I'd rather read the prompt

I've already asked a number of colleagues at work producing insane amount of gibberish with LLMs to just pass me the prompt instead: if LLM can produce verbose text with limited input, I just need that concise input too (the rest is simply made up crap).

I like the author's take: it isn't a value judgement on the individual using ChatGPT (or Gemini or whichever LLM you like this week), it's that the thought that went into making the prompt is, inevitably, more interesting/original/human than the output the LLM generates afterwards.

In my experiments with LLMs for writing code, I find that the code is objectively garbage if my prompt is garbage. If I don't know what I want, if I don't have any ideas, and I don't have a structure or plan, that's the sort of code I get out.

I'd love to hear any counterpoints from folks who have used LLMs lately to get academic or creative writing done, as I haven't tried using any models lately for anything beyond helping me punch through boilerplate/scaffolding on personal programming projects.

  Come up with a self-contained single room of a dungeon that involves an 
  unusual puzzle for use with a DND campaign. Be specific in terms of the 
  puzzle, the solution, layout of the dungeon room, etc. It should be totally 
  different from anything that already exists. Be imaginative. 

I'm far from the first to make this observation but LLMs are like anti-compression algorithms when used like that, a simple idea gets expanded into a bloated mess by an LLM, then sent to someone else who runs it through another LLM to summarize it back to something approximating the original prompt. Nobody benefits aside from Sam Altman and co, who get to pocket a cool $0.000000001 for enabling this pointless exercise.

I think the answer to the professor's dismay is quite simple. Many people are in university to survive a brutal social darwinist economic system, not to learn and cultivate their minds. Only a very small handful of them were ever there to study Euler angles earnestly. The rest view it as a hoop they have to jump through to hopefully get a job that might as well be automated away by AI anyway. Also viewed from a conditional reinforcement perspective, all the professor has to do is start docking grade points from students who are obviously cheating. Theory predicts they will either stop doing it, or get so good at it that it becomes undetectable-possibly an in-demand skill for the future.

There are many ways to use LLMs.

The issue, IMO, is that some people throw in a one-shot, short prompt, and get a generic, boring output. "Garbage in, generic out."

Here's how I actually use LLMs:

- To dump my thoughts and get help organizing them.

- To get feedback on phrasing and transitions (I'm not a native speaker).

- To improve tone, style (while trying to keep it personal!), or just to simplify messy sentences.

- To identify issues, missing information, etc. in my text.

It’s usually an iterative process, and the combined prompt length ends up longer than the final result. And I incorporate the feedback manually.

So sure, if someone types "write a blog post about X" and hits go, the prompt is more interesting than the output. But when there are five rounds of edits and context, would you really rather read all the prompts and drafts instead of the final version?

(if you do: https://chatgpt.com/share/6817dd19-4604-800b-95ee-f2dd05add4...)

Ever since AI came out I’ve been talking about the prompt to output ratio. We naturally assume that the prompt will be smaller than the output just because of the particulars of the systems we use, but as you get more and more particular of what you want, the prompt grows while the output stays the same size. This is logical. If instead of writing an essay, I just describe what I want the essay to say, the description is necessarily gonna be a larger amount of text than the essay itself. It’s more text to describe what’s said, than to just say it. The fact that we expect to do less effort and get back more effort indicates exactly what we’re getting here: a bunch of filler.

In that way, the prompt is more interesting, and I can’t tell you how many times I’ve gone to go write a prompt because I dunno how to write what I wanna say, and then suddenly writing the prompt makes that shit clear to me.

In general, I’d say that AI is way more useful to compress complex ideas into simple ones than to expand simplistic ideas in to complex ones.

I fully support the author’s point but it’s hard to argue with the economics and hurdles around obtaining degrees. Most people do view obtaining a degree as just a hurdle to getting a decent job, that’s just the economics of it. And unfortunately the employers these days are encouraging this kind of copy/paste work. Look at how Meta and Google claim the majority of the new code written there is AI created?

The world will be consumed by AI.

