Nostr notes by Ulrike Hahn

first talk today #cogsci2026 Judith Fan’s Gleitman Prize ...

2026-07-23T14:55:02Z

first talk today #cogsci2026 Judith Fan’s Gleitman Prize lecture - it was a beautiful overview of her work trying to understand the way humans build and use cognitive tools

here a recent, publicly available talk by Judith Fan

https://www.youtube.com/watch?v=AF3XJT9YKpM

https://cognitivesciencesociety.org/gleitman-prize/

we kind of lost momentum as the reception wasn’t particularly ...

2026-06-12T16:42:45Z

In reply to nevent1q…0q2h
_________________________

we kind of lost momentum as the reception wasn’t particularly warm trying to publish this in mainstream journals, but maybe that was just the wrong route! It would be nice to talk about this and maybe get some better ideas about possible paths forward…

Peter, this is the first time I’ve come across your idea! we ...

2026-06-12T06:38:33Z

In reply to nevent1q…ckxs
_________________________

Peter, this is the first time I’ve come across your idea! we had a similar one at the start of the pandemic: crowd sourced argument/evidence maps….

https://arxiv.org/abs/2212.14714

when the “gold standard” is anything but …. ...

2026-06-11T08:20:21Z

when the “gold standard” is anything but ….

https://www.bmj.com/content/393/bmj-2026-555297

one week of using Qwant instead of Google ….seems fine, ...

2026-06-09T14:10:23Z

one week of using Qwant instead of Google ….seems fine, doesn’t feel like I’m missing things, and generally far fewer results to wade through

(once AI feature was turned off)

👆This is a great paper on the vexed topic of what constitutes ...

2026-05-23T12:56:48Z

In reply to nevent1q…qepz
_________________________

👆This is a great paper on the vexed topic of what constitutes ‘understanding’ …

to me, it’s also a really good illustration of how (often polarized and often quite pointless) debate about these systems can be improved by not ignoring decades of work within cognitive science that is already aimed at these questions

just to summarise, then: Sara, you disagree both with me and with ...

2026-05-23T12:04:53Z

In reply to nevent1q…97tc
_________________________

just to summarise, then: Sara, you disagree both with me and with Yogi but for different reasons: Yogi maintains that LLMs can’t (fully) reason (once we expand reasoning beyond logic) precisely *because* they are not living things, but you agree with him that they don’t; by contrast, I see no compelling reason why reasoning requires a living agent, and whether LLMs do or do not reason is largely a definitional matter, but on the common use of the word in both psychology and AI prior to the advent of LLMs, they arguably do

am on board both with the importance of ‘judgement’ and the ...

2026-05-22T20:06:31Z

In reply to nevent1q…q7q9
_________________________

am on board both with the importance of ‘judgement’ and the fact that we undeniably don’t have an algorithmic account of it- I diverge from you though, in as much as I feel like we don’t have *any* real account of it…

reasoning is, at heart, a transition between *information ...

2026-05-22T04:59:09Z

In reply to nevent1q…qgxe
_________________________

reasoning is, at heart, a transition between *information states*: from this information this other information follows ….

it’s not at all obvious why this or that physical substrate would matter deeply to it….

Sara, I don’t think I was saying that there is no gulf between ...

2026-05-22T04:43:02Z

In reply to nevent1q…q5yk
_________________________

Sara, I don’t think I was saying that there is no gulf between humans and LLMs. We can start from the premise that there is, but that still leaves open whether and in what way that gulf is specifically down to the fact that human beings are living things and LLMs are not (as opposed to any of the other numerous ways in which they differ…), no?

and understanding that requires us to have some kind of account of what “being a living thing” contributes to a capacity like ‘reasoning’, last but not least because the majority of ‘living things’ don’t actually reason

and, at the same time, there are other activities like conducting numerical computations (that seem in many ways related) that we have built non-living artefacts for that are much better than humans performancewise

5/ it makes some concerns less plausible, while also pointing out ...

2026-05-21T11:49:42Z

In reply to nevent1q…g4ey
_________________________

5/ it makes some concerns less plausible, while also pointing out that the decoupling of these two functions in current AI systems doesn't preclude the possibility of comparable causal feedback loops toward (adverse) greater capacity in practice

At the same time, locating that function of 'living thing' brings into sharper relief (for me) the distinction to intuitions about 'living thing' and capacity itself- an intuitive sense that if it's not alive it's not *really* thinking or planning or reasoning, say, because it's not 'conscious'

Just heard the first talk at the conference on reasoning and ...

