, there is the reality that once you are enrolled in a booking database, you are accused of every crime that happens in any linked jurisdiction until the system goes offline. The probability that you will have an encounter with the police is not uniformly distributed across the population, nor is the probability that such an encounter will lead to a booking where your photo is taken and enrolled in a database forever. This is not something the biometrics community really discusses at all, but every enrollment in one of these systems is hundreds of thousands of opportunities for a false match per year, hanging over your head like bad debt.
The end result is that it's not surprising that just under one in six people in America are Black, but all six of these false arrests were Black. It's also not easily fixed in one place by "giving everyone implicit bias training."
Biometrics do freak people out more than tracking cookies do, largely because they represent your actual body (e.g. this article refers to an IR photograph of your iris as an "eyeball scan" because it gives the reader a more visceral reaction - "eyeball" makes you think of something wet and biological, and "scan" sounds invasive and like it involves lasers or something.).
But they can solve some serious technology difficulties in places where other identity infrastructure isn't in place, and often the concerns are a bit philosophical compared to the tangible benefits. As with all privacy-invasive tech, the issue is exposure to rare but nasty failure modes, and biometrics has the difficulty of being non-revokable.
Converting an image to a template requires similar images to result in similar templates. That's usually the opposite of what you want from a hash. There is a concept of "biometric hashing" that would accomplish the one-way claim, but it's theoretically hard and as far as I know none have hit the accuracy levels to make the biometric part feasible for a large scale systems.[1]
The real claim biometrics vendors tend to make if you push them is that you can't reconstruct a photo-realistic image of the face / finger / iris from the template. Which is generally not a true claim either, but at least would require some work with generative tech. And a photorealistic image that generates the same template is likely to look a lot like the actual person.
Source - 10 years of consulting in biometrics for government, academia, and industry, including spoofing and vulnerability assessments. Most people in the industry don't know what they are talking about and just repeat marketing claims.
[1]https://www.cse.unr.edu/~bebis/CS790Q/PaperPresentations/Bio...
That's actually the one that helped me visualize the theorem. If you look at your scalp from above, you can divide all the hairs into "points left" or "points right" and draw a boundary between them of hairs that point neither left or right. Then you can do the same thing with "points up" and "points down." Where the two kinds of boundaries cross, you have a hair that doesn't point up, down, left, or right - it points straight out of your scalp.
This one overestimates uniqueness because it doesn't consider stability (e.g. it uses your current battery charge level as a uniqueness measure, which is obviously not stable minute-to-minute let alone day-to-day).
I think this bit here drives the love-hate views on him. Most mature orgs are overly bureaucratic with small visions and strongly conservative views about what can be accomplished. Some play-act at having big visions for marketing and ego, but internally few line employees believe that they are achievable, and there is rarely a plan that could achieve the vision.
Musk seems to come in - similarly to Trump - and simply disbelieves the expertise of the managerial class, and comes with enough accumulated power to bulldoze them instead of playing political games. The managerial class is the connective tissue of all but the tiniest of orgs, so encountering them is a given.
The pitfall to me is that he extends that to all expertise - engineering, science, marketing. His opening belief seems to be that most professionals are bullshit artists who spend their whole day lying to justify their jobs. And while the professionals are often wrong about what organizations of people can achieve, they are much less often wrong about how engineering needs to work, and they are hardly every wrong about the core science.
So my general view is that he is pathologically convinced that everyone but him is a parasite. And in a world full of managerial parasites, that's a pretty effective strategy.
But he can't turn it off once he breaks through the connective tissue and gets to the actual expertise. It doesn't matter if you are an expert on cave diving, or scalable architecture, or the behaviors of your core customers. I think it says more about our organizations than it does about him that just spamming a strategy of saying "I think you're a parasite with no actual expertise" and only believing in someone in extremely rare cases can make you the richest man in the world.