That's because the instructor is asking questions that merely require the student to regurgitate the instructor's text.

To actually teach this, you do something like this:

"Here's a little dummy robot arm made out of Tinkertoys. There are three angular joints, a rotating base, a shoulder, and an elbow. Each one has a protractor so you can see the angle.

1. Figure out where the end of the arm will be based on those three angles. Those are Euler angles in action. This isn't too hard.

2. Figure out what the angles should be to touch a specific point on the table. For this robot geometry, there's a simple solution, for which look up "two link kinematics". You don't have to derive it, just be able to work out how to get the arm where you want it. Is the solution unambiguous? (Hint: there may be more than one solution, but not a large number.)

3. Extra credit. Add another link to the robot, a wrist. Now figure out what the angles should be to touch a specific point on the table. Three joints are a lot harder than two joints. There are infinitely many solutions. Look up "N-link kinematics". Come up with a simple solution that works, but don't try too hard to make it optimal. That's for the optimal controls course.

This will give some real understanding of the problems of doing this.

I used to teach, years before LLMs, and got lots of copy-pasted crap submitted. I always marked it zero, never mentioning plagiarism (which would require some university administration) and just commenting that I asked for X and instead got some pasted together nonsense.

As long as LLM output is what it is, there is little threat of it actually being competitive on assignments. If students are attentive enough to paraphrase it into their own voice I'd call it a win; if they just submit the crap that some data labeling outsourcer has RLHF'd into a LLM, I'd just mark it zero.

> I should hope that the purpose of a class writing exercise is not to create an artifact of text but force the student to think; a language model produces the former, not the latter.

It's been incredibly blackpilling seeing how many intelligent professionals and academics don't understand this, especially in education and academia.

They see work as the mere production of output, without ever thinking about how that work builds knowledge and skills and experience.

Students who know least of all and don't understand the purpose of writing or problem solving or the limitations of LLMs are currently wasting years of their lives letting LLMs pull them along as they cheat themselves out of an education, sometimes spending hundreds of thousands of dollars to let their brains atrophy only to get a piece of paper and face the real world where problems get massively more open-ended and LLMs massively decline in meeting the required quality of problem solving.

Anyone who actually struggles to solve problems and learn themselves is going to have massive advantages in the long term.

Call it the iron law of LLMs:

"No worthy use of an LLM involves other human beings reading its output."

If you use a model to generate code, let it be code nobody has to read: one-off scripts, demos, etc. If you want an LLM to prove a theorem, have it generate some Coq and then verify the proof mechanically. If you ask a model to write you a poem, enjoy the poem, and then graciously erase it.

Relatedly, there was a major controversy at work recently over the propriety of adding something like this to a lengthy email discussion:

> Since this is a long thread and we're including a wider audience, I thought I'd add Copilot's summary...

Someone called them out for it, several others defended it. It was brought up in one team's retro and the opinions were divided and very contentious, ranging from, "the summary helped make sure everyone had the same understanding and the person who did it was being conscientious" to "the summary was a pointless distraction and including it was an embarrassing admission of incompetence."

Some people wanted to adopt a practice of not posting summaries in the future but we couldn't agree and had to table it.

> They are invariably verbose, interminably waffly, and insipidly fixated on the bullet-points-with-bold style.

No, this is just the de-facto "house style" of ChatGPT / GPT models, in much the same way that that that particular Thomas Kinkade-like style is the de-facto "house style" of Stable Diffusion models.

You can very easily tell an LLM in your prompt to respond using a different style. (Or you can set it up to do so by telling it that it "is" or "is roleplaying" a specific type-of-person — e.g. an OP-ED writer for the New York Times, a textbook author, etc.)

People just don't ever bother to do this.

I just want to point out that AI generated material is naturally a confirmation bias machine. When the output is obviously AI, you confirm that you can easily spot AI output. When the output is human-level, you just pass through it without a second thought. There is almost no regular scenario where you are retroactively made aware something is AI.