2026-05-21T10:11:46Z

Just heard the first talk at the conference on reasoning and agency in AI at the Institute of Philosophy by Nick O’Shea. 🧵

https://www.sas.ac.uk/news-events/events/philosophy-ai-conference-2026-reasoning-agency-ai

He looked at outcome-directedness and goal-directedness in biological systems and compared them to AI systems. The crucial point was that both biological and current AI systems show can show outcome-directedness and goal-directedness, and both show capacities for improving themselves by finding “better” solutions to those outcomes or achieving those goals, but biological systems are forced, simultaneously, to find *both* new solutions and the resources to effect those solutions *simultaneously*. Current AI systems do not and this has implications for ‘doomsday scenarios’

#philosophy #AI. 1/

really? it’s an empirical question, I guess, to what extent ...

2026-05-12T10:19:39Z

In reply to nevent1q…5p89
_________________________

really?

it’s an empirical question, I guess, to what extent sloppily “recycled references” that authors did not read misrepresented topics as opposed to being broadly indicative of what they were taken to indicate…there must be research on this?

manual vs. automatic is a tricky one in as much as it’s seeming ...

2026-05-12T10:03:58Z

In reply to nevent1q…2kjj
_________________________

manual vs. automatic is a tricky one in as much as it’s seeming increasingly likely to me that we will be moving to a future where we need LLMs to guard against the damage wrought by LLMs….when the more desirable state of affairs would just be to curtail the availability and use of LLMs….

presumably part of the problem is that there is no agreed upon ...

2026-05-12T06:08:10Z

In reply to nevent1q…6fez
_________________________

presumably part of the problem is that there is no agreed upon procedure yet regarding whose job it is to go through all the references… historically, this was a copy editing/proof reading task, but it seems like it should come sooner?

new preprint on the spread of hallucinated citations in the ...

2026-05-12T05:10:40Z

new preprint on the spread of hallucinated citations in the scientific literature:

- origin primarily from small and early-career author teams

- hallucinated references disproportionately assign credit to already prominent and male scholars

- 𝟴𝟱.𝟯% of hallucinations in preprints persist into the published version

- hallucinations appear across the range of journals, including high impact ones

- fake citations are becoming listed in search engines like Google scholar

https://arxiv.org/abs/2605.07723

a revealing experiment in AI based data analysis…… ...

2026-05-10T11:25:34Z

a revealing experiment in AI based data analysis……

https://kucharski.substack.com/p/real-signals-or-artificial-stereotypes

well, I care about the integrity of our information environment, ...

2026-05-10T05:40:16Z

In reply to nevent1q…d0sn
_________________________

well, I care about the integrity of our information environment, both professionally and personally. That means I care if I receive intentionally deceptive content in my time line and I care if there is coordinated inauthentic activity in the fediverse….

trolling is one thing, the injection of co-ordinated, well funded disinformation by hostile actors into our information systems is quite another

Newsmast, I think it is via your boosts that I received this?

2026-05-10T05:26:28Z

In reply to nevent1q…wqjy
_________________________

Newsmast, I think it is via your boosts that I received this?

that’s your prerogative but I do care: it’s a piece of ...

2026-05-10T05:24:05Z

In reply to nevent1q…vn8w
_________________________

that’s your prerogative but I do care: it’s a piece of targeted disinformation and I care about who is making it and why and I particularly care about how it ended up in my timeline

RE: https://mastodon.social/@geoworldpolitical/116547144621089171 ...

2026-05-10T05:16:15Z

RE: https://mastodon.social/@geoworldpolitical/116547144621089171

strange account with even stranger post: the ‘article’ linked ‘reports’ a huge rise in AI systems refusing or circumventing human instructions in safety relevant ways….

but the linked references don’t exist ….

a UK gov webpage, a Bloomberg article, an FT article, an Anthropic report …. 403 errors

what is this? who is this? and why?

quoting
note1p0u…a28r

AI Models Are Disobeying Humans 500% More Than Six Months Ago

AI models are disobeying humans 500% more than six months ago, according to UK data. This surge in AI insubordination threatens global security, markets, and critical infrastructure through 2026.

https://theboard.world/articles/technology/ai-models-disobeying-humans-500-percent-more

#Technology #Tech #AI

I just spent two really thought provoking days at this amazing ...

2026-05-02T10:04:16Z

I just spent two really thought provoking days at this amazing event co-organized by Jens Madsen from the LSE and the Center for Countering Digital Hate, and it’s made me feel a little bit more hopeful again that there is enough will and expertise that we might still succeed in building information environments that reflect reality as opposed to commercial interests or the whims of autocrats - even though that task is being made even harder than it already was by the advent generative AI

https://info.lse.ac.uk/staff/divisions/research-and-innovation/innovation-and-impact/InformationAndDemocracy

Since taking office in January 2024, the current US ...

2026-05-02T04:45:25Z

Since taking office in January 2024, the current US administration has terminated over 100 scientific advisory committees

https://www.nature.com/articles/d41586-026-01301-5

postscript: interesting or not, I ended up using the example in ...

