Speaking of... It's interesting. It's a sociological issue as much as a technical one. Profit motive is hinted at, but ordinary people still have to remain consumers of businesses that lower their standards in this way. I don't fully understand it myself.

Borrowing from something Colin Fraser said on Twitter some while ago, but it is strange that different models are even offered for the same tasks, if some of those models are better (somewhat more reliable/accurate) than others at the task in question.

Added to the reading list (open-access is nice too). link.springer.com/article/10.1...

I have awoken from this particular slumber. Just in time for the holidays

The Ilya interview is worthwhile, though most of it leans into a very lengthy discussion of the problems (he sees) with imagining a hypothetical AGI and its impacts. Which gets a little too far from being grounded IMO.

"The amount of pre-training data is very very staggering, and somehow a human being...with a tiny fraction of pre-training data, knows much less, but whatever they do know they know much more deeply. Somehow. Already at that age, you would not make mistakes that our AIs make."

I just don't understand how the FedEx and UPS websites are as bad as they are. Pages that never load, endless loops, sign in credentials not recognized. It has to be intentional.

I don't think there's some deep meaning behind this, but there is some convergence among the most vocal language-is-for-thought scholars and the language-is-for-communication scholars on LLMs not being sufficient for general intelligence. Not total, but it's there. Worth observing.

Props to @benjaminjriley.bsky.social for bringing some CogSci to the tech world. Worth a read 👇

More a parlor trick than anything, but I guess they trained Gemini to explicitly say it hallucinated when it a user seems like it's calling out a hallucination (I imagine a lot of user prompts look like mine here in response to an obviously wrong answer).

Just thinking about this...like I say in the thread, I think the targets are good, as ideals. I'm just struck by how *direct* the pathway is from SV hype framing to the White House. This EO doesn't get written - at least the way it is - if there wasn't a very open channel between the two.

"The Genesis Mission will build an integrated AI platform to harness Federal scientific datasets...to train scientific foundation models and create AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs." www.whitehouse.gov/presidential...

Of the opinion that the state of the art has improved since 2022 (I dont care if we call it "reasoning" or whatever for this), but I've been thinking about how this would have played out if expectations were set appropriately by Altman and co. at that time.

I'll always instinctively side with those being threatened with obsolescence, but damn there really is a need for better AI critics

Interesting listen www.youtube.com/watch?v=DteP...

Ended up having somewhat mixed feelings about the first half of this paper, but it turned out to be pretty solid re: testing and deployment of genAI for safety-critical systems. Echoes a bit of what I've written about before wrt to lower standards for deployment of DL models in particular.

Journalist challenge: Use “Machine Learning” when you mean machine learning and “LLM” when you mean LLM. Ditch “AI” as a catch-all term, it’s not useful for readers and it helps companies trying to confuse the public by obscuring the roles played by different technologies. 🧪

Worth a try if you've had more trouble finding quality papers recently 👇Finding it useful already with a single, not very refined prompt.

Related panel discussion, also good (more wide ranging). Choice quote from Tomasello: "If all life on earth was wiped out, ChatGPT would sit on my desk doing nothing, but my home heating system - run by a thermostat - would keep heating my house...it has a goal..." www.youtube.com/watch?v=oJ9Z...

Some people saying she wants to be the GA governor, but this doesn't read like a statement in the run up to a governor's race...

To be read. Possibly low-hanging fruit, but Missy Cummings tends to be better than most. openreview.net/forum?id=uEY...

Thinking about the Yale professors who fled the US out of fear of prosecution while Mamdani and Trump pal around in the Oval

Pretty significant for those interested: "[The Public Investment Fund's] representatives have begun telling international investors that it is all but unable to allocate any more money for the foreseeable future, six people with knowledge of those discussions said." www.nytimes.com/2025/11/19/b...

One ironic consequence of this (referencing G. Marcus) is that, if you want substantive, focused research in safety and reliability in ML/AI these days, you're more likely to find it in academia than industry.

As someone who was very supportive of the original motivation for the ARC-AGI benchmark, I've come around to thinking there's two ways of looking at it: (1) It is directionally significant that the first version could not be saturated until test-time "adaptation" techniques came around.