> I’ll now cover the opposite case: my peers who see generative models as superior to their own output. I see this most often in professional communication, typically to produce fluff or fix the tone of their original prompts. Every single time, the model obscures the original meaning and adds layers of superfluous nonsense to even the simplest of ideas.

I'm going to call out what I see as the elephant in the room.

This is brand new technology and 99% of people are still pretty clueless at properly using it. This is completely normal and expected. It's like the early days of the personal computer. Or Geocities and <blink> tags and under construction images.

Even in those days, incredible things were already possible by those who knew how to achieve them. The end result didn't have to be blinking text and auto-playing music. But for 99% it was.

Similarly, with current LLMs, it's already more than possible to use them in effective ways, without obscuring meaning or adding superfluous nonsense. In ways whose results have none of the author's criticisms apply. People just don't know how to do it yet. Many never will, just like many never learnt how to actually use a PC past Word and Excel. But many others will learn.

> I have never seen any form of create generative model output (be that image, text, audio, or video) which I would rather see than the original prompt.

I've used LLM before to document command-line tools and APIs I've made; they aren't the final product since I also tweaked the writing and fixed misunderstandings from the LLM. I don't think the author would appreciate the original prompts, where I essentially just dump a lot of code and give instructions in bullet point form on what to output.

These generated documentation are immensely useful, and I use them all the time for myself. I prefer the documentation to reading the code because finding what I need at a glance is not trivial nor is remembering all the conditions, prerequisites, etc.

That being said, the article seems to focus on a use case where LLM is ill-suited. It's not suited for writing papers to pretend you wrote a paper.

> I say this because I believe that your original thoughts are far more interesting

Looking at the example posted, I'm not convinced that most people's original thoughts on gimbal lock will be more interesting than a succinct summary by an LLM.

Hate the game not the player. For the moment we continue to live in a world where the form and tone of communication matters and where foregoing the use of AI tools can put you at a disadvantage. There are countless homework assignments where teachers will give better grades to LLM outputs. An LLM can quickly generate targeted cover letters dramatically increasing efficiency while job hunting. Getting a paper accepted requires you to adhere to an academic writing style. LLMs can get you there. Maybe society just needs a few more years to adjust and shift expectations. In the meantime you should probably continue to use AI.

> Why do we write, anyway?

> I believe that the main reason a human should write is to communicate original thoughts.

in fairness to the students, how does the above apply to school work?

why does a student write, anyway? to pass an assignment, which has nothing to do with communicating original thoughts-- and whose fault is that, really?

education is a lot of paperwork to get certified in the hopes you'll get a job. it's as bereft of intelectual life as the civil service examinations in imperial china. original thought doesn't enter the frame.

LLM cheating detection is an interesting case of the toupee fallacy.

The most obvious ChatGPT cheating, like that mentioned in this article, is pretty easy to detect.

However, a decent cheater will quickly discover ways to conduce their LLM into producing text that is very difficult to detect.

I think if I was in the teaching profession I'd just leave, to be honest. The joy of reviewing student work will inevitably be ruined by this: there is 0 way of telling if the work is real or not, at which point why bother?

Is the author forgetting a baby in the bathwater he is throwing out? Especially on coding. He points out that vibe coding is bad, and then concluding that any program written through the use of an AI is bad.

For example if you already have a theory of your code, and you want to make some stuff that is verbose but trivial. It is just more efficient to explain the theory to an LLM and extract the code. I do like the idea of storing the underlying prompt in a comment.

Same for writing. If you truly copy paste output, it's obviously bad. But if you workshop a paragraph 5 or 6 times that can really get you unstuck.

Even the euler angles example. That output would be a good starting point for an investigation.

> I should hope that the purpose of a class writing exercise is not to create an artifact of text but force the student to think; a language model produces the former, not the latter.

This so much. A writing exercise sharpens your mind, it forces you to think clearly through problems, gives you practice in both letting your thoughts flow onto paper, and in post-editing those thoughts into a coherent structure that communicates better. You can throw it away afterwards, you'll still be a better writer and thinker than before the exercise.

