huac

22 ms·

211.
by huac 7y ago
Yeah, was the model trained using train/test splits? otherwise, the model likely has been severely overfit. I wonder how this performance would have compared to a simple random forest or MLP model.
212.
by huac 7y ago
yes, true. there has been progress on NLP SOTA, just not from graph NN architectures. I have thought about using graph NN's where I am currently using RNN's, but they seem fairly immature / difficult to train over very large
213.
by huac 7y ago
> Also were there any new real SOTA on any NLP tasks since last summer? I feel like accuracy progress has frozen.. Just in the last week there are two papers which claim SOTA on different tasks. Microsoft released Turing NLG ( https:&#x2
214.
by huac 7y ago
I can absolutely assure you that FB and G's ad targeting algorithms are more complex and significantly more performant than "cluster and sort by click rate per keyword." For one, you rank ads on an eCPM basis, which includes
215.
by huac 7y ago
> To receive rewards from Redneck Rewards personal checking account, participants must: Make 10 Redneck Bank VISA debit card transactions per monthly statement cycle. Please refer to #9 below. Heh...that's quite a bit, considering t
216.
by huac 7y ago
The Overwatch example reminds me of Playing to Win ( http://www.sirlin.net/ptw ), and in particular, "the scrub" section ( http://www.sirlin.net/ptw-book/introducingthe-scrub ). "A scrub is
217.
Rich people can't build social networks(hua.substack.com)
30 points by huac 7y ago|11 comments
218.
by huac 7y ago
I feel data.frame and tibble are mostly compatible (you can use tidyverse tools on dataframes), and nearly all R users use one or the other, while data.table is used by a few finance folks who grumble about how slow tidyverse is.
219.
by huac 7y ago
PyTorch is a project primarily funded/developed by Facebook/FAIR. Pandas is fully an open-source project, without corporate control. Corporate control means tighter development schedules and consistent API's. It also means th
220.
by huac 7y ago
I wrote a post about the so-called "law of advertiser atrophy": https://hua.substack.com/p/everything-i-have-ever-learned-ab... . TLDR: the author correctly identifies an issue, but misdiagnoses its root cause
221.
by huac 7y ago
Here's a mainstream article: https://www.wired.com/story/democrats-fix-crumbling-data-ope... , I have heard that both the previous setup and the new setup are worse than reported. On a different level, Obama drasti
222.
by huac 7y ago
no, because the Google ads would be served in low CPM countries (eg Turkey) whilst the Bing ads would be served as if the user was from the US (high CPM).
223.
by huac 7y ago
the concepts are simple but doing it at scale is hard. there are some good insights on scaling here: * scaling to more engineers/products: IGQL is an interesting way to compose ML pipelines with straightforward syntax * scaling to more
224.
by huac 7y ago
ha, I am very curious if GDPR does apply to the offline snapshot. it will literally be on EU soil!
225.
by huac 7y ago
I wonder if one of the major pharma companies will develop an anti-dementia drug from this finding, and then if she (or her family) will see any of the billions in revenues they'll make.
226.
by huac 7y ago
the US government says it's a US company, so...
227.
by huac 7y ago
steam market didn't exist then. the other comment links to a steamrep thread - steamrep was a reputation tracker for traders doing real world transactions, which was crucial to mitigate counterparty risk without steam market. the thing
228.
by huac 7y ago
this has been the case since at least 2012 (russian mob, i think), it was a very efficient payout mechanism because you can buy keys for $2.50 and resell for $2 ish, for around a 80% recovery rate. in tf2 we had other mechanisms, e.g. bills
229.
by huac 7y ago
Very similar to smelvetica ( https://github.com/tholman/smelvetica ) which did get takedown notices from Monotype.
230.
by huac 7y ago
how much tax should Amazon pay?
231.
by huac 7y ago
SSF is much less accessible by public transit - going from a major Caltrain station to a minor/local station is a big step down.
232.
by huac 7y ago
say you want to try adding two features, which don't you think interact with each other, e.g. a change to the "pick this room button" and a change to the checkout flow. then, you can randomly assign users to two experiments,
233.
by huac 7y ago
the Vickrey auction mechanism itself is incentive compatible - each agent maximizes their utility by bidding their true value. the article wasn't explicit on this front* but it's true! * "All buyers will likely all pay less
234.
by huac 7y ago
the commercial version: https://preset.io/
235.
by huac 7y ago
Imagine being dunked into a tub of water as your indoctrination. And yet - churches (and other religious organizations) have had high member retention for centuries!
236.
by huac 7y ago
that's a very expensive vision (though I agree that it can be meaningful!), and getting there requires a lot of capital. even though it might not be about pizza in the future, it definitely is right now, and a $2B+ valuation is very ri
237.
by huac 7y ago
uber (see stock), cruise (tbf, everyone overvalued the AV space), slack (see stock), zume pizza (come on), brandless ( https://techcrunch.com/2019/06/26/report-softbank-backed-bra... ), compass ( https:/&#
238.
by huac 7y ago
if everyone had that specific knowledge, there would be no specialists. no, different countries develop different specialties. famously, Kobe Bryant went to Germany (a country with national healthcare) to get a knee procedure from a special
239.
by huac 7y ago
don't you owe taxes on the stock appreciation?
240.
by huac 7y ago
My guess is that Stripe has a service internally which predicts "how much of each currency am I going to have, and how much of each currency do I need to pay out at the end of the day?" Then, they balance whatever currencies they

More ›