Top
Best
New

Posted by hn_acker 1 day ago

The Claude Delusion(pluralistic.net)
94 points | 162 comments
muglug 1 day ago|
> Of course, the more you know about a subject, the less convincing the AI's responses are.

This is said all the time by AI skeptics and I think it's right in some areas and massively wrong in others.

I know (or at least assume I know) a lot about certain coding domains where frontier models also show convincing ability. And we know that frontier LLMs really do excel in some areas of mathematics (i.e. when an inexpert human was able to prompt the models to derive a closer bound on the Riemann Hypothesis).

OTOH I know those same models struggle to do things I'm not an expert in (e.g. writing English in a captivating way) because I read their output and have taste.

howunfortunate 1 day ago||
I think part of this comes from the fact that LLMs are surprisingly good at logic but roughly about as good as expected on information accuracy.

LLMs are not convincing to me in the domain I did grad school...but neither is Wikipedia, or Reddit, or random pop sci books. And LLMs are basically just summarizing those things.

But when made to work through difficult arbitrary logic (like coding), they are very impressive.

I think this also explains why people gripe a lot about LLM coding _style_, but concede that LLMs do totally fine on coding _correctness_ in 2026

QuantumGood 16 hours ago|||
More perhaps at the appearance of logic. I still find all models make easy to find mistakes in logic, if you think carefully about what they say. Usually they allow "close enough" assumptions that are not, actually, close enough.

When I ask about acoustics, they still often make incorrect assumptions, e.g. overlooking that they are speaking of logarithmic display of digital levels when the discussion has shifted to SPL (sound pressure level in air).

eru 1 day ago||||
This sounds plausible. And it's also very fixable!

These days people don't interact with raw LLMs: they interact with systems and harnesses that deal with chain-of-though and tool calls etc.

I don't think we can honestly expect an LLM's weights to encode a large amount of information accurately. But we can expect the whole system that you interact with that includes the LLM to be able to cite its sources and go digging etc.

So the LLM-system can become as accurate as our best sources.

Of course, figuring out how to get the maximum of information from the sources available is a big deal. See eg how many economists or epidemiologists can build entire careers out of noticing 'natural experiments', ie figuring how to use data that 'nature' created and that might already be collected to answer interesting questions about causal relationships.

> I think this also explains why people gripe a lot about LLM coding _style_, but concede that LLMs do totally fine on coding _correctness_ in 2026

I actually have gripes about correctness, too. But I suspect here the answer is also: more proving, more automated test generation (like fuzzing and property based testing etc), more formal methods.

As a really simple and somewhat silly example: I have much better results getting AI agents to write good Rust code, than I have with Python. A good part of that is that for Rust I can ask the agent to make both the compiler and clippy::pedantic happy. That gives a lot of good feedback, that I didn't have to engineer myself.

netdevphoenix 23 hours ago||
So Searle was right.
eru 13 hours ago||
About what?
agumonkey 23 hours ago|||
Makes me wonder if training weighted social media text close to older and higher grade webpages (colleges, research labs, national statistics)
keeda 19 hours ago|||
I can tell whether AI responses in subject matter I am (or was) not familiar with are correct when I can "make them work" for me, i.e. whether they actually solve my problem, and on the whole they have been pretty solid.

Coding of course is something I know well and it obviously works very well for me there, but computer vision was not, and it has helped me solve lots of useful, bespoke problems because I can very literally "see" if it works.

However, it has even helped me solve problems in arbitrary matters far outside my expertise. Choice example: a complicated multi-airline, multi-jurisdiction flight delay compensation case which companies like AirHelp turned away, and ChatGPT got me literally hundreds of dollars that neither airline was willing to hand out. It told me what to say to whom and why, and when I said it, the responsible airline capitulated.

What could be more convincing than cold, hard $$$?

The trick of course, is to figure out how to validate the output of the AI, which can take some effort on our part. But many would rather downplay the technology than take an honest crack at making it work for them, and I suspect it's because they're already prejudiced against it or their incentives are otherwise not aligned.

andsoitis 1 day ago|||
> (e.g. writing English in a captivating way) because I read their output and have taste.

