Language Log

Martha Gimbel et al., "Evaluating the Impact of AI on the Labor Market: Current State of Affairs", The Budget Lab (Yale) 10/1/2025:

Overall, our metrics indicate that the broader labor market has not experienced a discernible disruption since ChatGPT’s release 33 months ago, undercutting fears that AI automation is currently eroding the demand for cognitive labor across the economy.

While this finding may contradict the most alarming headlines, it is not surprising given past precedents. Historically, widespread technological disruption in workplaces tends to occur over decades, rather than months or years. Computers didn’t become commonplace in offices until nearly a decade after their release to the public, and it took even longer for them to transform office workflows. Even if new AI technologies will go on to impact the labor market as much, or more, dramatically, it is reasonable to expect that widespread effects will take longer than 33 months to materialize.

Would You Work ‘996’? The Hustle Culture Trend Is Taking Hold in Silicon Valley.
The number combination refers to a work schedule — 9 a.m. to 9 p.m., six days a week — that has its origins in China’s hard-charging tech scene.
By Lora Kelley, NYT (Sept. 28, 2025)

The inverse of involution.

Working 9 to 5 is a way to make a living. But in Silicon Valley, amid the competitive artificial intelligence craze, grinding “996” is the way to get ahead. Or at least to signal to those around you that you’re taking work seriously.

"Active Use of Latent Tree-Structured Sentence Representation in Humans and Large Language Models." Liu, Wei et al. Nature Human Behaviour (September 10, 2025).

Abstract

Understanding how sentences are represented in the human brain, as well as in large language models (LLMs), poses a substantial challenge for cognitive science. Here we develop a one-shot learning task to investigate whether humans and LLMs encode tree-structured constituents within sentences. Participants (total N = 372, native Chinese or English speakers, and bilingual in Chinese and English) and LLMs (for example, ChatGPT) were asked to infer which words should be deleted from a sentence. Both groups tend to delete constituents, instead of non-constituent word strings, following rules specific to Chinese and English, respectively. The results cannot be explained by models that rely only on word properties and word positions. Crucially, based on word strings deleted by either humans or LLMs, the underlying constituency tree structure can be successfully reconstructed. Altogether, these results demonstrate that latent tree-structured sentence representations emerge in both humans and LLMs.

GPT-5 is impressively good at some things (see "No X is better than Y", 8/14/2025, or "GPT-5 can parse headlines!", 9/7/2025), but shockingly bad at others. And I'm not talking about "hallucinations", which is a term used for plausible but false facts or references — such mistakes remain a problem, but every answer is not a hallucination. Adding labels to images that it creates, on the other hand, remains reliably and absurdly bad.

At least sometimes…

Philip Taylor sent a link to this Guardian article "West Point cancels ceremony to honor Tom Hanks as ‘outstanding US citizen’", with the comment

It was only on reading the article that I realised that West Point was/were not cancelling the ceremony in order to honour Tom Hanks (as I had originally thought/believed) but were in fact cancelling a ceremony intended to honour Tom Hanks …

I've been meaning to test GPT-5's parsing ability, ever since I discovered its surprising ability to represent semantic scope ambiguities in correct predicate logic (see "No X is better than Y", 8/13/2025, and the details of its analyses).

Or rather, messages from Penn's Office of the Vice Provost for Research, mysteriously tokenized and re-formatted by gmail.

The start of the Fall 2025 OVPR email newsletter, as displayed by MS Outlook, has 14 bullet points referencing hyperlinked subtopics:

But gmail (where I first read the newsletter) shows me the same information as 14 columns of (individually) hyperlinked textual tokens, with a bullet on the first token of each column:

And what are the consequences of the growing population of AI agents?

In "Agentic culture", I observed that today's "AI agents" have the same features that made "Agent Based Models", 50 years ago, a way to model the emergence and evolution of culture. And I expressed surprise that (almost) none of the concerns about AI impact have taken account of this obvious fact.

There was a little push-back in the comments, for example the claim that "There may come a time when AI is autonomous, reflective and has motives, but that is a long, long way off." Which misses the point, given the entirely unintelligent nature of old-fashioned ABM systems.

Antonio Gulli from Google has recently posted Agentic Design Systems, which offers some useful (and detailed) descriptions of the state of the agentic art, along with example code.

Back in the 1940s, Stanislaw Ulam and John von Neumann came up with the idea of "Cellular automata", which started with models of crystal growth and self-replicating systems, and continued over the decades with explorations in many areas, popularized in the 1970s by Conway's Game of Life. One strand of these explorations became known as Agent-based Models, applied to problems in ecology, sociology, and economics. One especially influential result was Robert Axelrod's work in the mid-1980s on the Evolution of Cooperation.  For a broader survey, see De Marchi and Page, "Agent-based models", Annual Review of Political Science, 2014.

In "Chain of thought hallucination?" (8/8/2025), I illustrated some of the weird text representations that GPT-5 creates when its response is an image rather than a text string. I now have its recommendation for avoiding such problems — which sometimes works, so you can try it…

I Used to Know How to Write in Japanese:
Somehow, though, I can still read it
Marco Giancotti, Aether Mug (August 14, 2025)

During the last thirty to forty years, two of the most popular dictionaries for mastering sinographs were those of James Heisig and Rick Harbaugh.  I was dubious about the efficacy of both and wished that my students wouldn't use them, but language learners flocked to these extremely popular dictionaries, thinking that they offered a magic trick for remembering the characters.

The latter relied on fallacious etymological "trees" and was written by an economist, and the former was based on brute memorization enhanced by magician's tricks and was written by a philosopher of religion.  Both placed characters on a pedestal of visuality / iconicity without integrating them with spoken language.

I have already done a mini-review of Harbaugh's Chinese Characters and Culture: A Genealogy and Dictionary (New Haven: Yale Far Eastern Publications, 1998) on pp. 25-26 here:  Reviews XI, Sino-Platonic Papers, 145 (August, 2004).  The remainder of this post will consist of extracts of Giancotti's essay and the view of a distinguished Japanologist-linguist on Heisig's lexicographical methods.

Forty-five or so years ago, my Chinese and émigré friends who knew Chinese language and were familiar with Chinese society and culture used to josh each other about these terms:

fūrén 夫人 ("madam; Mrs.")

wàifū 外夫 ("outside husband", but sounds like "wife")

nèirén 內人 (lit., "inside person", i.e. my "[house]wife")

The first term is an established lexical item, and the second two are jocular or ad hoc, plus there are other regional and local expressions formed in a similar fashion, as well as some japonismes.

All of these terms were formed from the following four morphosyllables:

fū 夫 ("man; male adult; husband")

rén 人 ("man; person; people")

wài 外 ("outside")

nèi 內 ("inside")

The following sentence in this Bloomberg story

I’m of the mindset that no car payment is better than a new car payment – hence why my 2017 Volvo will likely stick around for a few more years – but I’ve been enticed more about the electric vehicles on the market.

…could lead the reader down a garden path of wondering why a new car payment is the best car payment.

In an update to "Morpho-phonologically AI", I wrote

Ironically, since this puzzle was vocalically inspired by the term "AI" , I'm guessing that current AI systems are not very good at solving (or creating) puzzles like this. I'll give it a try later today.

But it seems that I was wrong.