gridit

Wolfram Alpha debuted in 2009. Just imagine what they could have been if they'd opened up "skill" creation to the public.

Given that devices equivalent to Alexa or Google Home could pretty much be launched on Kickstarter, is the ecosystem large enough to allow an open-source or more public version of the AI assistant? Is it needed?

The work of a scholar I follow has covers this topic a little. Chris Blattman recently co-wrote an op-ed [0] about his data covering sweatshops in Ethiopia. They found that it wasn't always so clear that these dangerous industrial jobs lifted people out of poverty.

In a really neat trial, they went to factory owners and randomized the acceptance for job applicants, tracking the outcomes of people who found employment elsewhere. Chris originally made his name studying the lives of former child soldiers in Uganda. Recommend his mostly-professional twitter feed [1]

[0] https://www.nytimes.com/2017/04/27/opinion/do-sweatshops-lif... [1] https://twitter.com/cblatts

I enjoyed reading this, and was surprised to learn it was written two decades ago in (1998). I wonder what we've learned since?

This seems pretty streamlined and well-documented, although I wonder if they could have delayed the Facebook login. Convincing users to do your data cleaning for you is quite the trick, like the Smithsonian.

I wonder, what more underhanded versions of data collection are possible? Maybe providing the opportunity to clean inline as users search the archive?

https://transcription.si.edu/phyllis-diller-cards

I was curious too and found this list (sorry about the parent site):

https://roboticsandautomationnews.com/2015/07/21/top-8-indus...

1. Fanuc

2. Yaskawa

3. ABB

4. Kawasaki

The first one mostly has openings in Michigan, concerning machine vision and embedded NN -- from what I can tell.

> It reinvents how people interact with artificial intelligence by allowing you to teach it instead of guessing what it'll understand.

Putting out an app like this seems like a great idea, if only to have a platform to start collecting and cleaning data (for your MTurkers, relatives, and bored employees). Speaking of which the privacy policy link goes nowhere...

On a related note, does anyone have details on the penalties for not having a privacy policy on a Google Play app? They seem to be a little more insistent on having one than Apple.

[As a layperson] the most convincing argument the Authors Guild makes is given I think in their SCOTUS petition: that Google's "fair use" is sidestepping a legitimate business opportunity for the rights holders. Books are not just the paper they're printed on, and authors already as a matter of course hold the rights to plays, movies, etc adapted from the text. Particularly when there is no new expression, it seems to me you are just getting away with not licensing the data. This argument is one of the least-covered in the briefs, however. So.

If I were to make a bar bet, based on my limited knowledge, I would say that any bolder attempt to use mass digitized books for a "transformative purpose" like a chatbot or AI would not pass scrutiny (which kinda sucks, because that would be awesome). That's what I mean by overturned -- perhaps the current GB usage is fine because of point 2) above.

Of course, like many Court issues, the best solution (as yohui alludes above) seems to be to have Congress fix things with real law, such as to create compulsory licensing scheme like in music.

I have been going down the rabbit hole of copyright, fair use, and the Google Books Settlement recently. This article is a great summary including a lot of the peripheral issues, but the "2003 law review article" linked in TFA is nigh unreadable to me, compared to the actual legal opinions and briefs[0].

They are a couple of fascinating documents. The Authors Guild seems gobsmacked by the final ruling, and so am I. Perhaps the SCOTUS was correct to turn down hearing the case, if only to let the issue settle a little more, but it really feels like it's likely to be overturned in the near future.

There are some interesting tidbits in the opinions: 1) In the definitive ruling, the judge decides that the harm done to the market for the books is negligible, or overcome by the transformative "purpose" of the the usage ("purpose" is significant because most examples of fair use include some type of new creative "expression"). This is surprising to me. 2) Google Books is ruled fair use in part because the book descriptions (and snippets?) are metadata describing the books, information that should not be controlled by the authors.

[0] http://www.scotusblog.com/case-files/cases/authors-guild-v-g...

Ok, here is one of the important opinions in the Google Books settlement, by Judge Chin in 2013 [0]. He basically says (paraphrasing), "I'm going to assume Google has violated copyright by creating digital copies and serving them. But it's fair use, because the new products are transformative".

For example, re:ngrams

""" Similarly, Google Books is also transformative in the sense that it has transformed book text into data for purposes of substantive research, including data mining and text mining in new areas, thereby opening up new fields of research. Words in books are being used in a way they have not been used before. Google Books has created something new in the use of book text-the frequency of words and trends in their usage provide substantive information. [...]

On the other hand, fair use has been found even where a defendant benefitted commercially from the unlicensed use of copyrighted works

"""

Oh man, this is mind-blowing.

[0] https://copyright-casebook.com/about/recent-cases-edited/aut...

"If Google could find a way to take that corpus, sliced and diced by genre, topic, time period, all the ways you can divide it, and make that available to machine-learning researchers and hobbyists at universities and out in the wild, I’ll bet there’s some really interesting work that could come out of that. Nobody knows what,” Sloan says. He assumes Google is already doing this internally. Jaskiewicz and others at Google would not say."

For books that are scanned, but with no extra licensing, would Google be allowed to do anything with the data? Create a very delocalized n-gram set? Use it as the "test" set (not even cross-validation, where it might influence hyperparams) for a ML algorithm?

Edit: would love to know where google's authorization derives from, with the ngram set. Somewhere in the Judge's orders? A negotiated fee with the Authors Guild?

Here is a more proper link, if anyone wants to avoid going straight to pdf:

https://hal.inria.fr/hal-01501735/

Interesting that they provided an official incarnation of the WaveNet model for the github code repo https://github.com/tensorflow/magenta/tree/master/magenta/mo...