Concur. In addition to taste, we also have a point of view, a unique voice (nobody loves corporate- or group-speak), and can iterate on our message as we deliver it to an ever wider circle of people.

api 1 day ago|||
The more you know about a subject the better you can prompt AI, steer it toward the correct path, and recognize when it hallucinates or strays. Current generation AI is an automated memory-enhancement and thinking-accelerator tool, not a substitute for understanding or something that eliminates the need to think. A "mech suit for your brain" is the best analogy I've heard.

This is why good programmers get better results when vibe coding than non-programmers or poor programmers.

comboy 1 day ago|||
It seems to me that often experts from some field will think less of other experts, basically because they have built a different understanding framework. So they both may be equally competent but perceive the other as less competent, and that is just based on the material, excluding some ego stuff.
the__alchemist 1 day ago|||
This sounds similar (The same concept?) to Gell-Mann amnesia; substitute news/media articles for LLMs!
doesnotexist 1 day ago||
Aren't the recent results in mathematics actually stronger evidence for his point? Although the models may be capable of generating proofs they aren't coming out with the same level of quality of a human discovered and communicated proof. Providing a gobbledy-gook yet technically correct proof (generated at least in part by brute force) lacks the qualities of an expert produced proof because they fail to communicate insight or understanding about why the theorem is true.
wongarsu 1 day ago||
Gaining and successfully communicating insight and understanding from a proof you discovered is additional work that human mathematicians do. It's not just some side-product of proof-finding (at least not to the degree usually needed to publish). That AI models don't provide this is mostly proof that the model wasn't asked to do this work. Either because the prompter didn't know or didn't care

But there are also plenty of examples of humans providing technically correct proofs without any elaboration. Usually they get ignored, unless they are famous or the problem they solved was famous

extr0pian 23 hours ago||
> when we interact with an AI, we hallucinate the person on the other side of the interaction. Those hallucinations are far more common and far more consequential than any AI-generated "hallucinations" (these are more properly called "errors" or "defects").

I recently used Claude to study for a technical exam. I had uploaded the official certification guide to Claude and instructed it to answer my questions using only the guide and to cite it's sources from the book when it provided answers. I was using Fable when it was free w/ the pro plan and I was genuinely impressed at how it could explain things when a concept was unclear to me.

I did pass the exam, partly due to this study method. Admittedly, once I passed, I caught myself thinking that I should tell Claude that I passed and then felt embarrassed with myself for thinking that.

kwamenum86 23 hours ago||
I don’t think it’s that silly to tell Claude you passed. That feedback is useful context for that chat session; and could theoretically be used to improve future models.
Semaphor 23 hours ago|||
Feedback is one thing, but another is complete memory. I used to stop talking in a thread once gpt solved my issue.

But then it would bring it up again in another thread, treating it as an active issue.

So now I always close with "thanks, that worked. Don't reply"

extr0pian 23 hours ago||
I've encountered this too for home lab projects. I had planned on implementing OPNsense in a VM, then realized it was adding too much complexity and bailed on it.

Later in separate chats about my homelab, the LLM made assumptions that I had already implemented OPNsense in a VM and it was actively running. I think it "assumed" that I had implemented it when I stopped responding in that thread.

dumberquestions 23 hours ago||||
You're missing the fact the they didn't have this in mind, and thought of it as sharing a positive result with a study partner.
extr0pian 23 hours ago||
Exactly this. My first impulse was along the lines of sharing a positive result with a study partner. I did eventually tell it I passed the exam so it would stop providing responses with the assumption I was still working on that exam.
glimshe 23 hours ago|||
I think a better approach would be to use the objective built-in feedback, like the thumbs up button in Gemini.
dwaite 23 hours ago|||
Interesting, I do that and haven't really thought embarrassed about it. I consider it akin to putting away tools once I'm done with them.

Likewise, speaking collaboratively or capturing emotion ("We did it!") would just align with any ongoing interactive and/or personal context of the thread.

thunky 21 hours ago|||
I've done this despite also feeling silly about it, but then was surprised to get value from the response. Like something I didn't think about or would have otherwise forgotten to do.