Time to go back to writing essays in exams, live, on paper.

> I'd rather read the prompt

Yeah, to recycle a comment [0] from a few months back:

> Yeah, one of their most "effective" uses is to counterfeit signals that we have relied on--wisely or not--to estimate deeper practical truths. Stuff like "did this person invest some time into this" or "does this person have knowledge of a field" or "can they even think straight." [...]we might have to cope by saying stuff like: "Fuck it, personal essays and cover letters are meaningless now, just put down the raw bullet-points."

In other words, when the presentation means nothing, why bother?

[0] https://news.ycombinator.com/item?id=41675602

I’ve used ChatGPT as an editor and had very good results. I’ll write the whole thing myself and then feed it into ChatGPT for editing. And then review its output to manually decide which pieces I want to incorporate. The thoughts are my own, but sometimes ChatGPT is capable of finding more succinct ways of making the points.

While I agree with the thrust of the article being that students are cheating themselves by relying on LLMs, it's important to reflect on ways in which educators have encouraged this behavior. Anyone who has been to college in the age of the internet knows that many professors, particularly in the humanities, lazily pad out their class work with short menial writing assignments, often in the form of a "discussion board", that are rarely even graded on content. For students already swamped with work, or having to complete these assignments for general ed courses unrelated to their major/actual interests, it is totally understandable why they would outsource this work to a machine. This is a totally fixable issue: in-person discussions and longer writing assignments with well structured progress reports/check-ins and rounds of peer review are a couple ways that I can think of off the top of my head. Professors need to be held accountable for creating course loads that are actually intellectually interesting and are at least somewhat challenging to use LLMs to complete. When professors are constantly handing out an excess of low-effort assignments, using shortcuts becomes a learned behavior of students.

The level of cheating in college, pre-AI, is often overlooked in these articles.

Pre-AI, homework was often copied and then individuals just crammed for the tests.

AI is not the problem for these students, it's that many students are only in it for the diploma.

If it wasn't AI it would just be copying the assignment from a classmate or previous grad.

And I imagine the students who really want to learn are still learning because they didn't cheat then, and they aren't letting AI do the thinking for them now.

AI has changed how we learn by making the process of improving work much easier. Normally, learning involves writing a draft, finding mistakes, and fixing them over time. This helps build critical thinking. AI, trained on tons of refined data, can create polished work right away. While this seems helpful, it can skip the important step of learning through trial and error.

The question is: Should we limit AI to keep the old way of learning, or use AI to make the process better? Instead of fixing small errors like grammar, students can focus on bigger ideas like making arguments clearer or connecting with readers. We need to teach students to use AI for deeper thinking by asking better questions.

We need to teach students that asking the right questions is key. By teaching students to question well, we can help them use AI to improve their work in smarter ways. The goal isn’t to go back to old methods for iterating but change how we iterate altogether.

> [Not a student’s real answer, but my handmade synthesis of the style and content of many answers]

> You only have to read one or two of these answers to know exactly what’s up: the students just copy-pasted the output from a large language model, most likely ChatGPT. They are invariably

This is validating. Your imitation completely fooled me (I thought it really was ChatGPT and expected to be told as much in an entirely unsurprising "reveal") and the subsequent description of the style is very much in agreement with how I'd characterize it.

In previous discussions here, people have tried to convince me that I can't actually notice these obvious signs, or that I'm not justified in detecting LLM output this way. Well, it may be the case that all these quirks derive from the definitely-human training data in some way, but that really doesn't make them Turing-test-passing. I can remember a few times that other people showed me LLM prose they thought was very impressive and I was... very much not impressed.

> When someone comments under a Reddit post with a computer-generated summary of the original text, I honestly believe that everyone in the world would be better off had they not done so. Either the article is so vapid that a summary provides all of its value, in which case, it does not merit the engagement of a comment, or it demands a real reading by a real human for comprehension, in which case the summary is pointless. In essence, writing such a comment wastes everyone’s time.