2026-04-29T20:28:25Z

In reply to nevent1q…hg3y
_________________________

postscript: interesting or not, I ended up using the example in my cognition lecture today, so thanks both of you, for the exchange!

but that’s kind of my point: I don’t think we have (or have ...

2026-04-29T20:25:49Z

In reply to nevent1q…waas
_________________________

but that’s kind of my point: I don’t think we *have* (or have ever had) ‘definitions of learning’ by which NNs ‘don’t learn’, precisely because “learn” is a the (crude) description of a behaviour, not something defined in mechanistic terms…

I’ve never understood NNs as attempting to provide a ‘definition of learning’ itself, though they provide a mechanism, which might or might not have relevant similarities to human learning depending on level of description or detail one is interested in….

“adjusting weights via backprop” is a mechanism, not a system level behaviour and my original reply was that the statment “LLMs don’t “think” they just compute probability distributions” confuses behaviours and behaviour generating mechanisms…

to my mind, that’s not a ‘semantic debate’ or argument about meaning but about a deeper conceptual confusion that seems common in those kinds of articles…

whether pointing *that* out is interesting or enlightening is, of course, a different matter ;-)

I’d also take issue with the way “learning” is ...

2026-04-29T09:24:15Z

In reply to nevent1q…gssx
_________________________

I’d also take issue with the way “learning” is conceptualised in the Zador paper. From the perspective of someone who started doing cognitive science not long after the PDP books came out, I would say that “learning” in ANNs meant learning in an intuitive everyday sense that is contrasted with “hard coded” as in, say, expert systems of the time…

ie, it’s a broad behaviorial descriptor as opposed to the *specific way* they learn ….

it’s odd to me to say “learning” in neural networks “means” extracting statistical regularities as opposed to saying “extracting statistical regularities is how ANNs learn”….

and “thinking”, on a mechanistic level, by contrast is what ...

2026-04-29T07:58:26Z

In reply to nevent1q…6cj7
_________________________

and “thinking”, on a mechanistic level, by contrast is what precisely? 😉

I guess my intuition is that unless the reversal is a strict ...

2026-04-27T17:36:23Z

In reply to nevent1q…rq9w
_________________________

I guess my intuition is that unless the reversal is a strict “undo” nothing about the results is a) surprising and b) limited to LLMs nor even c) meaningfully classed as an error…. so it’s really all down to the actual operations themselves, so it’s all in the detail (and if “add Rembrandt lighting” really is an example they used, I see problems)

…for what it’s worth, I also thought “oh this is a clever ...

2026-04-27T17:02:22Z

In reply to nevent1q…z70y
_________________________

…for what it’s worth, I also thought “oh this is a clever operational test” but then I looked at the methods section and now am really not sure

Michael, that I think would be more problematic (and what I ...

2026-04-27T16:59:30Z

In reply to nevent1q…p9wf
_________________________

Michael, that I think would be more problematic (and what I originally assumed) whereas what I now *think* (?) they are doing is Markdown -> docx -> Markdown; Resultant Markdown -> Latex -> Markdown; Resultant Markdown -> troff -> Markdown

(the comment about order independence in the set of reversible operations is still weird though…)

actually, I think I have now understood the source of (my!) error ...

2026-04-27T15:31:09Z

In reply to nevent1q…0lcq
_________________________

actually, I think I have now understood the source of (my!) error - the “chain” consists of strung together immediate forward/backward steps not concatenation of steps and their undoing?

so linear transformations would be fine as one is only ever doing and undoing one operation at a time?

that would still leave the original problem of potential ambiguity in the instructions (add/remove Rembrandt lighting?)

interesting that it would seem intuitive to a computer scientist ...

2026-04-27T15:06:30Z

In reply to nevent1q…4ye6
_________________________

interesting that it would seem intuitive to a computer scientist given that function composition is typically non-commutative…

I suspect 2) is simply impossible- even very basic, reversible, ...

2026-04-27T12:25:26Z

In reply to nevent1q…fmp2
_________________________

I suspect 2) is simply impossible-

even very basic, reversible, transformations (such as scaling, rotation, reflection) aren’t commutative!

(am still thinking this is so obvious, I must have misunderstood something…)

hmmm… feels more like a potential problem with the method of ...

2026-04-27T11:44:10Z

In reply to nevent1q…lx73
_________________________

hmmm… feels more like a potential problem with the method of the paper to me?

the conceptual core is the idea that all operations are reversible and symmetrical so do not give rise to order effects, but I don’t see how that could be the case?

unless the reverse operation is strictly “undo the specific Rembrandt lighting you added” removing Rembrandt lighting need not take you back to the same place, and there can then also legitimately be order effects….

and *if* the reverse operation is “undo the specific Rembrandt lighting you added” those “undos” would necessarily *only* work in reverse sequential order (by definition)

either way I don’t understand how the actual test lives up to the expressed constraints

won’t the way you route, aggregate and store messages translate ...