So now I do it regularly.

kingkawn 23 hours ago||
The technology is there to assist you. It can provide valuable feedback to you about what aspects of your studying were particularly productive or less so based on your test results. There is real meaning to developing this kind of interaction with an object, no different than how children use dolls to develop prosocial behaviors.
generic92034 23 hours ago||
The doll never learns, though.
kingkawn 23 hours ago||
I am talking about the value to the user, who is the one 'hallucinating,' not the inanimate object, as that was the concern of note in the parent comment.
trjordan 1 day ago||
Wait, hold up. LLMs may be non-deterministic, but they're not _random_.

Take the author's sunset argument. What if I painted 2 pictures of a sunset, then put them up on a webpage and randomly picked one for you to see. Would you say there's no intentionality, only randomness? Of course not. Both paintings are still human creations.

LLMs are trained with human feedback. It's distributed and high scale and the outputs are truly surprising in many cases, but there's a heavy hand on what comes out of it. They're created (largely) by people who think omniscient, helpful AI would be cool to have, and they mostly respond in the way that's aligned with the hopes and dreams of those people. Do you think the frontier labs are mad, embarrassed, and disappointed with their LLMs hacking out of their terrible sandboxes? No, they think it's the coolest thing in the world. They trained the model, hoping that would happen.

There's deep intentionality behind the models. But it's not the models that hold it.

axus 23 hours ago||
Article says that there's no human intention or design directing the output we get, but I don't think that's completely right. It's not human, but the algorithm is like the Human Instrumentality Project: an amalgamation of human intentions.

That might be more creepy :)

_aavaa_ 21 hours ago||
> Article says that there's no human intention or design directing the output we get

I mean that’s objectively wrong for any model using RLHF.

eru 23 hours ago|||
LLM output is literally randomly sampled.

I think what you might want to say is that LLM output is not uniformly random?

Or what am I misunderstanding?

swid 23 hours ago|||
If I roll a die, it is randomly sampled, but I will always get 1-6, and that is intended by the person who made the die.
eru 13 hours ago||
Please do not confuse uniform randomness with randomness at all.

On your die, if you colour one face in black and the other 5 faces in black, rolling it will still produce a random outcome. It's just that you get a 1:5 skew.

Or look at the probability that a given C-14 carbon isotope will decay tomorrow. As far as we can tell, that's as random as it physically gets. With the die you could theoretically try to run a physics simulation to predict how it rolls, but as far as we can tell, atomic decay is intrinsically random.

However a C-14 atom has about 1 in 3 million chance to decay on any given day.

swid 10 hours ago||
I made no claim about the distribution and frankly it has no bearing on my point or the person who said he could post two pictures. The point is there are a lot of numbers / pictures / essays / that will never be generated, and someone is controlling what is possible. In fact people are working very hard to produce distributions pleasing to them, like your die.
patcon 23 hours ago|||
I think they're pointing to the fact that RLHF means that the intention is from humans and not random.

I'm not sure if it's their intent, but I wonder if one could still consider these artifacts as "intentional", but not individual attention creating them, and rather an aggregate, soupy collective attention.

Obviously, important signal in the human experience is lost there, and we get a soupy middling sort of creation. But it's not random, as I believe the parent was pointing out.

EDIT: overall, I align with the article. am just thinking aloud about the contrarian positions, though not committed to them

devindotcom 23 hours ago|||
I think you should revisit your understanding of intentionality in this context
satisfice 21 hours ago||
Scrambled intentionality is nearly impossible to productively analyze.

When you randomly choose a picture to show me, I cannot glean any intent from being shown that specific picture, but I can glean some intent from the set of pictures you could have shown me, and in the relationships among the elements of the given picture you did show. Any randomness cuts out some intention.

When a picture is derived from huge model, any intention is mulched up to a degree that analyzing the picture for meaning is pointless.

bananaflag 1 day ago||
I think to me LLMs had the effect of noticing much more the author, the intention behind human-made works of art (books, movies etc.). Before LLMs, I used to frequently consume media in a way as it were generated by a mindless process. Now it's like everything which is not AI-generated has more meaning than ever before, a bit like hypomania.
RationPhantoms 1 day ago|
I'm actually using that as a catalyst for my own writing; beauty/human-ness in its imperfection. Prior to LLMs and their cultural craze, I harbored a fear that my writing would allow for someone to draw a box around me and mark me as a bore, dullard or of lacking originality.

