Stuart Shieber (@[email protected])

If you'd like me to continue to follow you on Mastodon, make sure to put your Mastodon username in your own profile, so that tools like debirdify, twitodon, and fedifinder can do their work. 4/4

What happens in the longer term is up in the air. Further action by Twitter to stop Mastodon linking and other aggressions will speed up my transition. 3/4

Having a serious allergy to drama, I'm trying out Mastodon as a Twitter alternative. For any followers who would like to follow me at the new location, I've put my Mastodon username in my Twitter display name. 1/4

Since we first distributed our #goodOA guide 10 years ago today, over 70 institutions have enacted simpatico #openaccess policies. We'd love to know (harvard.goodoa@gmail.com) if it's helped you too! bit.ly/goodoa h/t @petersuber

@SplitSeason1981 All records pre-2023 increase in size of bases.

Moderating principles blogs.harvard.edu/pamphlet/2022/…

Which patents get accepted? How do the standards for innovation evolve over time? What kinds of technologies turn over more quickly? In joint work with Mirac Suzgun @lukemelas @skominers and @pmphlt, we introduce a dataset that may help address questions like these [1/n]

I do not think it is possible to say very much with confidence about the sentience of LaMDA — or of any other computer system — on the basis of transcripts. @pmphlt’s “The Turing Test as Interactive Proof” convincingly explains why not. dash.harvard.edu/handle/1/20272… 1/

@srush_nlp @redpony @yoavgo …but they'd probably argue that matching up of words and corresponding letters is not a *syntactic* requirement, much as in their argument against "respectively" constructions in section 4 of <jstor.org/stable/25001071>. Rather, it's a semantic or even pragmatic phenomenon. 2/2

@srush_nlp @redpony @yoavgo I'm assuming @yoavgo is thinking of an argument that the full phrase and its acronym display a cross-serial dependency, as in the CSDDSG example? I'd defer to Pullum and Gazdar (GaP) on this, ... 1/2

why do people always go all the way to cross-serial dependencies in dutch and swiss-german ( CSDDSG ) to show human language is not context free, where abbreviation definition does the trick just as well (and universally in all (?) languages)?

Exactly 11 years ago today I was on this trip: shbr.link/scouring #uffington #whitehorse #preservation #libraries

If the last two years were a movie, in which scientists develop a near-perfect vaccine for a deadly plague, but half the country won't take it, you'd be like: "This is too dumb to watch."

The results are in this new #ACL2021NLP paper: arxiv.org/abs/2106.06087 Hoping this encourages more work using mediation analysis to reveal internal mechanisms in language models. 4/4

Sometime in the last month, the guide to good practices for university #openaccess policies that I maintain with @pmphlt passed the milestone of 300,000 page views. bit.ly/goodoa

@CarlosODonell That’s the most important thing I learned in college. (Unless you count meeting my wife. But we’re still married because I try to make it easy for her to give me an “A”.) Giving credit where it’s due, I learned it from @pmphlt.

I've never asked for retweets before, but please retweet this in an effort to put an end to this madness: tinyurl.com/x3ywt3ca @CNN et. al. Please. It's "MLB," not "the MLB." It stands for "Major League Baseball." No one says "The Major League Baseball is working with. . . ."

@ThomasScialom @srush_nlp Yep. That was Alan Turing's point in his 1950 Mind paper "Computing machinery and intelligence".

@srush_nlp I tend to think that truly solving NLG would be very close to an AGI, isn't it ? Would be curious to know your opinion on that Sasha?

Every time someone says we are close to solving NLG they should be forced to read 5 of these summaries from Shashi Narayan's paper. huggingface.co/datasets/viewe…