2026-04-18T17:55:40Z

In reply to nevent1q…sska
_________________________

won’t the way you route, aggregate and store messages translate directly into energy used?

I see virtually no discussion of the relationship between ...

2026-04-18T15:09:49Z

I see virtually no discussion of the relationship between protocol design and carbon footprint…

shouldn’t that be one of the most important considerations in thinking about social media design right now?

#ActivityPub #atproto

in what way do you see Bluesky an attempt to defeat surveillance ...

2026-04-18T13:55:31Z

In reply to nevent1q…ythe
_________________________

in what way do you see Bluesky an attempt to defeat surveillance capitalism? from their 2024 arXiv preprint:

“….as Bluesky grows, there are likely to be multiple professionally-run indexers for various purposes. For example, a company that performs sentiment analysis on social media activity about brands could easily create a whole-network index that provides insights to their clients. Web search engines can incorporate Bluesky activity into their indexes, and archivists such as the Internet Archive can preserve the activity for posterity.” [pg. 5]

https://arxiv.org/pdf/2402.03239

or maybe this is more helpful... I think the examples of rival ...

2026-04-09T18:55:44Z

In reply to nevent1q…xvzn
_________________________

or maybe this is more helpful...
I think the examples of rival means that both support the 'task' provided in the paper are pretty clear:

1. sounding out letters versus whole word recognition for 'reading' (task 'extracting meaning from text')

2. model based inference versus natural deduction in the case of deductive inference

no?

the only `confusion', such as it is, arises when we start to think about 'inappropriate means', because these will typically also negatively impact attainment on the task.

that is *to be expected* because the 'dimensions' or aspects are functionally related/mutually constrain one another.

The point of the framework isn't to provide a surgical tool that allows us to neatly partition all edge cases (why should it? why should there be failure cases that are failures on just one or the other dimension in the first place?). The aim of the framework is to illuminate the very different focus different disciplines have in how they define reasoning and why.

it feels to me like you understand the distinction perfectly and ...

2026-04-09T18:27:53Z

In reply to nevent1q…x9fq
_________________________

it feels to me like you understand the distinction perfectly and the only confusion/disagreement we have is that you seem to want to read the example of inadequate means with respect to reading as being an example *only* of inadequate means....

there are many examples of the distinction throughout the paper, including the worked example from the psychology of reasoning and there is good reason.

the premise of the paper is that what we think of as a 'behaviour' has different aspects to it - there is no cl;aim that that these can always be separated cleanly any more than that there is a clear border between our concepts of 'bowl' and 'cup'

not sure why I need to “provide definitions” of means, task, ...

2026-04-09T17:46:37Z

In reply to nevent1q…t0ug
_________________________

not sure why I need to “provide definitions” of means, task, and attainment instead of illustrative examples?

also, can I draw your attention again to the passage that very explicitly says that (bottom of page 9, top of 10) that these dimensions, while conceptually distinguishable, clearly constrain each other - not all combinations, ie not all points in space, are realisable - it will be extremely hard to find an example such as the way you are wanting to read the ‘wrong means’ passage memorisation example (ie not failing more than ‘wrong means’)

Tiota, I’m just not seeing where you think I disagree. If you ...

2026-04-09T13:08:45Z

In reply to nevent1q…6m5c
_________________________

Tiota, I’m just not seeing where you think I disagree. If you reread the actual text you will see that the very point of the memorization example is to illustrate “not task (reading), because not appropriate means”

just a quick comment: the paper doesn’t give a definition of ...

2026-04-09T01:31:17Z

In reply to nevent1q…2pmu
_________________________

just a quick comment: the paper doesn’t give a definition of reading, let alone an attainment based definition. The whole premise of the paper is that typical cases are typical *along all three dimensions* and that all three matter, but that different fields give different weight to each of these dimensions in ways that make sense, given their constraints and goals.

What you are reading as an “attainment based definition” of reading is simply part of the exposition of what the attainment definition *is*, alongside what the task and means dimension are.

hope that clarifies things!

and this is one of the earliest (?) papers on calibration in LLMs ...

2026-04-08T17:48:24Z

In reply to nevent1q…ae2y
_________________________

and this is one of the earliest (?) papers on calibration in LLMs

at the same time, calibration in humans has been the topic of many decades of study within judgment and decision-making research….

https://arxiv.org/pdf/2207.05221

I haven’t been following this literature of late and it depends ...

2026-04-08T17:40:06Z

In reply to nevent1q…nlyz
_________________________

I haven’t been following this literature of late and it depends on what you want to call a “mode” but there is a literature on the signal distinguishability of hallucinated versus non-hallucinated states …my crude take away, last time I looked, is that there was uncertainty information there if one wanted to leverage it….

https://aclanthology.org/2023.findings-emnlp.68.pdf

if I may… https://zenodo.org/records/18231172

2026-04-08T04:26:06Z

In reply to nevent1q…lkhg
_________________________

if I may… https://zenodo.org/records/18231172

Vibe coding for the win… ...

