Why ask forum members, when you can ask AI?

[MarkRhodes]

I find it useful.
I write a newsletter for the condo where my wife and I live, and I needed to mention something specific about a roofing project that will begin in April of 2025. (The budget was recently approved, and the roofing project was one item.)
I asked Grok about this because I did not understand what the term meant, much less how to explain it to someone else.
Grok told me.
Had I googled the query, I would have received many hits, the top ones being sponsored and weighted toward people selling something. (If you click on those, you must scroll through ads to reach the info you want.)
This was faster, free of distractions.
It has its uses, and the tech will doubtless improve rapidly.

[omphalopsychos]

CliffR, I appreciate the follow-up and your observations about the code. Let’s address the points you raised, including the errors in the snippet, and I’ll also take the opportunity to clarify some misunderstandings about the value of AI in this context.


On C vs. SQL


I see where you’re coming from about C being “richer” than SQL, but I think the comparison is misplaced. While C’s flexibility can indeed hide subtle errors (e.g., memory access issues or infinite loops), SQL presents its own challenges—especially for AI. Writing SQL queries often requires an understanding of data structures, relationships, and schemas, which involves context-specific reasoning. By contrast, the n-queens problem is a well-documented, canonical challenge with clearly defined rules.


For a human, the C vs. SQL comparison might feel significant because of their differing levels of abstraction, but for AI, both tasks boil down to processing patterns and rules. If anything, SQL might pose a harder challenge for AI because of the need to interpret data-specific context, whereas solving n-queens is largely syntactic. That said, I think this debate distracts from the real point.


Errors in the Code


After reviewing the snippet, I identified two critical issues:
1. Premature return true; in Recursive Function:
The code halts further exploration after finding a single solution, which violates the problem’s goal of finding all solutions. Removing this ensures the loop continues to explore other possibilities.
2. Static Memory Allocation:
Using a fixed-size array (e.g., MAX_SIZE) wastes memory and limits scalability. A dynamic allocation approach using malloc/calloc is more efficient and allows for larger values, limited only by available memory.


Improved Code


Here’s the improved code with both errors resolved, optimized memory usage, and clearer modularity:


#include<stdio.h>
#include<stdlib.h>
#include<stdbool.h>


int solutions = 0;
int** board;
int n;


// Function to allocate memory for the board
void allocateBoard() {
board = (int**)malloc(n * sizeof(int*));
for (int i = 0; i < n; i++) {
board[i] = (int*)calloc(n, sizeof(int));
}
}


// Function to free allocated memory
void freeBoard() {
for (int i = 0; i < n; i++) {
free(board[i]);
}
free(board);
}


// Function to print the board
void printBoard() {
for (int i = 0; i < n; i++) {
for (int j = 0; j < n; j++) {
printf("%d ", board[i][j]);
}
printf("\n");
}
printf("\n");
}


// Function to check if a queen can be placed at board[row][col]
bool isSafe(int row, int col) {
for (int i = 0; i < row; i++) {
if (board[i][col] == 1) return false;
}
for (int i = row, j = col; i >= 0 && j >= 0; i--, j--) {
if (board[i][j] == 1) return false;
}
for (int i = row, j = col; i >= 0 && j < n; i--, j++) {
if (board[i][j] == 1) return false;
}
return true;
}


// Recursive function to solve the N-Queens problem
void solveNQueens(int row) {
if (row == n) {
solutions++;
printBoard();
return;
}
for (int col = 0; col < n; col++) {
if (isSafe(row, col)) {
board[row][col] = 1;
solveNQueens(row + 1);
board[row][col] = 0; // Backtrack
}
}
}


int main() {
printf("Enter the number of queens (n): ");
if (scanf("%d", &n) != 1 || n <= 0 || n > 20) {
printf("Invalid input. Please enter a number between 1 and 20.\n");
return 1;
}


allocateBoard();
solveNQueens(0);
printf("Total solutions: %d\n", solutions);
freeBoard();
return 0;
}


Key Improvements:
• Removed return true; to allow continued exploration of solutions.
• Dynamically allocated the board to optimize memory usage and scalability.
• Modularized the memory management logic for clarity.


