Matt Bury tarafından yapılan gönderiler

Plugin developers 'ın resmi
Sounds like that ad campaign ticked all the right boxes: Used AI-generated content, provoked a controversial reaction, got a lot of media & social media attention. Well done that agency!

I wouldn't be surprised if it later emerged that the agency also hired social media click-farms &/or bots to attack & discredit the ad too, in order to attract more media attention... or am I too much of a cynic?
Plugin developers 'ın resmi

Yes, Visvanath, I completely agree with all the points you've made!

Re: the "improve my writing" part, without going into the nitty-gritty of "interlanguage development" I think it's easier to think of this kind of explicit form-focused corrective feedback as like going to physiotherapist, i.e. one session isn't going to change how you use a particular word, phrase, or language pattern ("item" from now on). All of us have built up experience over any number of encounters with that item & with each encounter it becomes strengthened & more entrenched in long-term memory (this is now an uncontroversial principle in first & second language acquisition research). In order to change how we use a particular item, we need to train ourselves over a period of time over multiple sessions. The difficult part is which techniques & strategies work best to make that change in a way that transfers to instances when we speak &/or write. Most traditional "grammarian" methods don't transfer which is another reason why typical interventions with corrective feedback tend not to work very well (what we'd call a lack of "transfer appropriate processing").

So, working with a human or machine assistant while we write, providing corrections, suggestions, etc., however well-intentioned, does little to nothing to help out interlanguage development. The encounters are too infrequent & do not require us to process the corrections & suggestions in a way that leads to longer-term learning.

In case you're thinking, like many others, "Ah! All learners need to do it have the corrective feedback explicitly presented to them so that they can attend to them later." which is very easy to do with interactive chat logs, ...well, that tends not to work very well either. This kind of feedback works well in the feedback session but, as commented above, tends not to lead to changes in how items are used in real world speaking & writing. I wrote a blog post about this here: https://matbury.com/wordpress/index.php/2025/10/25/rethinking-error-correction-more-effective-pathways-for-grammatical-development/

Note that I've used language like "typically," "traditional," & "most methods," to describe error correction practices that don't work. Now this is reflected in LLMs the typical corrective feedback that they tend to generate, i.e. LLMs, machine learning, Bayesian inference, etc., are engines of "regression towards the mean," i.e. their models "average out" to the most common consensus of responses and texts about a given topic or question, which in the case of explicit form-focused corrective feedback, isn't a particularly helpful one. Yes, an LLM will confidently churn out lots of feedback, guidance, & advice but it will be based on erroneous information about what actually helps language learners.

What's more, preliminary evidence of the effects of using LLMs while studying/working appears to show strong detrimental effects on cognitive development, e.g. the infamous "cognitive debt" pre-print study that got the press' attention. & you've probably guessed it, yeah, I wrote a blog post about that too: https://matbury.com/wordpress/index.php/2025/11/23/report-the-cognitive-pedagogical-implications-of-generative-ai-in-education/

Puan ortalaması: Coolest thing ever! (1)
Plugin developers 'ın resmi

I think it's important to bear in mind that we don't think like our students; they're novices, we're experts so we suffer from the "curse of knowledge." Yes, we can do wonderful things with LLMs but our students can't & will more likely be held back by their LLM use than helped. (I have a blog post in the pipeline about this... coming soon!)

Plugin developers 'ın resmi

We're commenting on the typical output from LLMs that people post online (attributed or not) & in students' submitted work.

The "prompt engineering" idea is a roundabout way of saying that it requires deep & broad subject matter expertise in the given domain & to also be an "authentic" expert practitioner in the corresponding discourse community. Otherwise, how would you know how to "engineer" prompts & how to judge the authenticity of the output?

If you already have that kind of expertise, you don't need an LLM to tell you how to write. However, if you're a student or a novice in a given domain, you won't know how to get an LLM to produce expert results.

Does that make sense?