2026-04-06T14:57:43Z

Vibe coding for the win…

good piece! “the development and application of methods, the ...

2026-04-03T15:45:52Z

good piece!

“the development and application of methods, the training of minds, the creation of people who know how to think about hard problems. If you hand that process to a machine, you haven't accelerated science. You've removed the only part of it that anyone actually needed.”

https://ergosphere.blog/posts/the-machines-are-fine/

well worth the read!

2026-04-02T18:02:35Z

In reply to nevent1q…5r6c
_________________________

well worth the read!

I recommend (for multiple reasons) this piece on coding with AI ...

2026-04-01T19:45:49Z

I recommend (for multiple reasons) this piece on coding with AI (Claude Code):

https://taggart-tech.com/reckoning/

mir aus der Seele gesprochen….

2026-04-01T15:46:21Z

In reply to nevent1q…9z0j
_________________________

mir aus der Seele gesprochen….

congratulations!

2026-03-31T19:57:50Z

In reply to nevent1q…h7g2
_________________________

congratulations!

there’s always two parts to this: 1) your use of the tool, and ...

2026-03-30T15:36:03Z

In reply to nevent1q…5wnc
_________________________

there’s always two parts to this: 1) your use of the tool, and 2) the use of your data by such tool when used by others. The fact that Bluesky users are blocking attie on mass and encouraging others to do so, suggests to me that they dislike 2) as well.

nor does the response here seem very different to Bluesky itself ...

2026-03-30T15:29:13Z

In reply to nevent1q…pwzy
_________________________

nor does the response here seem very different to Bluesky itself - Jay Graeber’s announcement had 1700 replies when I looked and it was just one endless stream of negative responses

but people feel sufficiently strongly about AI as a tool that ...

2026-03-30T15:24:26Z

In reply to nevent1q…0vzh
_________________________

but people feel sufficiently strongly about AI as a tool that they simply don’t want the benefit if it’s by those means. I don’t see any evidence of misunderstanding in that regard

what is it you think people are misunderstanding? (including on ...

2026-03-30T15:18:58Z

In reply to nevent1q…tfnk
_________________________

what is it you think people are misunderstanding? (including on Bluesky itself where the comments on the announcement are overwhelmingly negative…)

“With Attie, anyone will be able to build their own custom feed ...

2026-03-30T13:16:00Z

“With Attie, anyone will be able to build their own custom feed just by typing in commands in natural language, the same as if they’re chatting with any other AI chatbot. To use the app, people will sign in with their Atmosphere login (meaning their login for any app that runs on atproto, which includes Bluesky). Attie will immediately understand what you’ve been talking about, what sort of things you like, and more, because Bluesky and the wider ecosystem are open systems that share data across apps.”

https://techcrunch.com/2026/03/28/bluesky-leans-into-ai-with-attie-an-app-for-building-custom-feeds/

there will hopefully be restricted visibility posts in ATproto ...

2026-03-20T14:11:49Z

In reply to nevent1q…gcep
_________________________

there will hopefully be restricted visibility posts in ATproto eventually. But they weren’t part of the original design

I think the legal aspect matters. And once someone downloads an ...

2026-03-20T14:08:50Z

In reply to nevent1q…unue
_________________________

I think the legal aspect matters. And once someone downloads an ATproto post it’s not going to vanish by the author deleting it either. There is a freely available archive tool that lets you download your entire Bluesky post history with two clicks. And given how ATproto works, that means you can use the same tool to download, archive and search another users entire history of posts as well

there isn’t a permission less firehose, though? it seems both ...

2026-03-20T13:56:24Z

In reply to nevent1q…37fc
_________________________

there isn’t a permission less firehose, though?

it seems both legally different to me and different in design:

https://write.as/ulrikehahn/bridging-to-bluesky-the-open-social-web-consent-and-gdpr

reminder, the Bluesky team explicitly mentioned commercial use of ...

2026-03-20T13:49:45Z

In reply to nevent1q…jmzh
_________________________

reminder, the Bluesky team explicitly mentioned commercial use of their data in their 2024 arXiv preprint:

“…as Bluesky grows, there are likely to be multiple professionally-run indexers for various purposes. For example, a company that performs sentiment analysis on social media activity about brands could easily create a whole-network index that provides insights to their clients. Web search engines can incorporate Bluesky activity into their indexes,…” pg. 5

it’s been part of the model from the start…

I don’t think banning is likely or feasible

2026-03-14T20:55:07Z

In reply to nevent1q…rq35
_________________________

I don’t think banning is likely or feasible

Is that what is happening here? I think these arguments stem from ...

