https://www.404media.co/teachers-are-not-ok-ai-chatgpt
A fantastic article.
Soon after I read the Jason Koebler article at 404 Media I had my first intersection with AI mucking up education. A student group used it to generate an experimental plan.
This has made me realize how much of a nightmare it would have been if I had gone through the system.
I have clinically-diagnosed dysgraphia and effectively can’t handwrite. The quality of anything I write suffers dramatically when I have to handwrite it, or when I have a time constraint. I was provided accommodations when I entered university which allowed me to type most all exam essays.
Now there’s a renewed emphasis on writing essays in-class:
Now the essays are written in-class, and treated like mid-term exams. My quizzes are also in-class.
from the article
But it’s not just education.
A key factor that allowed me to distingusih myself when I applied for this job was, essentially, along with the regular interview, to write an essay proving that I had a sound understanding and could rapidly teach myself the relevant expertise. This would be unworkable nowadays.
Despite the enormous success of these computer approaches, our simple models are still crucial for developing understanding.
To quote the Nobel laureate Eugene Wigner
“It is nice to know that the computer understands the problem, but I would like to understand it too”.
*The Oxford Solid State Basics *
Steven H. Simon
Diversity of opinion, preference, and experience is a fundamental trait of being human. If you were to ask “Tell me a joke” or “What is the best book of all time?” to the next five people you talk to, with high likelihood you would receive five different answers. It is reasonable to expect language models to generate responses that have the same level of diversity as humans. Yet, when we ask models such as GPT-4 to recommend a movie or Claude 3 to suggest several vacation destinations, we often receive variations of the same few ideas—a phenomenon known as mode collapse Hamilton (2024).
Paraphrasing and system prompting are only marginally effective—indicating
that asking a model for “creative outputs” does not work very well. In-context regeneration is the most successful approach, with all three models roughly matching the diversity of human writers (the dashed lines in figures). This suggests that state-of-the-art LLMs can indeed generate diverse answers when explicitly constrained by their previous outputs in context. Under this strategy, GPT-4o and Gemini 2.0 Pro even surpass the cumulative utility scores of human writers due to the added diversity.* While our findings indicate that prompt engineering could partially address diversity limitations, they also reveal that this diversity is not inherently built into the models’ output distributions. Rather, it must be deliberately elicited through specific prompting techniques.
Because datacenter companies can afford to pay a lot more for electricity than your average consumer, and because datacenters are typically built in places that have unusually low power costs, they very quickly end up driving up the price of power for everyone else. Indeed, we’re already seeing that power bills for some consumers near power grids have gone up by more than double.
Oh now this is an interesting complication on the energy use discussion which I was pretty dismissive of above.
https://nicholas.carlini.com/writing/2025/are-llms-worth-it.html
These papers actually try to consider the legal implications of machine learning and copyright directly. Katherine Lee, who helped lead several of the papers I mentioned above, wrote a massive 186 page report on this topic with A. Feder Cooper and James Grimmelmann. They also wrote a followup report discussing things in more detail.
https://nicholas.carlini.com/writing/2025/privacy-copyright-and-generative-models.html
I believe that even current models are largely sufficient to allow the vast majority of people to solve meaningful tasks they could never have solved before just by asking for the solution.
https://nicholas.carlini.com/writing/2024/how-i-use-ai.html
Yeah, I’ve seen this. Colleague was able to patch a bug in a piece of software.
https://blog.plover.com/tech/gpt/claude-xar.html
Will I lose anything from having Claude write that complex
parser.add_argumentcall for me? Perhaps if I had figured it out on my own, on future occasions I would have remembered theconst=5anddefault=1specifications and how they interacted. Perhaps.But I suspect that I have figured it out on my own in the past, more than once, and it didn’t stick. I am happy with how it went this time. After I got Claude’s explanation, I checked its claimed behavior pretty carefully with a stub program, as if I had been reviewing a colleague’s code that I wasn’t sure about.
I immediately set out programming and debugging, by which I mean “I pasted the error messages into Aider and let Gemini figure out what the hell was wrong”. The AI very quickly fixed all the stuff that was wrong, and the display sprang to life, showing the TRMNL logo! Isn’t the future amazing?
Really, TRMNL’s firmware required minimal changes to work with the Waveshare driver. Most of the changes were just changing the pin numbers from one board’s to the others, as well as adapting for the fact that the Waveshare board doesn’t have a button or a battery, and uses an ESP32 instead of an ESP32-C3.
https://www.stavros.io/posts/making-a-trmnl-device/
https://www.cira.ca/en/resources/news/domains/10-best-ai-tools-every-business-should-be-using/
I appreciate the work the author put into this article, it’s not just fluff, but I don’t see that I would want any of the services they recommend playing a role in my life.
