jjani

"Agenticness" depends so much on the specific tooling (harness) and system prompts. It mentions Copilot - did it use this for both? Given it's created by Microsoft there's good reason to believe it'd be built yo do especially well with GPT (they'll have had 5 available in preview for months by now). Or it could be the opposite and be tuned towards Sonnet. At the very minimum you'd need to try a few different harnesses, preferably ones not closely related to either OpenAI/MS or Anthropic.

This article even mentions things like "Sonnet is much faster" which is very dependent on the specific load at the time of usage. Today everyone is testing GPT-5 so it's slow and Sonnet is much faster.

We're 4 months later, a century in LLM land, and it's the opposite. Not a single other model provider asks for this, yet OpenAI has only ramped it up, now broadening it to the entirety of GPT-5 API usage.

This web app wouldn't survive unless it already had backing of the Microsoft level or did offer immense unique value to its users.

They're already doing so, just not receiving it. Microsoft made a yearly profit of 101 billion dollars, or better phrased, 101,000 million dollars. The solution isn't "the consumer pays more", it's "investors earn less".

LLMs with web search enabled also give sources.

There's not much reason to believe LLMs are incapable of outputting a similarly well-structured lesson plan as those provided by third parties. It fits well into what they're capable of.

The problem is, again, that the teachers in question clearly don't care about the quality of their teaching in the slightest. Are LLMs a tool that could amplify their awfulness further? Sure, but at the same time they can amplify the goodness of the hopefully larger group of teachers who do care.

First time learning that scraping and training on data that they have often been explicitly disallowed to obtain for free or for that purpose by the rights holders is "fundamental to how LLMs work". If not, then there is no reason those who gathered the information wouldn't stand to profit by selling LLMs this data.

Yes, this proposal has been around since long before those parties got as big as they are now, but even back then the quadrants were roughly similar, and as such the level of support (including now looking to pass, unlike before) has also roughly been in step with the growth of those parties.

IME the only reliable way around it when using an LLM to create blog-like content is to have actual hard lists of slop to rewrite/avoid. This works pretty well if done if correctly. There's actually not that many patterns (not hundreds, more like dozens) so they're pretty enumerable. On the other hand, you and me would still be able to tell if only rewriting those things.

Overall the number one thing is that the writing is "overly slick". I've seen this expressed in tons of ways but I find slickness to be the most apt description. As if it's a pitch, or a TED presentation script, that has been pored over and perfected until every single word is optimized. Very salesy. In a similar vein, in LLM-written text, everything is given similar importance. Everything is crucial, one of the most powerful X, particularly elegant, and so on.

I find Opus to have the lowest slop ratio, which this benchmark kind of confirms [1], but of course its pricing is a barrier.

[1] https://eqbench.com/creative_writing_longform.html

> As a former teacher, I know you need to have a good grasp of the material you are using in order to help students understand it. The material should also be in a similarly structured form thoughout a course, which will reinforce the expectations of the students, making their mental load lesser. The only way to do this is to prepare the material yourself.

It's absolutely necessary to have a good fundamental understanding of the material yourself. These teachers abusing AI and not even catching these obvious issues, clearly don't have such an understanding - or they're not using any of it, which is effectively the same. In fact, they're likely to have a much worse understanding than your average frontier LLM, especially given this post is about high school level teaching.

> The only way to do this is to prepare the material yourself.

As brought up in other comments, what is yourself? For decades teachers have been using premade lesson plans, either third-party, school supplied or otherwise obtained, with minor edits. All teachers? Of course not, but it's completely normalized. Are they doing it themselves? If not, then the remainder did it together with Google and Wikipedia. Were they also not doing it themselves? Especially given how awful modern Google is (and the worldwide number of high school teachers using something like Kagi will be <100 people), simply using a frontier model, especially with web search enabled, is simply a better version of doing that, if used in the same way.

It's all PR, entrenched interests.

The majority of times I see things like this it turns out that it's either:

- The "they've built it wrong" case; this one is the most common. People using - or in this case being made to use at work - tools that behind the scenes all use very cheap models (e.g. 4o-mini) with little context, half vibe-coded up, to save costs. The company making "MagicSchool" doesn't care, they want to maximize those profit margins and they're selling to school administration, not teachers, who only look at the costs and don't ever actually use the products themselves. Just like classic enterprise software in traditional companies. They need to tick boxes, show features that only show the happy path/case. It is perfectly possible to make it high quality, in a way that adds value, doesn't make shit up, and is properly validated. But especially in this niche, sales trumps everything. The hope is that at some point, this will change. We've seen the same play out with enterprise software to an extent; new such software does tend to be more usable on average than it used to be. It has taken a long time to get there though.

- The "you're holding it wrong" meme; users themselves directly using tools like Microsoft Copilot, 4o and friends (very outdated, free tiers, miles behind Claude/Gemini 2.5 pro/o3/etc.), along with having zero idea about what LLMs can and can't do, and obviously even less of an idea about inherent biases and prompting to prevent those. This combined with a complete lack of caring, along with a lack of competency - people lacking the basic critical thinking skills necessary to spot issues - is a deadly combo.

Of the problems with tasks and outcomes named in that thread, the large majority can indeed be done already with LLMs in a manner that both saves time and provides better quality than the level of those teachers rightly being criticized there. Teachers who are not even checking the output obviously don't give a single damn anyway, and that tells you enough about what the quality of their teaching would've been like pre-LLMs.

They might not fit your personal definition of "openness", but they do fit many other equally valid interpretations of that contept.

[1] https://www.bbc.com/news/articles/ce9jv1xyxm0o

But really, being deeply ingrained into society and having personally worked at such places here in Korea is worth more than flawed statistics, which really all statistics are when it comes to these topics. I've literally personally interviewed the very few applicants we got as a small company outside of Seoul. They didn't know the difference between a GET and a POST; these were people applying as a backend API developers. We had to train people from absolute zero. And this was at a "modern" workplace ran by young leadership, which should theoretically be much more attractive - one of the, if not the number 1 most commonly named reason for not wanting to work for smaller companies (and even moreso outside of Seoul) is "old(-school) working culture".

https://www.segye.com/newsView/20250327517804

These companies aren't like Silicon Valley startups abusing H1Bs, hiring foreigners because they're cheaper, Koreans genuinely aren't applying.

Go ask her and her friends how many 중소 outside of 수도권 they've applied to. Again, if they've studied something like German language or archaeology then sure, but people aren't getting jobs in those fields in Sweden (to name a completely random high HDI country) either.

In fact, the parents often actively enable and support this strategy. Unthinkable in the West.

This is nothing like "I sent 300 resumes to every place in the country and can't get a job". Of course, such people do exist, but those are in majors/fields in which it was already near-impossible 10 years ago to get a job in most of the world. If everyone in Denmark would go "Maersk, Novo Nordisk, Carlsberg or bust", you'd see the exact same.

Honestly, your story indicates that your sister-in-law and her friends are likely at least middle class. Those who simply can't afford to sit around for 2 years (lower class) do just get a job straight away.

I'm also pretty shocked how HNers don't seem to notice or care, IMO it makes it unreadable.

I'd write an article about this but all it'd do is make people avoid just those tells and I'm not sure if that's an improvement.