Jelle Zuidema 🟥

I agree on the problem, and on the limits of the kind of AI that dominates the headlines. But solar & wind energy thrive regardless, I am hopeful about the use of physics-informed predictive models --also AI-- for discovering new materials and industrial processes for cheap carbon capture.

The left is missing out on AI As a movement, it has largely refused to engage seriously with AI, ceding debate about a threat and opportunity to the right

Just spent two hours talking w/ 30 (likely left-leaning?) doctoral students about the opportunities and perils of AI. Marx was quoted; the phrase “zero-shot” was used; “stochastic parrot” was not. If this complex reality isn’t visible in thinkpieces / social media, we need to make it visible.+

I need one too!

Transparency at the Source: Evaluating and Interpreting Language Models With Access to the True Distribution We present a setup for training, evaluating and interpreting neural language models, that uses artificial, language-like data. The data is generated using a massive probabilistic grammar (based on sta...

Without guarantees for faithfulness, the generation of explanations is often mere theater ("explainwashing"), but demonstrating faithfulness of explanation methods is hard. In our 2023 paper, @jumelet.bsky.social and I made a herculean attempt to develop a framework for this
arxiv.org/abs/2310.14840

With all the talk about "the blackbox problem" and "the right to explanation", I continue to be suprised by how little attention there is to what I have started calling *the faithfulness problem*: the problem of ensuring that generated explanations are faithful to the underlying causal mechanism.

Much work to do, indeed. And, unfortunately, even the model's explanations for the model's behavior cannot be trusted.

E.g., openreview.net/forum?id=4ub...

Zij "hecht aan land, landschap, taal, een historisch gegroeide gemeenschap". Code-taal voor wensen die men liever niet expliciet opschrijft?

I agree that would be desirable. I'm not involved, but found
this iCal link: calendar.google.com/calendar/ica...

Interessant! Is er consensus dat landbouwgiffen hiermee worden vrij gepleit? De opvallende uitzonderingen -- Zeeland, kop Noord-Holland, tip van Friesland -- zijn precies de gebieden met veel brak grondwater, waardoor boeren vast anders besproeien. Ben benieuwd of dat een factor kan zijn.

Bookmarked - looks very interesting! Is there a connection to the "evolution of modularity" literature from +-20 years ago?

Here's NLP1 from the master of AI in Amsterdam, with lots of materials:
cl-illc.github.io/nlp1-2025/sy...

14th century painting of St. Margaret of Antioch attacking a devil woman in a virid dress hitting a demon with a hammer

Just making sure that 2025 is definitely over

"When a single company can implement a rights-respecting, consent-based access regime at scale, it’s worth asking why public institutions have failed to do the same."

Everyone replies to the "AI" part of you message, but the problem is with "place": Bsky is not one place, but what you make if it. In my experience, you need to follow lots of people on Bsky to have enough interesting posts, but also unfollow those who keep posting on topics I don't find enjoyable.

Big news! 🗞️ I defended my PhD thesis "From Insights to Impact: Actionable Interpretability for Neural Machine Translation" @rug.nl @gronlp.bsky.social

I'm grateful to my advisors @arianna-bis.bsky.social @malvinanissim.bsky.social and to everyone who played a role in this journey! 🎉 #PhDone

It's quite amazing how oblivious the people in charge of these services seem to be of the very real concerns people have with being force-fed AI services everywhere.

The fact that LLM-probabilities mix P(m) and P(g), doesn't justify the practice of posthoc correcting these probabilities with information that humans don't have available. I think it simply means that the search for a more adequate model of grammaticality judgments is still open.

I agree, of course, with your warning about P(m) being a confound, and your 3 predictions (which should be uncontroversial). But that doesn't take away the fact that humans do have strong grammaticality judgments for sentences in isolation, which should not be compared with LLM-judgments on pairs.

Nice! You might be interested in our BLIMP-NL paper, which also reports correlations of SLOR + other metrics with human acceptability judgments, for another language: Dutch. In it, we're also critical of the apples-and-oranges comparison that Leivada et al complain about
bsky.app/profile/wzui...

Cf. @glupyan.bsky.social "Understanding whether the ability to make explicit grammatical judgments is causally related to being able to use language would be a fascinating development."

arxiv.org/pdf/2512.12447

works in practice. E.g., does it explain why people judge some sentences with an entire clause missing as more grammatical? (I think this happens in hippo-sentences with center embedding: daß A B schwimmen helfen sah).

I also wonder if it's relevant for LLM's weaknesses in judging grammaticality.//

Interesting thread. If I understand you right, your theory differs from "syntax emerges from pragmatics" accounts (e.g., bidirectional OT) in that you focus on the *logical* possibility of optimal relevance, not whether or not relevance has been achieved. I'm trying to get my head around how that /1

(Not agreeing with everything in that paper, but looking forward to reading the full discussion on those topics)

Check out www.cambridge.org/core/journal...

We agree on that claim being controversial (and with my "all of" I was trying to refer back to your "all of"), but might disagree on your criticism of Wired's lead. Congrats again on the nice Verge essay; it's an important discussion to be had with all the uncritical talk of imminent AGI/SI.

"All of" neuroscience and cultural evolutionary psychology would be a high bar for a popsci essay! But you rely heavily on one (interesting) slice, but the neuroscience and evolution of cognition literature is much broader. In it, the Wired premise we started with is not uncontroversially disproven

bsky.app/profile/wzui...

I.e., it could be that even if language and thought are district in the adult brain, the development of thought much friends on the development of language in the developing brain.

Or (cf. Stan Dehaene), they both depend on brain structure that evolved for language.

The argument overlooks the possibility of ontogenetic, "glossogenetic" (~cultural) and phylogenetic interdependence of language and thought. And these could also be relevant for discussions on how far we'll get with LLMs. e.g., for you argument about taking away language from LLMs.

I now a read it; I don't disagree (too much) with your analysis of LLMs. But I disagree with: "But if it were true that language is essential to thought, then taking away language should likewise take away our ability to think.". Shared neural infrastructure is not the only possible dependence

I'll have a look at your essay! But I'm skeptical that it will convince me that the "premise is simply not true"; at best, it will be "a premise that has been heavily debated for centuries and that still has proponents and opponents among experts".

My own take: archive.illc.uva.nl/LaCo/CLAS/cl...