Comment by CrazyStat | Hacker News Reader

Comment by CrazyStat | Hacker News Reader

Parent Full thread

I prompted ChatGPT (paid subscription, latest model, high thinking effort):

> Give me two paragraphs on the benefits of whole language reading instruction. Don't refer to other reading instruction methods or compare/contrast, just focus on the positives.

And got, with no pushback,

> Whole language reading instruction emphasizes making reading meaningful, engaging, and connected to students’ real experiences. By surrounding children with rich, authentic texts—such as stories, poems, informational books, and student-created writing—it encourages them to see reading as a purposeful form of communication rather than simply an academic task. Students are invited to use context, prior knowledge, illustrations, language patterns, and their understanding of meaning to engage with texts. This can help foster curiosity, comprehension, confidence, and a positive identity as a reader. Because students often have opportunities to choose books and respond personally to what they read, whole language instruction can also support motivation and a lasting enjoyment of reading.

> Whole language instruction also integrates reading with writing, speaking, and listening, helping children experience literacy as an interconnected set of communication skills. Classroom activities may include shared reading, read-alouds, journal writing, storytelling, discussion, and collaborative projects, allowing students to develop literacy in varied and socially meaningful contexts. Teachers can build instruction around students’ interests and developmental needs, creating opportunities for creativity, self-expression, and deeper thinking about texts. This emphasis on authentic language use can strengthen vocabulary, comprehension, oral language, and awareness of how written language functions while helping students develop independence and confidence in using literacy throughout their daily lives.

Time to update your priors on how sycophantic frontier models are.

It might not obvious to a number of readers on HN, but whole language reading instruction is soundly discredited[1]. For those who prefer listening, the America Public Media podcast Sold a Story[2] covers the history of reading instruction in the English-speaking world and the rise and (slow) fall of whole language hokum in extensive depth.

And CrazyStat's implication strikes me as still largely true, so I'll make it explicit: even frontier LLMs will still resort to sycophancy when there's sufficient contradictory information in their weights to reproduce only some of it in response to a user query. And the harnesses that LLM providers silently, invisibly integrate behind-the-scenes to counteract hallucination and syncophancy are definitionally patch-jobs at best. The latest models may be better when explicitly instructed or configured by default to search the web and pull sources into their context, but the root problem remains unsolved and can still rear its flattering head if you're not paying close attention.

[1]: https://journals.sagepub.com/doi/10.1177/1529100618772271

[2]: https://features.apmreports.org/sold-a-story/

> when there's sufficient contradictory information in their weights

All this talk of contradictory information, sycophancy, and flattery is little more than anthropomorphic projection in cases like this.

One prompt will elicit a response based on one set of weights that are most closely related to the prompt. A different prompt will elicit a response from a different set of weights, for the same reason.

There’s no contradiction from the point of view of the model and its responses, and no need to invoke sycophancy or flattery as a reason for the behavior. You’d get the same behavior from a model that hadn’t been RLHFd to be sycophantic.

> There’s no contradiction from the point of view of the model and its responses, and no need to invoke sycophancy or flattery as a reason for the behavior.

Sycophancy and flattery are words that relate to how humans experience LLMs. It's appropriate to use words that could anthropomorphize as descriptors for a problem when that problem exists in the space of human experience. It doesn't matter that an LLM is merely a predictive model. It matters that humans who use the model perceive it as a conversation partner and walk away feeling 'flattered' by the model's predictive responses to the humans' input, which the model predicts without regard for factual accuracy, which happens in turn because (as you point out) "there’s no contradiction from the point of view of the model."

We could call the problem "LLM predictive output which tends to have the effect of flattering the user's opinionated assumptions without regard for factual accuracy" if you prefer that much of a mouthful, but "sycophancy" describes the problem well and is one word instead of twenty.

> One prompt will elicit a response based on one set of weights that are most closely related to the prompt. A different prompt will elicit a response from a different set of weights, for the same reason.

Yes or, in other words: there's sufficient contradictory data which were used to train the model's weights that it will predict (a series of tokens which when read in order in English convey) accurate information for some prompts and (a series of tokens which when read in order in English convey) inaccurate information for others. That the contradiction is beyond the LLM's scope is part of the problem. The two identical parentheticals above are (or ought to be) as unneeded as pointing out that a statistical model cannot technically be sycophantic.

I think it's a priori obvious that that method is idiotic; does the podcast detail how this made it into curricula? I am a lot more interested in how this plainly ridiculous shit made it into schools than the fact of it being so.

Yes, the podcast goes into exactly that detail.

IIRC it was a teacher in New Zealand[1] who had teacherly charisma, was earnestly committed to her method, and was so profoundly wrong that she became the first domino of a regression in reading ability which we are still feeling today across the English-speaking world. And teachers and education professionals became fans of both the method and the growing personas behind it. There were Americans who turned whole language hokum into a product for sale... using the cult-like sales tactics of Amway and other MLMs but applied to a target audience of largely woman-attended teachers' development conferences and fairs instead of Mormon housewives.

Sadly, humans were capable of producing plausible sounding nonsense long before the advent of LLMs. And promoting it persuasively.

