Twitter | Pretraživanje | |
Gary Marcus
"If right doesn't matter, we're lost. If the truth doesn't matter, we're lost."
14.633
Tweetovi
3.819
Pratim
39.575
Osobe koje vas prate
Tweetovi
Gary Marcus proslijedio/la je tweet
The New York Times 2 h
"You can’t trust this president to do the right thing, not for one minute, not for one election, not for the sake of our country, you just can’t," Rep. Adam Schiff said during his final pitch in the impeachment trial. "He will not change and you know it."
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Gary Marcus 2 h
the word model gets used in a lot of ways; an ngram model is a model of word usage space but NOT a model of entities in the world; same for Meena, though it is in other ways more sophisticated than n-grams. clear thinking on this point is essential for getting to the next level
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 2 h
forget good conversations, which is maybe partly a matter of social cognition. if you don’t have a model the speaker describes, you are basically verbally nodding your head (“awesome!”) without grasping what is going on.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 2 h
and the main problem is that progress in these metrics doesn’t constitute progress towards extracting dynamic models of the world from discourse.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 2 h
the word model gets used in a lot of ways; an ngram model is a model of word usage space but NOT a model of entities in the world; same for Meena, though it is in other ways more sophisticated than n-grams. clear thinking on this point is essential for getting to the next level
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 2 h
no, i really don’t think so, certainly not in the sense that i mean, any more than a n-gram mode with high n has an actually mode of the events it might describe
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 3 h
no; i don’t think that Meena ever has an internal cognitive model of different actors & their properties & how they evolve over time it’s just that usage history is often but not always correlated w what a system with an internal cognitive model would do. hence zero reliability
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 5 h
Odgovor korisniku/ci @vrandezo @karmacondon
indeed, that's the case here; the author(s) gave me permission to post a link but not to deanonymize, per conference rules. i will repost with fuller information once I am allowed.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Ben Dickson 7 h
Does another huge language model prove anything? I spoke to and about Google's new chatbot Meena
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 5 h
Odgovor korisniku/ci @StenderWorld @eripsa i 3 ostali
it actually has no motive at all, but it refracts to some degree the human dialogs it is built upon.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Jonathan Mugan 6 h
Agreed. You can't have a meaningful conversation without a conceptual understanding of the world. Let's focus on building that.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Gary Marcus 7 h
Odgovor korisniku/ci @tdietterich @vdignum
Meena has exactly the same core issue as ELIZA: it doesn't build a model of what it or the interlocutor has said, and it often contradicts what happened a few turns earlier. Topic without understanding in 1965, topic without understanding in 2020. Here's a sample:
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 7 h
Odgovor korisniku/ci @tdietterich @vdignum
Meena has exactly the same core issue as ELIZA: it doesn't build a model of what it or the interlocutor has said, and it often contradicts what happened a few turns earlier. Topic without understanding in 1965, topic without understanding in 2020. Here's a sample:
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Virginia Dignum 17 h
ELIZA could do the same, or better, in a tiny fraction of the cost. Over 50 years ago! Yet an example of how is being oversold. is not the solution to all your problems! Proper engineering is: thinks first about requirements and design options!
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 18 h
Odgovor korisniku/ci @hyonschu
not in the respects i care about, or wrote about, as far as i can tell.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 18 h
you just can’t make this stuff up.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 20 h
thanks. it also quite relevant to the just-announced Meena...
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Nima Dehghani 2. velj
It is amazing that some fathom that GPT-2 just needs more data to reach comprehension and intelligent. well portrays how misguided is this notion.
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus 2. velj
Odgovor korisniku/ci @vladquant @DrEigenbastard
BERT is totally Chinese room in my eyes, except that it's failures show you that the Chinese room experiment presumes something that is unrealistic. (I have a brief mention of this in my forthcoming Next Decade essay.))
Reply Retweet Označi sa "sviđa mi se"
Gary Marcus proslijedio/la je tweet
Piero Molino 31. sij
Odgovor korisniku/ci @w4nderlus7
2) the model has 2.5B parameters, 79% score. Adding a sibgle rule to avoid repetition improved about 7% the performance. What do you think twitterspere, can we say that a rule is worth 200M parameters? :) (3/3)
Reply Retweet Označi sa "sviđa mi se"