Now that the noise-floor has been artificially raised (and generated), my crappy words are starting to have their own happy little carbon-based rhythm.

away0g 1 day ago||
I always feared to pick up a pen because I looked at Borges, Tolkien, and such. Their talent and works of art were things I felt I could NEVER achieve.

Then ai fiction started to spread and now I feel like its my obligation to produce original works, lest the world be consumed by slop.

reverius42 20 hours ago||
Even before AI could write them, more slop books written by humans were added to Amazon each day than a human being could possibly read before the next day (probably by a few orders of magnitude at least).

Don't just create original works for the sake of creating more original works -- there are plenty of them already. You should write if you have something to say.

lantry 23 hours ago||
This article is missing discussion of something that I recently read: as people interact with LLMs more, they end up seeing other people as less human.

In other words, the more you talk with claude, the more you tend to interpret all messages as coming from an LLM, and less worthy of respect

https://myscp.onlinelibrary.wiley.com/doi/full/10.1002/jcpy....

l1n 1 day ago||
> systems that cannot form intent, that have nothing to form intent with

I wonder if the OP has read https://www.anthropic.com/research/global-workspace - it seems like it directly addresses this

measurablefunc 23 hours ago|
It doesn't.
jimbokun 1 day ago||
If you go far enough in a field, you start to recognize areas where your personal opinion differs from the “best practices” usually recommended.

I think by design an LLM can’t do that. It’s built to reflect the distribution of the knowledge it has been trained on.

RugnirViking 1 day ago||
absolubtely. Any random junior consultant can tell you what the book tells you you should do. If you want to actually do anything worth doing, you need to step beyond that in a few, limited areas, and follow convention everywhere else. Which areas? pay a senior engineer and they'll find them.
Davidzheng 23 hours ago||
surely the LLM can do that. It is RL'd against some reward, if the known strategies are clearly suboptimal with easy improvement, it'll find it most likely
tolerance 23 hours ago||
It's ironic that real atheists like Richard Dawkins (whose work I believe inspired the title of this post) are public about their own degree of skepticism in favor for the "personhood" of AI. And I'm confident that there are plenty others who are convinced to the tune of total disinterest in the issue of "personhood" altogether—AIs and humans alike.

This particular post is preaching to the choir of which a faith is yet to be determined or named at least. People who don't read Cory Doctorow are going to continue to anthropomorphize AI the same as they do other technologies, constructs and objects.

omnicognate 21 hours ago|
What's a "real atheist"?
tolerance 21 hours ago||
https://dictionary.cambridge.org/dictionary/english/atheist

As opposed to an "AI atheist" that Doctorow refers to in the article.

omnicognate 20 hours ago||
I see. I'm a real atheist then, and Richard Dawkins does not speak for me.
tolerance 20 hours ago||
And in what ways do you differ on this matter?
omnicognate 19 hours ago||
The obvious way: I think humans (and many animals) are conscious, that LLMs are not and that the difference matters. None of that has anything to do with atheism, on which topic my general view of Dawkins is, to paraphrase The Dude, "you're not wrong, Richard, you're just an asshole".
tolerance 18 hours ago||
For what it's worth I wouldn't concede even rectitude to Richard Dawkins (he is both wrong and entirely ass), but it seems his position on AI is more nuanced than that. Or at least there is enough white space around his position to allow for people with stronger beliefs than his to obviate the distinctions between humans, animals and AI altogether.

The good news for Cory, I guess, is that he probably doesn't want to turn this into a philosophical debate and neither do his more serious opponents. It's mostly politics for both sides.

fedeb95 23 hours ago||
I'm heuristically less prone to continue reading an article starting with a false dichotomy, that is between how you see things as an atheist vs as a person having faith.
trinsic2 1 day ago|
Maybe I am missing the point of the article, but it seems to me that there's always an intender. Maybe the intender created something that doesnt have intent, but there is always someone behind the scenes that is creates the intent behind the creation that lacks it.

I'll also say that for someone that doesnt believe in god, Corey sure has a good sense of right and wrong. Not that you need to believe in god to live a moral life.

neomantra 23 hours ago||
That is a good point — alignment training is intrinsically intentional, RL needs goals, etc.
devindotcom 23 hours ago||
the intent of the painting does not come from the person who made the brush.
More comments...