ianhorn

13 ms·

91.
by ianhorn 6y ago
What blows my mind even more than that is the fact that the universe is apparently 93 billion light years in diameter despite being only 14 billion years old. Intuitively, I'd think that the radius of the universe would be smaller than
92.
by ianhorn 6y ago
I think my discomfort with the idea comes from my inability to know where to draw the line. If I think a person should be fired for a 2020 Trump donation in 2020, why would I think that person shouldn't be fired for a 2019 donation or
93.
by ianhorn 6y ago
> The survey found that many Americans think a person’s private political donations should impact their employment. Nearly a quarter (22%) of Americans would support firing a business executive who personally donates to Democratic presid
94.
by ianhorn 6y ago
Nice work! It's fun to play with. How do you make it reasonably performant without a backend that costs much? I've thought it'd be fun to do stuff like this, but I wouldn't want to have server costs that exceed the price
95.
by ianhorn 6y ago
If your job consisted of calculating with sequences of changes of programs, with no copy paste available, you'd probably feel differently. For example, a fairly roundabout derivation of a change of base for logarithms (pretending we fo
96.
by ianhorn 6y ago
I found the idea of (ab)using mathematical notation to talk about mathematical notation to be entertaining and got the sense that he was kinda just having fun for himself. I also got the impression that he was potentially nerd sniped. The r
97.
by ianhorn 6y ago
> But why not just 1 lifetime of different kinds of data? Heck, why not an environment of 3 years of multi-media data that a child would experience? That wouldn't cost insane amounts of money (or probably anything even close to what
98.
by ianhorn 6y ago
> Why then, the continued obsession with building single-media models? I think it's literally just working with available data. With some back of the envelope math, GPT-3's training corpus is thousands of lifetimes of language
99.
by ianhorn 6y ago
The app you linked has some issues with data entry. If you re-enter the minutes field for SF, it will give a straight line distance of about 7800km, instead of the 1500km you listed. As another commenter pointed out the worse case is ratio
100.
by ianhorn 6y ago
One thing I'd add to this is that this kind of thinking makes your coordinate system really matter. Consider a measurement of some uncertainly sized cubes. You could describe them with their edge length or their volume. Learning one te
101.
by ianhorn 6y ago
GPT-3 was trained on half a trillion words (common crawl, webtext, two book corpuses, and wikipedia, IIRC). At about 100 words per minute, that's almost ten thousand years of continuous speech. By my estimate it's probably a few t
102.
by ianhorn 6y ago
Isn't what you said impossible by construction? You're saying ~80 "valuable" mental mutations wouldn't make us smarter, but if 80 of them wouldn't make us smarter, were they really valuable of mutations?
103.
by ianhorn 6y ago
This is a cool exercise, but there's a broken assumption. Or an assumption of no outwear I missed: > Where is there enough support to prevent items swinging around uncomfortably? He answered this with a box basically around the thig
104.
by ianhorn 6y ago
Cool idea, and thanks for the answer! I'll have to look more closely at the paper :)
105.
by ianhorn 6y ago
What do you mean "carry it out?"
106.
by ianhorn 6y ago
I'd be surprised and saddened if it were legal for a company to offer a contract that they couldn't legally commit to, like saying they'd keep the data out of the U.S. but also give it to the U.S. gov if asked. I guess the le
107.
by ianhorn 6y ago
Thanks for the answer! The local vs systemic bias thing is particularly interesting for a system like snorkel. I have a clarifying question. I'm imagining a dumb example like recipes, where "1 tsp salt" is a common format for
108.
by ianhorn 6y ago
I have a question for the snorkel folks if they're lurking here. With noisy label functions for large amounts of data, I could easily see the cases where the label functions fail correlating with classes that are already having a hard
109.
by ianhorn 6y ago
I saw a talk on Snorkel a few years back, so I don't remember perfectly, but it seemed to be an iterative process. It's a tool for you to build and refine simple rules. If you have ingredients, a simple heuristic "<number&
110.
by ianhorn 6y ago
That got me googling and I found this gem: > Today, Munich is Germany’s most expensive city in which to buy property, with asking prices reaching €35,000 per square metre in some exclusive developments, according to Engel & Völkers.
111.
by ianhorn 6y ago
Maybe the simplest law to make it less of a shit deal for society is forbid the forbidding of charging different prices for cash vs card. Or something halfway, like forcing card companies to allow price differences up to x% of their charge.
112.
by ianhorn 6y ago
Don't other countries limit the amount payment providers can charge?
113.
by ianhorn 6y ago
If your currency is going to undergo hyperinflation, aren't equity or physical goods exactly where you want to keep your money?
114.
by ianhorn 6y ago
Month over month volatility is a poor reason to put money there. Long term average growth is a good reason. Over the 10-20 year periods in the last century, what widely available asset classes perform better?
115.
by ianhorn 6y ago
> Professionals get to opt in and get commensurate rewards. More like professionals are forced in, but get commensurate rewards. Employment isn't flexible enough for people to really choose extra work for extra rewards. The "ty
116.
by ianhorn 6y ago
Was "hacking" as a lay person term ever the same as "hacking" as used by makers? If it was a shift, then yeah you're right. If it was a demonization as the public learned the term at all, then I'd say there&#x2
117.
by ianhorn 6y ago
I wouldn't miss it. If they can find a viable non-tracking business model, I'd imagine most would. People used to pay for newspapers so there's pretty good precedent.
118.
by ianhorn 6y ago
Today, a common "good" chain of custody of facts is a broken game of telephone that's hard to unwind anyways. How often on social media do you see something like someone on reddit/fb/wherever posting an article repo
119.
by ianhorn 6y ago
Take active learning versus usual learning. Often with active learning you can learn much faster. That's a kind of "experience." Out of distribution problems where it fails to generalize could be dealt with much more efficien
120.
by ianhorn 6y ago
For more interesting problems than "which country has the highest GDP?", it's about more than just sloppy data. If you want to include any covariates, how do you know which ones to include? You could try to include everything

More ›