Michael Littman

@TobyWalsh @BrownCSDept @BrownUniversity Thanks, Toby! I start this summer.

@BrownCSDept @mlittmancs @BrownUniversity Congrats @mlittmancs

@mlittmancs has been appointed @BrownUniversity's first Associate Provost for Artificial Intelligence, a new leadership position with a campus-wide charge to advance, in a responsible manner, Brown’s engagement with AI across its academic missions: cs.brown.edu/news/2024/12/0…

This month is the one-year anniversary of the publication of my book, "Code to Joy". I'm happy to announce it's also the month of the release of the audiobook version of the book, which I narrate. Enjoy! @mitpress amazon.com/dp/B0DKG4KPWY

@shimon8282 I see. I was thinking of "Honest Abe". Still...

@mlittmancs Wasn't that Geor Gewashington?

The US Secretary of State cannot tell a lie; after all, he's A. Blinken.

@tdietterich I see. I was thinking of "Honest Abe". Still...

@mlittmancs Wasn't that G. Washington?

I got to help shape this document, providing guidance about how AI researchers collaborate globally. It was unveiled at the UN General Assembly yesterday by the Secretary of State.

You'd think that, since I'm working in government, my id card would count as "government id". It does not. Very confusing when I tried to get through security at the Pentagon. (My Rhode Island drivers's license worked.)

@jeffbigham @ShriramKMurthi @mm_jj_nn This gives me an idea for a future message... "what do the different ratings mean?"

@ShriramKMurthi @mlittmancs @mm_jj_nn hey, at least i didn't get Not Recommended for funding! still a chance!

If you're an NSF PI (or aspirant) not in AI, you probably don't know that @mlittmancs sends out a monthly newsletter educating the community about the workings of the NSF. With a healthy dose of Littmania (jokes, puzzles, etc.). 💯 recommended! littmania.com/courses/my-nsf…

@HamedZamani I saw you position yourself to take this photo. :-)

@mlittmancs's perspective on the current AI landscape in academia! "no one better positioned than the #SIGIR community to help lead the way!" (see the slide for context)

@mm_jj_nn @ShriramKMurthi haha, that's funny... I feel like NSF folks (1) use a lot of acronyms (my list is creeping up to 500 that I've seen used in my 2 years), and (2) like to pronounce them (probably because the alternative is just too hard to understand).

@ShriramKMurthi @mlittmancs huh, TIL

@Evan_ec @svlevine Hi Evan!

I wanted to better understand how language models can help with decision-making and planning and worked with a great team to produce this nifty piece of work!

@bai_liping It's interesting, though, right? Because the best computational tool we have for processing language is the transformer, which isn't really structured the way we make decision systems.

@mlittmancs I always assume that the commonality between language and decision making is the understanding of structured system.

Great topic, great speakers!

In my book, "Code to Joy", I include a Marvel-movie-like post credit scene teasing a future in which AI capabilities are available on an Arduino chip. I just learned that life is imitating art! theverge.com/2024/6/4/24170…

@stanfordnlp @NSF Thanks for posting! I few things I'd add: Susan Dumais, George Furnas, Tom Landauer, Scott Deerwester really pioneered these ideas in the late 80s. And one thing Sue and Tom did that I think is worth new attention: Modeling the language acquisition process using human-scale data!

.@mlittmancs opens the @NSF Workshop on New Horizons in Language Science, noting his work on Latent Semantic Indexing, “the LLMs of the early 1990s”. We made this connection in the GloVe paper, but “neural” work often doesn’t. aclanthology.org/D14-1162/ newhorizonsinlanguagescience.github.io

Dear Schrodinger, the cat's alive! nytimes.com/2024/04/29/us/…

@_max_entropy My concern is with the notion of counting hyperparameters/constants. Sagan said: If you wish to make an apple pie from scratch, you must first invent the universe. Where you draw the lines (algorithm/parameters/hyperparameters) matters.

@mlittmancs Okay, let me rephrase it. If it's possible to generate LLM like algorithms using some reward function, what would it be? How many things (constants) we'd need to specify?

Any LLM design experts here? How many hyper-parameters in total do LLMs use on avg? Say we want to include all that contributed to their working form. Include anything that isn't obvious/ couldn't be derived from first principles of learning (assuming there's such a thing).

And "escape codes" is "cape cod" surrounded by "es" on both sides.

You probably know that an escape code is something that tells a computer not to interpret something literally, but to execute the command following it. But did you ever notice that "escape code" has "cape cod" RIGHT INSIDE it??

@_max_entropy Ah, I apologize. I misread "hyper parameters" as "parameters". I'm not sure your actual question is well-defined, though.

@mlittmancs Oh so you mean all the weights of the network. Shouldn't we use the term "parameters" for that?