2026-03-11T14:03:51Z

In reply to nevent1q…hzn0
_________________________

Is that what is happening here? I think these arguments stem from deep concern; unfortunately that isn’t enough to make them compelling. I genuinely wish it were as simple as “this seems bad, let’s not do this”, but unfortunately I think it really is not. As @SyliviaFysica said, the very fact that developments are so concerning makes it all the more important that we don’t fool ourselves. In that context, I don’t feel it’s the ethical issues that I’m struggling with, it’s that the introduction of LLMs seems both so transformational and so uncontrollable.

We really need ways to cope with generative AI and control it, ...

2026-03-11T09:00:55Z

In reply to nevent1q…08m3
_________________________

We really need ways to cope with generative AI and control it, but they are going to have to be able cope with its unique features.

OK, let’s destroy the LLMs themselves, the Luddites can be our ...

2026-03-11T08:56:30Z

In reply to nevent1q…c2e2
_________________________

OK, let’s destroy the LLMs themselves, the Luddites can be our inspiration!

never mind that they did not halt the advance of the mechanical loom, the only “thing” one needs about LLM for use is the weights. These too are infinitely copy-able.

‘Technology is not inevitable. We can just not do LLMs. Tobacco ...

2026-03-11T08:50:24Z

‘Technology is not inevitable. We can just not do LLMs.
Tobacco apparently is an illustrative example. Asbestos too.’

Cigarettes are still on sale. Tobacco companies still exist. Despite decades of monumental efforts. And the parties deeply invested in developing AI technology are not just large corporations but the most powerful state actors on the planet.

Asbestos is a distinct physical substance, that we can detect and remove. LLMs produce linguistic output which is not. Text is infinitely replicable and can seamlessly be transformed into another modality, speech. The linguistic output of LLMs is impossible to reliably distinguish from human output by its very nature.

Judith Fan is the recipient of the 2026 Lila R. Gleitman Prize. ...

2026-03-10T15:24:57Z

Judith Fan is the recipient of the 2026 Lila R. Gleitman Prize. Dr. Fan is an Assistant Professor of Psychology (and, by courtesy, Education & Computer Science) at Stanford University. She received her Ph.D. in Psychology at Princeton University in 2016, following an A.B. in Neurobiology and Statistics at Harvard College.

Congratulations!

https://cognitivesciencesociety.org/gleitman-prize/

#cogsci

sadly had same experience with Libreoffice including losing ...

2026-03-10T12:00:06Z

In reply to nevent1q…pd85
_________________________

sadly had same experience with Libreoffice including losing documents…

I’m also inclined to think that online social media has played ...

2026-03-09T12:18:06Z

In reply to nevent1q…8zr0
_________________________

I’m also inclined to think that online social media has played a big part in this and that it was a consequential step into a negative feedback loop when the news started reporting on “what people are talking about on Twitter right now” - not only is virality on social media not epistemologically grounded in the right way, the growth of social media undermined the advertising model of legacy media, leading to shrinking resources for direct reporting, and ever greater journalistic reliance on what is being said, which in turn….

agreed: I think we fundamentally broke the link between “how ...

2026-03-09T12:06:38Z

In reply to nevent1q…662z
_________________________

agreed: I think we fundamentally broke the link between “how impactful is this” and “how much noise does this make” over the last decades. So a formerly reasonable attention-based indicator (how much are people talking about this, responding to this?) has now lost much, if not most, of its diagnostic value….

the same goes for any strategy that rests largely on trying to ...

2026-03-09T10:59:52Z

In reply to nevent1q…xy86
_________________________

the same goes for any strategy that rests largely on trying to dissuade people from using AI by telling them it doesn’t work and it’s evil, because that strategy’s effectiveness presupposes that people’s evaluations of the actual system capabilities, what they personally could or could not do with these systems, as well as what utilities/values are involved in using them will ultimately match your own.

indeed, and I think we should all be able to agree that that ...

2026-03-09T09:42:37Z

In reply to nevent1q…cssv
_________________________

indeed, and I think we should all be able to agree that that horror indicates the limits of “ridicule as praxis” as a strategy for countering the development of AI systems and their impact, however well-intentioned that strategy might have been

I wish I knew…

2026-03-09T08:09:52Z

In reply to nevent1q…qmud
_________________________

I wish I knew…

unfortunately, one of the things current AI systems can do ...

2026-03-09T08:09:32Z

In reply to nevent1q…e3yc
_________________________

unfortunately, one of the things current AI systems can do extremely well is increase the scale of responding. That limits the efficacy of a strategy focussed primarily on dissuading people from using these systems, because comparatively few actors can have very large impact

the negative transformative effects that AI is currently having ...

2026-03-09T07:46:18Z

In reply to nevent1q…davf
_________________________

the negative transformative effects that AI is currently having stem as much from the things these systems can do well as they do from the things that they can’t, and focussing exclusively on the “can’t” has, I think, been counter-productive to anticipating threats, countering them or adapting to them

it’s unfortunate that so many of the voices concerned about the ...

