nah i think youre underselling it tbh 🤔 identity matching is genuinely hard and theyre at least trying the real problem instead of another chatbot wrapper 🙃

yeah same vibe honestly. the demo is pretty but i kept getting weird scores on names with accents. felt like it wasnt really built for anything outside latin script

not buying the hype. matched two of my test records that shouldnt match. thats a dealbreaker for the use case

别急着唱衰啊 这个MIID方向是真有用的 身份匹配这块刚需 confidence调一调就起飞了 别光盯着BUG看 🚀

hard disagree, this thing matched two totally different people as the same id for me 😬 thats not a calibration problem thats a scary problem 🚩

whats the actual throughput on the matching endpoint though. and is the confidence a calibrated probability or just a raw similarity score normalized to 0-1? makes a big difference for how you set thresholds

i ran my own name through it just for fun and it matched me to someone in a totally different country with 0.81 confidence. so either ive got a long lost twin out there or the calibration is doing comedy. either way i was entertained for a solid ten minutes

agreed with the worry but honestly for a v1 its decent. clean data works, edge cases need polish, thats kinda every early subnet

agreed. fine for clean data, falls apart on the messy stuff

一开始我也觉得没啥 后来发现它去重的能力是真的强 之前手动对账漏的它都抓出来了 改观了 就是那个confidence别太当真

i actually agree with the skepticism here and i want to like it. the core idea of doing identity matching as a subnet task is solid and the team clearly knows the domain. but you are right that the confidence calibration is the weak spot. when i fed it transliterated names the scores were all over the place and there was no obvious way to tell a real match from a coincidence. thats not a small thing if anyone wants to use this for compliance work

yeah the confidence numbers are noise. ignore them and just look at the ranking

demo theater is my favorite genre of subnet honestly. pull up a chair

i dont think its fair to call it demo theater though. theater implies its faking the result. its real, its just not calibrated. those are different complaints and lumping them together makes the team look worse than they are

came in ready to dunk on it but ran my own test set last night and the recall was better than i expected. not perfect but not theater either. taking my doubt down a notch

false positives on identity matching is the one error you cant ship with. push back hard on this one

agree. ranking good, scores bad. usable if you know that going in

idk man i think youre being too generous calling it slick even. the demo lagged for me half the time and the matching missed an obvious dupe. not ready

yeah agreed with larica, demo is doing a lot of heavy lifting. it works but the moment you go off the happy path it gets shaky. needs more time in the oven

我给它喂了我前同事三个名字的变体 它都认出来是一个人 看来比HR还懂我同事 哈哈哈

still think its the most honest subnet in this category fwiw. low bar but yeah

the confidence score and i have a lot in common. neither of us knows whats going on and both of us are very sure about it

yeah i gotta push back, the demo being shaky is one thing but the docs are basically nonexistent. i couldnt even figure out how to set a threshold properly. thats a worse problem than the calibration imo

本来觉得就是个画饼的demo 结果自己跑了下 中文名匹配居然还行 收回一半之前说的话 confidence那块确实拉胯

别吹了 我测出来误匹配率不低 这种东西用在合规上出事谁负责

ok i was being a jerk earlier. ran it on a real list and it caught dupes i missed by hand. its fine. confidence still wonky tho

WAS fully team skeptic on this. then i actually READ how they do the matching and ok yeah theres real engineering here. confidence still needs work but im UP on it now 🔥

我现在管它叫薛定谔的置信度 你不看它就是对的 一看就崩 哈哈

shocking, the demo that was built to look good looks good and the rest doesnt. who could have predicted this. truly nobody

started out nodding along with the skeptics but after a week i changed my mind. the edge cases are real but for the bulk of normal records it just works and thats more than most subnets can say

genuine q for the people running it 👀 are you using the raw confidence at all or just the top k ranking? trying to figure out if the score is even meant to be trusted 🤷

同意 干净数据没问题 脏数据就翻车

demo theater is harsh. it does the matching fine for clean inputs. you just found the fuzzy ones