Reframing the Discussion


Your question about how long it would take a human to spot and fix the bugs is valid, but I think it misses the larger point. AI isn’t meant to replace human expertise—it’s a tool to augment it. For tasks like this, AI accelerates the scaffolding process, delivering a functional draft in seconds. Humans can then refine it, addressing subtle issues like the ones you highlighted.


This isn’t about AI vs. humans or which language is “harder.” It’s about leveraging AI to achieve incremental gains. In aggregate, these gains—whether in debugging, scaffolding, or iterating—can result in far more efficient code production at scale. That’s where the real value lies: not in replacing human insight but in empowering it.

[GuyBoden]

My question would be:

Please find the sequence for Prime Numbers.

[omphalopsychos]

BTW if you feel I missed the specific bugs you had in mind, that’s fair—but it highlights the larger point. With the AI-generated draft, you were able to quickly review, identify issues, and engage in improving the solution. That ease of iteration is exactly what makes AI valuable.


The purpose isn’t to generate flawless code but to accelerate the process. By handling the structure and basic logic, AI gives us a starting point we can refine, saving significant time compared to writing everything from scratch. The n-queens example is a great case: the AI produced a functional draft in seconds, and from there, the effort shifted to debugging and enhancing.

[CliffR]

I understand the sentiment, but I just don't think we're there yet. I tried telling Gemini there were bugs. It's taken three attempts now to find and fix them. It correctly identified the problem was related to the return in the recursive function, just as you've identified above. However, none of its proposed solutions were correct. The code above appears to be correct, but not for the reason you (or the LLM) state - just returning without the bool doesn't affect the search space. But ignoring the return result in the calling function means it correctly reports the number of solutions found.

[omphalopsychos]

Try Chat GPT o1. There are huge performance differences among the different ones. If you're using a free AI tool, you're using tech from 2 years ago. Also if you can't get the AI to fix your two bugs, I'm confident this is a user error. Would love to see your prompts since it should be easy to say. "Bug 1 = x, Bug 2 = y, fix it."

[CliffR]

Try Chat GPT o1. There are huge performance differences among the different ones. If you're using a free AI tool, you're using tech from 2 years ago. Also if you can't get the AI to fix your two bugs, I'm confident this is a user error. Would love to see your prompts since it should be easy to say. "Bug 1 = x, Bug 2 = y, fix it."

Bug 2 was a figment of my imagination . Bug 1 it found, but failed to fix for several iterations. It did eventually get there after much prompting. I mean, I don't want to be in the position as a professional programmer of letting the AI do the fun bit of putting a solution together and then me do the hard bit of working through the bugs. This will only get harder and more tedious with more complex, real-world problems.

[omphalopsychos]

Was the bug the premature return?

Yeah it sounds like your review was slowed down by your two imagined bugs, but put a senior pro software in the loop and you’ve got a really tight process.

[CliffR]

Was the bug the premature return?

Yeah it sounds like your review was slowed down by your two imagined bugs, but put a senior pro software in the loop and you’ve got a really tight process.

Lol - I am a senior pro. I've been writing code commercially since 1980. There never was a premature return. The bug was the return 'false' at the end of the recursive function. This means all invocations would be seen as failures and the code would assume 0 solutions.

[omphalopsychos]

By senior I meant someone with 4-6 years of experience. They tend to be the most adaptable and versed in modern stacks. For someone of your experience level I’d say principal or retired. I think it makes sense though, usually we wouldn’t ask a principal engineer to do code reviews for small scripts since they’re usually working at a much higher level of abstraction, and less experienced reviewers can spot them quickly.

Anyway it seems like an easy fix, I can also tell you have a prejudice that seemed to muddy up your interactions. Just be glad you’re either already retired or close to it before this new tech affects you too directly.

[Litterick]

Right now, if writing any kind of paper, thesis or project, ChatGpt can turn the time required from weeks to days, or from a week to a weekend.

