OpenAI CEO Sam Altman testifies before U.S. Senate: live cnb.cx/3WaNvDi
Of course, understanding is key. For ex. you can't hope to address bias without talking about the underlying sources of statistical regularities. (See the following, inspired by @HadasKotek )
Danny Fox retweeted
"...discovering that LLMs teach us about how humans work would be startling, akin to discovering that a newly designed drone accidentally solves an open problem in avian flight."
Note by @roni_katzir on the limits of the language modeling paradigm, despite recent claims.
A paper by @spiantado claims that “Modern language models refute Chomsky’s approach to language.” I wrote a short reply explaining why they do no such thing. ling.auf.net/lingbuzz/007190
Danny Fox retweeted
New paper with @_Danny_Fox on distributive inferences - the inferences that speakers draw from a sentence like “every student is French or Spanish” that some student is French and some student is Spanish. (1/3)
Bar-Lev & Fox: On fatal competition and the nature of distributive inferences ift.tt/gBFuMTA #implicatures
Danny Fox retweeted
🗞️ New paper on Lingbuzz with Emmanuel Chemla & @roni_katzir:
How much of our linguistic knowledge is innate? We test four neural networks, including GPT-3, on key aspects of syntactic dependencies and find that all four models fail.
ling.auf.net/lingbuzz/006829
1/3 >>
ALT GPT-3 surprisal values for the sentence "Who did the fact that Mary remembered surprise yesterday/*you?"
Lan et al: Large Language Models and the Argument From the Poverty of the Stimulus ift.tt/1O2e3U7 #syntax
Danny Fox retweeted
New paper on Lingbuzz with @roni_katzir. Many languages have simple words for "and" and "or". Some have words (usually complex) for "nor" (= “not or”). None have a word for "not and".
🧵
Bar-Lev & Katzir: Communicative stability and the typology of logical operators ift.tt/BM7Zn0q #typology
Danny Fox retweeted
Can iterated rationality lead to strengthening when grammatical exhaustification doesn't? Read my new paper w/ @_Danny_Fox and @roni_katzir to find out. (Spoiler: No. When exhaustification does not apply, strengthening does not arise.)
Asherov et al: Reference games and the nature of exhaustification ift.tt/3lFySHQ #scalarimplicature #exhaustivity
Danny Fox retweeted
New paper, with Paul Marty and Jacopo Romoli. Quick summary: we run two experiments to compare selectional (Stalnaker-style) views and homogeneity views of conditionals. Selectional views come out on top. semanticsarchive.net/Archive…
These are very hard systems to understand. One can view this paper as an attempt to report on our efforts.
New paper with @_Danny_Fox with our best attempt to evaluate models of iterated probabilistic reasoning and whether they can replace traditional grammatical explanation of semantic/pragmatic phenomena.