2026-03-09T07:36:38Z

it’s unfortunate that so many of the voices concerned about the adoption of AI have been so singularly focussed on “hype” as the main problem and pushing back against “hype” as a core strategy, because that hasn’t helped us deal with the very real, and in many cases very destructive, transformative effects the technology has actually been having

literature "If we're not prepared to deal with the fraud ...

2026-03-08T14:04:06Z

In reply to nevent1q…x8rw
_________________________

literature

"If we're not prepared to deal with the fraud that's already occurring, then we're certainly not prepared to deal with what generative AI can do to scientific literature”

the only bit I’d take issue with in the article is the comment ...

2026-03-07T08:28:42Z

In reply to nevent1q…7hqv
_________________________

the only bit I’d take issue with in the article is the comment about the brokenness of search - I think that’s actually functionally connected to what has made the fediverse what it is (I’m aware that’s a controversial position)

thanks! hadn’t come across that! will read asap!!

2026-03-06T15:44:59Z

In reply to nevent1q…g0th
_________________________

thanks! hadn’t come across that! will read asap!!

but I also think that evidence that LLM output varies in ...

2026-03-06T11:56:51Z

In reply to nevent1q…n5mg
_________________________

but I also think that evidence that LLM output varies in surprising ways in response to changes in features that we would not consider to be substantively relevant to the “task” as we envision it goes way beyond the moral judgments, there’s so many examples, and they seem really pertinent to evaluating science

the difficulty with low probability outputs highlighted in the embers of autoregression paper seems pretty relevant to the science context, I think

https://www.pnas.org/doi/10.1073/pnas.2322420121

pretty much the only thing I’ve been saying for the last three ...

2026-03-06T11:42:34Z

In reply to nevent1q…n5mg
_________________________

pretty much the only thing I’ve been saying for the last three years about LLMs is that claims about their performance and behaviour are empirical claims and we need to treat them as such ;-)

indeed

2026-03-06T09:48:24Z

In reply to nevent1q…cz7m
_________________________

indeed

Relatedly, the most productive way to deal with baseless comments ...

2026-03-06T09:47:25Z

In reply to nevent1q…x65t
_________________________

Relatedly, the most productive way to deal with baseless comments is to try and see what about the manuscript might have led them to go wrong….

not always, but quite often, there are fairly simple fixes that cut off entire paths to error. These can, in fact, make papers better, last but not least, because scientific papers are attempts to *communicate to a specific audience*

it is conceptually odd to have the evaluation of communicative success happen by anyone other than that target audience, and it feels like it would require lots of empirical work to show replacing a sample of the intended audience with some “other” (human or machine) didn’t lead to outputs that were ultimately less fit for purpose

THAT is an excellent point

2026-03-06T09:40:21Z

In reply to nevent1q…rtar
_________________________

THAT is an excellent point

that made me laugh on a gloomy morning… thank you!

2026-03-06T09:35:41Z

In reply to nevent1q…ej8e
_________________________

that made me laugh on a gloomy morning… thank you!

indeed… here’s the paper: https://doi.org/10.1098/rsos.241229

2026-03-06T09:29:40Z

In reply to nevent1q…e5wa
_________________________

indeed…

here’s the paper:

https://doi.org/10.1098/rsos.241229

the COI issue is an interesting one… given the current state of ...

2026-03-06T09:15:46Z

In reply to nevent1q…8khy
_________________________

the COI issue is an interesting one… given the current state of the literature on “deceptive” behaviour in LLMs should one allow an AI to review papers about AI performance from that perspective?

Tom, I wouldn’t class the repeated findings whereby LLMs give ...

2026-03-06T08:55:45Z

In reply to nevent1q…jwgj
_________________________

Tom, I wouldn’t class the repeated findings whereby LLMs give very different answers in response to accidental features of the input as compatible with “consistency” in the way it is implied here.

Just a few examples of many: the evaluation of moral judgments differing as a function of whether the two alternatives are labelled a) and b) or 1) and 2) (see Demberg et al), or the many examples in the Embers of Autoregression paper such as the accuracy of shift sequences depending on the size of the shift, or the accuracy of arithmetic operations depending on the actual numbers etc. etc.

there’s a fair bit of conjecture about the actual behaviour of ...

2026-03-06T08:14:21Z

In reply to nevent1q…n484
_________________________

there’s a fair bit of conjecture about the actual behaviour of LLMs as reviewers here, isn’t there?

the idea that AI will stay more relevant and be less biased than Reviewer 2 feels like it needs some empirical data…

the number of scientists has long been growing exponentially, and ...