I see such papers often. They are always banal, contributing nothing new because no thinking took place. Research in the humanities is done by human minds, not algorithms.

[omphalopsychos]

I see such papers often. They are always banal, contributing nothing new because no thinking took place. Research in the humanities is done by human minds, not algorithms.

The key here is how AI is used. If someone relies on AI to write a paper without doing the thinking or research, the result will be shallow, as you said. But I believe Alter is suggesting a more thoughtful approach, where AI complements rather than replaces human effort.

Effective use of AI involves doing the intellectual work—researching, crafting arguments, and structuring ideas—while letting AI handle the mechanical tasks like drafting, rephrasing, or improving clarity. This saves time on writing without compromising the depth or originality of the work.

The issue isn’t the tool itself but how it’s used. Misuse leads to superficial results, but when used properly, AI enhances efficiency while leaving the critical thinking firmly in human hands. That’s the distinction I think is worth emphasizing.

[palindrome]

The key here is how AI is used. If someone relies on AI to write a paper without doing the thinking or research, the result will be shallow, as you said. But I believe Alter is suggesting a more thoughtful approach, where AI complements rather than replaces human effort.

Effective use of AI involves doing the intellectual work—researching, crafting arguments, and structuring ideas—while letting AI handle the mechanical tasks like drafting, rephrasing, or improving clarity. This saves time on writing without compromising the depth or originality of the work.

The issue isn’t the tool itself but how it’s used. Misuse leads to superficial results, but when used properly, AI enhances efficiency while leaving the critical thinking firmly in human hands. That’s the distinction I think is worth emphasizing.

I think I can agree with your general statement that intelligent interaction with those systems can speed up workflows, but I would definitely disagree with some of your more specific statements.

Rephrasing, and improving clarity in particular, are not "mechanical" tasks in my book. According to the Chicago Manual, both fall into the category of "substantive" as opposed to "mechanical" editing, these being distinct activities with very different mindsets at work.

And "improving clarity" is clearly tantamount to actual thinking. My experience with heavily AI-influenced texts has been that, aside from their logical inconsistencies, they tend to be all over the place even at the terminological level, which has a way of confusing things. Imagine an author who ends an article not having the slightest idea what was said in the first sentence or even two paragraphs previously.

I may be lagging behind, OK, but from all I have seen so far, once you *really* start plowing through these texts at a deeper level (and assuming that the messages to be conveyed are relatively complex), they do not hold up to scrutiny, and their somewhat "nice" appearances on the face of it have a way of costing thorough readers an awful lot of time, given a need to compensate for any lack of primary thinking by secondary thinking. Robbing Peter to pay Paul, so to speak. Anyway, I am convinced that much of the praise for LLMs is based on superficial reading.

[aleksandar]

First of all, I want to emphasize that my intention for writing this comment is far, far from being rude in whatsoever way. So to answer your question, well, look at your results! It's like generic add shit you'd read anywhere on the Internet, and that's what AI at the current state of affairs is, a cumulative, generic sum of experiences that PEOPLE have shared across the web. Now, if I'd post a question here, for example, what would you choose between a Tele and a 335, I'd like to read what the average Joe has to say from his personal point of view, because I am an average Joe myself, just with a different flavor, so to speak, and with a different name. In other words, I don't want random generic blabbering, but a personal, direct and real communication.

[omphalopsychos]

I think I can agree with your general statement that intelligent interaction with those systems can speed up workflows, but I would definitely disagree with some of your more specific statements.

Rephrasing, and improving clarity in particular, are not "mechanical" tasks in my book. According to the Chicago Manual, both fall into the category of "substantive" as opposed to "mechanical" editing, these being distinct activities with very different mindsets at work.

And "improving clarity" is clearly tantamount to actual thinking. My experience with heavily AI-influenced texts has been that, aside from their logical inconsistencies, they tend to be all over the place even at the terminological level, which has a way of confusing things. Imagine an author who ends an article not having the slightest idea what was said in the first sentence or even two paragraphs previously.