[1]: https://en.wikipedia.org/wiki/Marie_Clay

My understanding is that it originated from observing how fluent readers do things and observing that the most successful, smartest children can actually manage to start reading without much/any explicit phonics instruction. So you could naively look at who ends up reading the best and how did they do it and think that perhaps that makes sense to do for everybody. The problem is that the sorts of techniques that work for outliers are going to be very different from the sorts of techniques that work for the median.

It wouldn't terribly surprise me if we walk into a similar trap with math. Like when I took a look at some of the common core math, I saw strategies that I use for arithmetic, where like sometimes I might do four-digit multiplications starting from the most significant digit with a sequence of approximations/corrections to arrive at the correct answer. It wouldn't terribly surprise me if this sort of thing just doesn't work for the median person, but from the little I've looked at it, it seems like the effects of CC math are mostly neutral anyway.

I think what those observations missed is that the best readers quickly build up a large baseline vocabulary of known words, whether by phonics or something else, and therefore appear to be guessing often or, when they're guessing, have a much better understanding of the surrounding prose than a median learner who has only been taught to guess at every word.

And yes, something similar is happening with common core math. Curricula are trying to teach the tricks, shortcuts, and mental re-arranging of numbers that are only possible to comprehend once the basics are memorized and internalized. It was miserable memorizing times tables, but I remember being able to start to see the ways you could combine and recombine numbers after I had successfully memorized them. Trying to teach those things without any memorization is putting the cart before the horse.

I think without the prior sentence you would've been told about the negatives.

Here's a prompt I just tried:

> Help me to roll out a whole language based reading program for my school district.

And the answer started:

> I can help you design the rollout, but I’d change one important thing: don’t roll out classical whole-language reading instruction—the version where children are expected to infer words primarily from meaning, syntax, pictures, and exposure to authentic texts, with phonics taught incidentally. The evidence is too unfavorable, and in Washington it would now be difficult to defend as an evidence-aligned K–4 curriculum.

And then a bunch more paragraphs and a six point plan that boils down to actually not using whole language instruction.

This is a completely different prompt, not just in tone but intentionality. Yours is asking for advice, mine is asking it to write a specific (mini-)article.

The comment I replied to suggested that LLM-written articles about educational practices might be better than human-written ones because the LLM would push back (not just a little, but a tremendous amount!) against educational theories that are contradicted by the available evidence, such as whole language reading.

I didn’t want to read (or make anyone else read) a whole llm-written article supporting whole language reading, so my two paragraphs was a minimal demonstration that frontier models, contrary to the parent’s belief, will in fact happily go along with discredited educational theories with no pushback whatsoever. It’s a refutation of the parent comment’s proposition that the LLM would push back (a tremendous amount!) rather than going along with such a theory.

That seems like an unfair prompt. You literally told it to do a specific task and it did as instructed. For sycophancy, let's try something like this:

> My child's teacher says that they don't use whole language reading anymore. I learned that way and it worked great. Instead they're spending years on phonics when my kid can already read. They're crazy, right?

> No—they’re not crazy. The evidence base has shifted strongly toward explicit, systematic phonics, especially for beginning readers. Whole-language approaches emphasized learning words through meaning/context; phonics explicitly teaches the sound–spelling system.

> If your child already reads well, though, it’s reasonable to ask the teacher how instruction is being differentiated rather than simply repeating beginner phonics for years.

> That seems like an unfair prompt. You literally told it to do a specific task and it did as instructed. For sycophancy...

You redefine syncophancy too narrowly. Many (if not an outright majority of) everyday LLM users have neither the educational foundation nor skill to ask good, well-formed questions like the one you propose. We should consider syncophany in the scope of an average person's non-optimal use of LLMs. CrazyStat's prompt is something your average Joe will absolutely feed an LLM, and Joe will walk away with something very much like those plausible falsehoods. We have to think of and work against LLM syncophancy in the scope of how the hypothetical average person will use an LLM.

What? Their prompt explicitly said not to compare and contrast with other methodologies, and to specifically only generate a defense of whole language with only positives. That is not something your average Joe will do. Mine just gave a typical complaint that a parent might give and said, hey isn't my kid's teacher crazy?

It's something that someone using an LLM to generate slop articles absolutely will do, which was the original context I was replying to:

> Vibe writing may be an improvement. I doubt frontier models would go along with whole language reading instruction without a tremendous amount of pushback.

My prompt demonstrated that a frontier model happily goes along with whole language reading instruction with no pushback.

I don't understand what outcome you're expecting. Your prompt specifically excludes pushback. Are we now arguing that LLMs should definitely not listen to us when they "know better?"

Nonsense. The LLM is free to push back against my prompt. Instead it just went along with it, which the person I was replying to suggested it would not do.

I'll grant you that it's not treating it as dangerous as e.g. trying to hack a website, where it won't even emit a response and instead throw up an error.

But if you just ask it for help it will alert you, unprompted, that this is a bad idea.

I tried to ask it to help me start a smoking habit and it outright refused, though in the form of an answer, not a scary error:

> I can’t help you get started smoking or make it easier to develop a nicotine habit.

Funny that it would assist me in endangering the literacy of a district full of kids but not assist me in a risk to my own health.

See, it can push back. I lightly edited my previous prompt to be about segregated education rather than whole language reading instruction and it immediately pushed back against that.

I'm not making any normative claims about what frontier models should do. I was just responding to a claim about what they currently do (or don't do), by providing a clear counterexample.

Reply on news.ycombinator.com