2026-03-06T07:41:20Z

In reply to nevent1q…s85s
_________________________

the number of scientists has long been growing exponentially, and (rather unsurprisingly) so has the number of scientific publications, and now there are tools that allow each individual to become exponentially more ‘productive’…

I’m sure self-ranking will really help ….

that was exactly my response - everything we know about ...

2026-03-06T07:37:34Z

In reply to nevent1q…v42k
_________________________

that was exactly my response - everything we know about “impact” and the relative merits of funding foundational vs. applied research suggests to me that it is incredibly difficult, if not impossible to foresee how “useful” a piece of research will turn out to be….

(my favourite example is Boolean algebra, imagining Boole’s head of department and impact officers asking in 1847 “so what have you been working on, George? How exactly does that benefit society?”)

another band-aid for coping with AI induced productivity ...

2026-03-06T07:05:14Z

another band-aid for coping with AI induced productivity increases….

https://www.nature.com/articles/d41586-026-00059-0

and my perception is that, contrary to what one would hope, the ...

2026-03-06T06:38:46Z

In reply to nevent1q…l95y
_________________________

and my perception is that, contrary to what one would hope, the pre-print model looks set to collapse before commercial publishing does precisely because more information increases the need for more curation

https://write.as/ulrikehahn/is-ai-killing-scientific-reform

my question is more “what’s the feasible alternative” - at ...

2026-03-06T06:32:58Z

In reply to nevent1q…8y0j
_________________________

my question is more “what’s the feasible alternative” - at the core of the AI problem is the ability for (sheer endless) over-production relative to our capacity to process, evaluate, and integrate results. That structural problem doesn’t magically evaporate because you get rid of commercial publishers, or a particular unit of scientific record keeping and dissemination

given that science is a communal, inter-subjective, incremental endeavour it’s going to need *some kind of* records for information. And if the problem is “too much information”, how will the format alone solve it?

? and then what … how do we keep hold of and share scientific ...

2026-03-06T06:20:31Z

In reply to nevent1q…80y4
_________________________

? and then what …

how do we keep hold of and share scientific results?

I don’t understand the idea that papers are just units of measurement for job applications, frankly

I see- I agree that the context matters and you might also ...

2026-03-05T17:22:45Z

In reply to nevent1q…lkkg
_________________________

I see- I agree that the context matters and you might also definitely be able to `do activism' by presenting the evidence

but the core difference remains, to my mind, that not engaging with all of the evidence is 'bad science' while advocacy does not require that (rather, whether I do or do not, in communication or otherwise, is an instrumental decision: does it help or hinder..)

that is a helpful distinction and related to what I'm ...

2026-03-05T14:36:38Z

In reply to nevent1q…u2xh
_________________________

that is a helpful distinction and related to what I'm interested in here, but more in an implicit way, and I'm interested in capturing that implicit bit more clearly.
To put it differently, I could potentially be an expert on a topic but, say, publicly act as a lobbyist, so 'expertise' is relevant here by virtue of (genuine) expertise requiring a mode of engagement that engages with all of the evidence.

So expertise requires the particular mode of engagement that I think characterises science, but the fact that I'm an expert doesn't necessarily make all my communications in that mode.

And if what I'm trying to understand are lines like that between 'objective/subjective' or 'political/apolitical' I think that mode matters.

make sense?

this deserves a longer answer than I have time right now, but I ...

2026-03-05T11:18:13Z

In reply to nevent1q…7ns3
_________________________

this deserves a longer answer than I have time right now, but I think I'm not really in first instance thinking about the contrast between the scientist and the politician (ie the parlamentarian, or member of government) but between Me-the-scientist and Me-the-citizen.

As a citizen, I can freely advocate for actions/positions and there is no obligation (legal or otherwise) that in doing so I *have to* engage with other positions

at least, not in any straightforward way that I can see

I think they can blur, among other things, because I think there ...

2026-03-05T10:02:07Z

In reply to nevent1q…ys4v
_________________________

I think they can blur, among other things, because I think there can be contexts where Ulrike-the-scientist can rightfully say "this is the best course of action" (as a declarative statement)

so the action vs. fact distinction is only sort of correlated with the roles....

yes, I agree that's one way in which the roles are often ...

2026-03-05T09:58:22Z

In reply to nevent1q…2pef
_________________________

yes, I agree that's one way in which the roles are often different, but I think there's more to the contrast than that-

I am, after all, free a a citizen to engage in political advocacy (that's part of free speech), so there is an issue about the line between me saying things as Ulrike-the-scientist and Ulrike-the-citizen.

and, I guess what I'm arguing is that looking just at statements themselves isn't enough to distinguish these two roles

one immediate consequence of this is that issues around the ...

2026-03-05T09:48:06Z

In reply to nevent1q…gf7u
_________________________

one immediate consequence of this is that issues around the public pronouncements or stance taking of scientists will arise to the extent that it's not clear from statements themselves what mode of engagement underlies them