I may be lagging behind, OK, but from all I have seen so far, once you *really* start plowing through these texts at a deeper level (and assuming that the messages to be conveyed are relatively complex), they do not hold up to scrutiny, and their somewhat "nice" appearances on the face of it have a way of costing thorough readers an awful lot of time, given a need to compensate for any lack of primary thinking by secondary thinking. Robbing Peter to pay Paul, so to speak. Anyway, I am convinced that much of the praise for LLMs is based on superficial reading.

Thanks for your thoughtful critique. I think your points highlight some common concerns about AI, but there are a few misunderstandings worth addressing.

On your example of “an author who ends an article not having the slightest idea what was said in the first sentence or two paragraphs previously,” this reflects a misunderstanding of how large language models (LLMs) like GPT-4 work. LLMs retain context within their prompt’s scope, enabling them to reference and build upon earlier content. They don’t “forget” previous sentences unless the context window is exceeded. If coherence issues arise, they are typically due to unclear input or poor structuring by the user, not an inherent limitation of the model.

As for your broader skepticism, it might stem from older models or less advanced systems. For example, while Google’s Gemini produced buggy C code for CliffR, GPT-4 generated accurate solutions in seconds, even improving modularity and memory management. This is all while CliffR, with 40 years of experience spent 30 minutes determining whether there were three, two, or one bugs in the same code snippet. This contrast highlights how AI can achieve rapid, coherent, and high-quality results when used effectively.

Lastly, while the Chicago Manual draws a distinction between “mechanical” and “substantive” editing, it’s worth noting that this isn’t some absolute rule. These categories are human-made and don’t necessarily reflect how tasks should be divided between people and machines, especially as AI continues to improve. Machines today show the ability to achieve complex reasoning in specific domains, but this depends heavily on the availability of high-quality training data. AI excels in well-defined areas, like rephrasing and simplifying language, while lagging in tasks that require original synthesis or reasoning across unfamiliar domains. The distinction isn’t static—it shifts as AI capabilities grow, which is why it’s important to evaluate the tools based on what they’re good at now, not fixed assumptions.

[sgosnell]

AI only knows what it has read on the internet. Not everything on the internet is true, contrary to popular belief. The programming term long ago was GIGO.



Garbage In, Garbage Out.



Lots of garbage on the interwebz, and more every second.

[omphalopsychos]

AI only knows what it has read on the internet. Not everything on the internet is true, contrary to popular belief. The programming term long ago was GIGO.



Garbage In, Garbage Out.



Lots of garbage on the interwebz, and more every second.

You’re absolutely right that the internet contains a mix of valuable information and unreliable content, and the principle of “garbage in, garbage out” applies to AI as much as it does to humans. However, it’s important to recognize that the internet also contains an enormous amount of verified, high-quality information, including academic publications, textbooks, and other reliable sources. Models like GPT are trained on a vast range of data, and this diversity allows them to produce accurate and meaningful outputs in many domains.

It’s also worth noting that future cycles of AI research are increasingly focused on improving data quality by separating valuable content from noise. This is already happening to some extent: companies are exploring curated datasets and partnerships to refine their models. For example, Zuckerberg reportedly wanted to acquire Penguin Books to access its corpus of curated, high-quality texts. These efforts aim to ensure models are trained on robust, authoritative data rather than relying on the general messiness of the internet.

While the internet’s imperfections are real, AI’s ability to process and synthesize vast amounts of data—coupled with human validation—makes it a powerful tool. The key is thoughtful application and continued improvement in how we curate and refine the datasets we use. This is a challenge, but also an opportunity for AI to evolve into something even more reliable and transformative.