This might have been meant as a hot take, but I do think there's a very substantive point here. And I also think a similar point applies to LLMs (as he says in the thread).

Good talks by Mitchell and Gopnik (title only says Mitchell). Centering on the question of whether cognition is even the right frame to evaluate AI models. youtu.be/yKbeCvOKvMc?...

I think these blinded tests have lost the plot. I don't normally say something is "academic" as a criticism, but this sort of thing is academic. Fine-tuning and other sorts of stage-setting to boost the appearance of humanlike outputs in restricted domains is interesting yet not especially useful.

The issue with bubbles isn't that they're built completely on hoaxes. A few may be, but most are built on overexcitement about a very real technology and its promises. People just get way ahead of themselves, and think that everything that glitters must be gold.

Trying to write a tech policy piece for the first time since maybe March/April, and I just don't know how to write "so and so agency should implement such and such policy" knowing that no such thing could possibly happen right now.

"It is not enough, therefore, to build a machine that could use words (if that were possible), it would have to be able to create concepts and to find for itself suitable words in which to express additions to knowledge that it brought about. Otherwise, it would be no more than a cleverer parrot...

Larry Summers being on the board of the OpenAI nonprofit has kind of flown under the radar. Arguably a more influential position than his (current) Harvard affiliation.

One of the reasons why debates about phenomena like consciousness seem unending is because the only way to observe it is to be a subject with consciousness. So far as we know, only humans have consciousness. We can only be sure of our own. We assume others are because we're made similarly. 🧵

Some Fukuyama content for the Fukuyama enjoyers out there youtu.be/-GDAEzuHn1U?...

Muted "Nuzzi." Still seeing Nuzzi content.

Insightful piece, echoes much of what I've come to in my own writing over the years. I'd just suggest an additional problem for LLM writing beyond how they're trained: part of the oomph of writing, as Nathan discusses, is the person's voice. But that is a result of idiosyncratic motivations to use

You tell 'em Yann

Purely anecdotal, but I have been coming across more published articles that either don't pass the smell test or have hallucinated citations. Came across a piece that cites a real article I co-authored, but lists my name as Victoria Carchidi, who is a real but completely unrelated researcher.

What an absolute disgrace Chomsky made of his last years...

Silver lining of a possible AI downturn: it would give me an excuse to learn more about tech (and tech policy) I've always been interested in but never had time to dive into. Renewables are probably top of the list.

This past few months is probably the most difficult period I've had staying focused on any one thing. Feels like everything is in a state of transition, and doesn't make much sense to think about much beyond the next year or two.

Still thinking about this quote from Larry Summers, which is by no means a novel attitude, but it expresses an attitude that's usually kept a least somewhat covered so clearly. And I keep thinking about how destructive an attitude this is for the elite of an advanced society to hold.

Good stuff here. I saw something from Benedict Evans recently where he noticed that the term "AI slop" at some point shifted from its original meaning of "trash output" to "anything automated by an LLM." It seems like having a barrage of even *better* outputs in areas like recruitment becomes slop

Appreciate the general idea in this thread, but I think this answers the original question about why some people can't/don't use them appropriately: the thrust of the current agent craze is that the models are...agents. They shouldn't need handholding. Why learn when you can automate?

To be read. philpapers.org/rec/BLOCOM-3

I accepted some time ago that pretty much anyone I've been influenced by intellectually who was born before a certain year will have made...questionable decisions.

There's also the intellectual autonomy issue here, which I repeatedly bring up because so long as LLMs' quality of output are dependent on the competencies of the person prompting them, it's difficult to say the LLM has mastered the skills necessary to produce the output in any sense relevant to us.

I think it's true the paradox is less useful/clean cut than it used to be. I've been thinking recently that it also started getting misapplied in the 2023-present period. (gyges may not agree with this part, just my hot take)

Reminded of some commentary from the pre-ChatGPT era because of the LeCun news. AI Winters are seen as being in the rearview mirror, deep learning might be capturing general principles of intelligence, but it's useful enough either way to put those concerns aside. ojs.aaai.org/aimagazine/i...