I think you've overlooked some meta-level value here. By supplying such a comment, one signals that the article is vapid to other readers who might otherwise have to waste time reading a considerable part of the article to come to that conclusion. But while it isn't as direct as saying "this article is utterly vapid", it's more socially acceptable, and also more credible than a bald assertion.

I will often write a bunch of stuff and then use an LLM to pre-process it a little bit and suggest some improvements. I will then work through the differences and consider them individually and either accept them, or use them to write my own improvements. This is kind of like having an okay editor working for you. No substitute for a real editor, but it means that a: what I intended to say is preserved, b: there's no additional waffle (the prompt includes instructions not to expand on any topic, but only ever to summarise where possible), c: everything still goes by me in the end, and if it doesn't feel like something I would actually write then it doesn't get used.

I believe that it has improved my writing productivity somewhat, especially when I'm tired and not completely on the ball. Although I don't usually reach for this most of the time (e.g. not for this comment).

The problem isn't LLM, it's how universities are designed. With short terms and high pressure, students develop 'knowledge bulimia' (in lack of a better term). They have to study highly complex fields in short amounts of time, then move on to often unrelated fields quickly thereafter with no emphasize on persistent learning: the knowledge learned in previous exam can be mostly discarded. They may need to 're-learn' it for another exam, but that's fine, they are very good at learning new things which later on can get discarded.

Using LLMs to achieve this is just another step in the evolution of a broken education system. The fix? IMO, make the exams for the courses delayed by one semester. So during the exam study-period, the students have to 'catch up' on the lectures they had a few months ago.

An exception to test the rule with: people are generating lifelike video based on the pixel graphics from old video games. I have no interest in seeing a prompt that says "Show me a creature from Heroes of Might and Magic 3, with influences from such and so", but it's incredible to see the monsters I've spent so much time with coming to life. https://www.youtube.com/watch?v=EcITgZgN8nw&lc=UgxrBrdz4BdEE...

Maybe the problem is that the professor doesn't want to read the student work anyway, since it's all stuff he already knows. If they managed to use their prompts to generate interesting things, he'd stop wanting to see the prompts.

None of the people writing these sorts of posts seem willing to acknowledge how prevalent not doing your own work was before AI wad around.

The hardest hit industry by AI has been essay writing services.

If anything, it seems they're noticing because the AI is doing a worse job.

> A typical belief among students is that classes are a series of hurdles to be overcome; at the end of this obstacle course, they shall receive a degree as testament to their completion of these assignments.

I agree with the broader point of the article in principle. We should be writing to edify ourselves and take education seriously because of how deep interaction with the subject matter will transform us.

But in reality, the mindset the author cites is more common. Most accounting majors probably don't have a deep passion for GAAP, but they believe accounting degrees get good jobs.

And when your degree is utilitarian like that, it just becomes a problem of minimizing time spent to obtain the reward.

This is equivalent to students using AI to complete computer programming assignments. They misconstrue the purpose of an assignment as just one of generating output instead something to teach the principles and techniques they'll require later if they want a job in the profession. While they may believe they're fooling the teacher, all they're really doing is fooling, and cheating, themselves.

Whether it be writing or computer programming, or exercising, for that matter, if you aren't willing to put in the work to achieve your goals, why bother?

I've found really saddening to see students submit written-by-ChatGPT arguments to the department council when their university spot was on the line (for failing grades). This was their ultimate chance to prove their worth and they left it to ChatGPT.

At first, I thought they didn't care. However, it was so pervasive that it couldn't be the only explanation. I was forced to conclude they trusted ChatGPT more than themselves to argue their case... (Some students did not care, obviously.)

I think AI can be an amazing tool that can help us learn even better when used correctly and when not used as a substitute for learning and understanding.

It can be used as a personal tutor. How awesome is it to have a tutor always available to answer almost any question from any angle to really help you understand? Yes, AI won't get everything right 100%, but for students who are still learning basics, it's fair to assume that having an AI tutor can yield far better results than having no tutor at all.