[omphalopsychos]

Anyway as much as I've enjoyed this discussion, this debate will soon be moot. Many criticisms here target earlier-generation models, which I agree had significant flaws. But what’s being ignored are (1) the observable results where AI is already producing high-quality work and (2) the widespread adoption of AI tools that are enhancing productivity across industries. Of course, there are risks and a need for validation and testing, but even with those layers, the efficiency gains are massive. It’s reminiscent of the Industrial Revolution, where guild masters resisted mechanization, claiming it couldn’t produce quality—only to watch mechanized production transform the world. Similarly, companies and individuals who avoid adopting AI will likely be left behind, as their slow and expensive workflows can’t compete. The outcomes will speak for themselves, and in two years, I’d encourage anyone skeptical to revisit this discussion and see if their perspective has changed.

[dixiehwy25]

Well, I asked Google since I get some summaries attributed to AI often when I am doing a search.

What I got was a list of drop down potential search topics like- all time best jazz guitar, or price(s), or used versus new, etc. Old fashioned, but I think that's a lot more useful...

[sgosnell]

I completely agree that there are both reliable and verified data, and garbage, on the internet. The really hard part is determining which is which. Should I believe everything Q or RJK Jr has supposedly posted, or a peer-reviewed study on the effectiveness of vaccines? How would I know? How would AI know?

TBH, I know what I believe, but I'm not totally certain how or why. And I know that I'm not yet ready to blindly accept anything given to me by any AI bot. Predicting things is hard, especially about the future, so I'm also not ready to completely ignore AI. Ask me again in a few years.

[palindrome]

AI excels in well-defined areas, like rephrasing and simplifying language.

No time for a more elaborate response, so let me just point out that, in your previous post, you said "improving clarity" rather than "simplifying language."

Now, more simpler isn't invariably more clearer, as they say.

[palindrome]

One example, just because I cannot resist:

Here we have "Thousands of AI Authors on the Future of AI":

https://aiimpacts.org/wp-content/upl...2022ESPAIV.pdf

List on Page 5, second item: "Translate text in newfound language"

Heck, what does this even MEAN? Of course I found out (and remain doubtful), but shouldn't "thousands of AI authors" with all their intellectual capacities, whether natural or artificial, have managed to come up with a meaningful sentence at this prominent position in the paper?

See?

[Litterick]

Effective use of AI involves doing the intellectual work—researching, crafting arguments, and structuring ideas—while letting AI handle the mechanical tasks like drafting, rephrasing, or improving clarity. This saves time on writing without compromising the depth or originality of the work.

That kind of thinking keeps me in business. PhD candidates know their subjects, but they do not know how to write. Some do not know English very well. They want to save time writing. They rely on AI to handle the 'mechanical tasks' of expressing their ideas. AI lets them down. Their supervisors come to me to make their dissertations work.

[CliffR]

For example, while Google’s Gemini produced buggy C code for CliffR, GPT-4 generated accurate solutions in seconds, even improving modularity and memory management. This is all while CliffR, with 40 years of experience spent 30 minutes determining whether there were three, two, or one bugs in the same code snippet. This contrast highlights how AI can achieve rapid, coherent, and high-quality results when used effectively.

The irony being I asked *you* to find the bugs. Instead, you asked ChatGTP to do it, and it mis-identified the real bug. Then you repeated that misidentification to me - you were taken in by the false information ChatGTP provided to you.

(Also worth pointing out that the use of global variables and the redundant searching board squares make this far from a 'high quality' result.)

[RJVB]

I asked Grok about this because I did not understand what the term meant, much less how to explain it to someone else.
Grok told me.

I got curious and found out I don't grok that AI at all, imagining how it is supposedly modelled after HGG
and undoubtedly permeated with musky and other tweetxy "ideas".

Funny, years ago, at the first From Animals To Animats conference someone coined the term neuromancy. Not as something for use in hobby fantasy novel writing, but to describe the fact that you could take a big enough neural network and get it do almost any kind of input/output mapping via the right amount of backprop training. Back then we could probably only dream of current day LLMs and maybe not even that about the possibility to run them like we can now.

Anyway, I am convinced that much of the praise for LLMs is based on superficial reading.

Isn't the main target audience for the technology also the main target audience of Reader's Digest, people who'll rarely ever read something more profound than what can be found in that catalogue?