Mark J. Nelson

OMG, I recognize that diagram! Although I was still an undergrad then, that was the year I joined the Yale AI lab. Funny side note: the next year I got an "office" in the lab, which, as the lowest ranking member, I had to share with the laser typesetting machine, which smelled terrible.

a photograph of the Edmonton skyline with some hazy green and red bands visible

aurora in Edmonton, though too much light pollution to see with the naked eye (this is a 4-second phone camera exposure)

A diagram showing a narrative sequence involving John, Mary, and the events: "wants to buy", "wants to buy", "buys house", "house is sold", "desires fire", "can't set fire", "desires fire", "gets fire going", "house burns down"

Including (in the '80 volume) some cool diagrams assembled out of typewriter art and hand-drawn arrows cdn.aaai.org/AAAI/1980/AA...

Since I'm Zoom-teaching AI today from an AI conf hotel, I gave a tour of AI publication, what's the deal w/ conference proceedings & walked through some proceedings volumes. Kind of interesting to see the progression – AAAI '80 had 3-pg typewriter-typed "papers"! Up to 7-8 pgs and typeset by 1986.

that's why I got casters installed on my baby grand

After a weird ad experience like this, I religiously rip any online video I want to show in class with yt-dlp ahead of time. Works pretty well, but you have to update yt-dlp weekly to stay a step ahead in the arms race.

Brett Karlan, AI empiricism: the only game in town? - PhilArchive I offer an epistemic argument against the dominance of empiricism and empiricist-inspired methods in contemporary machine learning (ML) research. I first establish, as many ML researchers and philosop...

Only just got around to this, which turned out to be interesting in a different way than I expected. It's not a defense of a specific AI nativist research program so much as a defense that there *should be* an AI nativist program comparable to ML empiricism.

philarchive.org/rec/KARAET-4

Super interesting. For training small LLMs, they get results comparable to those trained on an order of magnitude more tokens by replacing normal pre-training (which ingests huge amounts of human-written text) with exclusively synthetic text derived in a structured way from Wikipedia.

Hmm for me, a conference like CHI is just too big to even read the ToC, let alone a substantial number of papers. Venues like that (along with arXiv, NeurIPS, etc.) do have good papers, but I think of them as more like firehoses to occasionally dip into via search than as publications to browse.

Had an interesting discussion with @sgware.bsky.social yesterday: what conference proceedings or journals do you find worth regularly browsing when they come out? Not necessarily read start to finish, but at least read the whole ToC and browse a subset of the papers? For both of us AIIDE is one!

A position on positions The complex evolution of academic process doesn't always lead us to better practice.

Based on a fun conversation on here, I wrote about the arXiv position paper controversy and the weird, unwritten, organic evolution of academic practice.

"No, seriously. Take command line parsing. You want common syntax for options, right? And you want some of them to accept values, right? And those values can be strings, and booleans, and integers, right? And integers can be decimal or hexadecimal, right? And they can be values of user-defined types, right? And they can have help strings, right? And you'd like to generate help screens from them, right? And GUIs with property pages? And read them from configuration files? And check the legality of flags or sets of flags, right?"

In a slightly different way, I think also ties in a bit to this older debate. yosefk.com/blog/redunda...

Or worse doesn't even really do anything. LLMs love to add try/catch blocks for unrecoverable errors in scripts where the only thing to do is terminate the script anyway.

Two people within earshot on my Air Canada flight ordered Clamato as their in-flight beverage. Is this a Canadian thing? Just sip on some refreshing clam juice at 9:30am? Not even in a cocktail or anything afaict.

Got through customs and made the connection! Are you staying at the Chateau Lacombe hotel? Share a taxi if so?

that's right!

I made it to Canada! Had a 30-min delay taxiing due to increased ATC takeoff spacing, but at least no cancellations. It does appear to be snowing in Toronto, so we'll see about the next leg. 😒

Oh no, I'm in an airport myself. I hope the Mike travel curse doesn't have this large a radius.

(P.S. When did you last go? Looks like it's been a decade for me too, since Santa Cruz 2015.)

Imo most of the policy goals there worth pursuing are barely related to AI – e.g. the LLM features aren't the core of the problem with Canvas style corporate walled gardens. But maybe the AI enclosure framing is ok if that's what it takes to get academics moving on edtech enclosure generally.

I think @mike-treanor.bsky.social, Ben Samuel & I discovered ppl using LLMs for "real" social sim after deciding that, in games, we wanted to use them for NLG but keep them out of the social sim. Different context ofc (social scientists don't? have authorial/design goals), but still funny in a way.

I have one of those too, artisanally generated by a Common Lisp script I wrote (github.com/anadrome/cl-...). Gotten a lot more intermittent that it gets indexed by things like Google Scholar though, maybe because it's not on an .edu (my employer phased out ~username academic websites years ago).

I use it for that too, though it's kind of an odd use of preprint servers. When I worked in Europe (in two different countries) we had institutional repositories for archival OA post-prints. But US institutions seem to either not have those, or they don't work or aren't long-term maintained.

This one's our paper:

AIIDE '25 proceedings are up! And the conference starts in two days. 🧳✈️

Salary is negligible, like 4% of the budget. Would have to stop paying Social Security, Medicare, or Medicaid to produce numbers large enough to impact bond issuance I believe.

This tracker doesn't show any noticeable slowdown in total expenditures yet: www.hamiltonproject.org/data/trackin...

Also seems somewhat different from what they do in coding, where it's not that common to want basically "the 'same' code, but more/less of it".

Has anybody written anything interesting about the duality of LLMs being used to both summarize and expand text? Seems like there is some kind of general function where their goal is to adjust the amount of prose, squeezing or stretching it.

Kind of fun results. Grid of adjective x noun; 10 images sampled from sd-turbo and averaged for each grid cell.

Maybe not as many practical uses, but I would love to have some Safe Mercury

The US government should subsidize Open AI rather than OpenAI