It can also be used as a tool for doing mundane work, so you can focus more on the interesting and creative work. Kind of like a calculator or a spreadsheet. Would math majors become better mathematicians if they had to do all calculations by hand?

I think instead of banning AI, education needs to reform. Teaching staff should focus less time on giving lectures and grading papers (those things can be recorded and automated) and more time on ORAL EXAMS where they really probe student's knowledge and there's no possibility of cheating.

Students can and should use AI to help them prepare. E.g. don't ask AI to write an essay for you, write it yourself and ask it to critique it. Don't ask it to give you answers for a test, ask it to ask you questions on the topic and find gaps in your knowledge. Etc.

I wish the author had state out right that they were not using LLMs much, since their opinion on them and their output has no value (its a new technology, and different enough that you do have to spend some time with them in order to be able to find out what value they have for your particluar work[0].

The is especially the case when you are about to complain about style, since that can easily be adjusted, by simply telling the model what you want.

But I think there is a final point that the author is also wrong about, but that is far more interesting: why we write. Personally I write for 3 reasons: to remember, to share and to structure my thoughts.

If an LLM is better then me at writing (and it is) then there is no reason for me to write to communicate - it is not only slower, it is counterproductive.

If the AI is better at wrangling my ideas into some coherent thread, then there is no reason for me to do it. This one I am least convinced about.

AI is already much better than me at strictly remembering, but computers have been that since forever, the issue is mostly convinient input/output. AIs makes this easier thanks to speech to text input.

[0]: See eg. https://www.oneusefulthing.org/p/centaurs-and-cyborgs-on-the....

> [...] but not so distinctive to be worth passing along to an honor council. Even if I did, I’m not sure the marginal gains in the integrity of the class would be worth the hours spent litigating the issue.

The school should be drilling into students, at orientation, what some school-wide hard rules are regarding AI.

One of the hard rules is probably that you have to write your own text and code, never copy&paste. (And on occasions when copy&paste is appropriate, like in a quote, or to reuse an off-the-shelf function, it's always cited/credited clearly and unambiguously.)

And no instructors should be contradicting those hard rules.

(That one instructor who tells the class on the first day, "I don't care if you copy&paste from AI for your assignments, as if it's your own work; that just means you went through the learning exercise of interacting with AI, which is what I care about"... is confusing the students, for all their other classes.)

Much of society is telling students that everything is BS, and that their job is to churn BS to get what they want. Early "AI' usage popular practices so far looks to be accelerating that. Schools should be dropping a brick wall in front of that. Well, a padded wall, for the students who can still be saved.

> Lately I’ve seen more people in their cars thwarting stoplight boredom—­that is, unable to sit unmediated for even the few moments that it takes a red light to turn green, they reach for their smartphones.

I wish it was only at stoplights. But then just a few days ago, I witnessed a totally unnecessary accident. Left-lane got green, and someone in the straight lane noticed the movement but didn't look up and drove right into the car in front of them...

The most brain intensive activity I ever did at school, happened in the very last two years (so, ~20 years at school lead to it), when I rubbed my brain to Lambda Prolog.

I almost had headaches after intense thinking of problems and ways to solve them in lambda prolog. That was the most interesting and satisfying to physically feel the effect of high focus combined with applying what was a new logic.

Computer science at the university, taught me how to learn and explore new ideas. I might sound like my grandpa who told me when I was 8yo that using calculator would lead to people not able to count... and here I am saying that LLM might lead to people who do not know how to write.

Actually, I am a bit concern that we might produce more text in the short term because it is becoming cheap to write tons of documentation with LLMs. But those feel like death by Terms and Conditions, i.e. text that no one reads. So not only we would lose our ability to write, but we can seriously affect our ability to read. Sure LLM can summarize as well, but then we lose the nuances.

Nature is lazy, but should we be lazy and delegate our ability to think (read/write), to a software ? Think about it :)