LLMs

Started by urge
This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.
Get help with your application

Use all the free resources available to you from SDN: articles, guides, expert advising, forums discussions, and school research.

Browse articles Compare schools My application plan Read Interview Feedback
Practice for interviews Ask an Advisor
urge

urge

Full Member
15+ Year Member
Advertisement - Members don't see this ad
What use have you found for LLMs?

1 Internet search, where you just want a summarized answer rather than 10 bloated websites.
2 Writing the occasional letter (recommendation, resignation..)
3 Bouncing ideas/entertainment.

Anything else?

Anything of real value? Anything worth billions of dollars for the data canters?
 
Advertisement - Members don't see this ad
The current actual use case is software engineering and development. Outside of that the use is limited at present but it is kind of interesting that the frontier models seem really conspicuously close to recursive self improvement in software and algorithm design… and that one model has brute forced solutions to decades or centuries old unsolved math problems
 
I guess I'm in the minority, as I find it extremely useful for almost everything. "Help me pick an outfit for (whatever occasion)." "Give me some ideas for pictures on this wall, rendering an edit of this picture." I'm usually a lot more specific in follow-up. I use it a lot, so it knows me well and curates answers based on what it knows about me. I ask it questions about politics and philosophy disagreements, asking for arguments for and against positions, as well as evidence and examples supporting each. For work, I ask about anesthesia considerations for cases to make sure I'm not missing anything (often times I dismiss its advice, but I appreciate the thoughts). I've even had it teach me entire subjects, structured like a college course with suggested reading and such. It's helped me with financial literacy, how to fix things around the house, and perfecting my morning espresso and steamed milk art. As you can tell, it's pretty random, but I use it all the time.
 
the frontier models are *incredible* as tutors to teach well documented subjects up to the post graduate level . It’s like having a pocket professor that never gets tired or irritated and if you know how to prompt it will teach you in any style you wish and do an excellent job (Socratic, QA, etc).

I’ve been taking actual courses and using the models to teach myself software engineering and they are incredible teachers
 
T-burglar said:
the frontier models are *incredible* as tutors to teach well documented subjects up to the post graduate level . It’s like having a pocket professor that never gets tired or irritated and if you know how to prompt it will teach you in any style you wish and do an excellent job (Socratic, QA, etc).

I’ve been taking actual courses and using the models to teach myself software engineering and they are incredible teachers
Click to expand...
I agree they are better teachers than any person I encountered in my training, if you are interested. They go at your pace. No judgement. Their breath of knowledge is bigger. Provides ramifications of thought.
 
T-burglar said:
The current actual use case is software engineering and development. Outside of that the use is limited at present but it is kind of interesting that the frontier models seem really conspicuously close to recursive self improvement in software and algorithm design… and that one model has brute forced solutions to decades or centuries old unsolved math problems
Click to expand...

Most of the software engineers I know see AI generated code (for non-trivial projects) as spaghetti that may function OK-ish now but is all but guaranteed to be un-maintainable, un-expandable, un-updatable in the future.

It's going to create a lot of work for human software engineers to redo and/or un**** the AI-generated slop when that day comes.

The believers say Well then the AI will be good enough to fix that, when that day comes.

I'm agnostic about that.

But I will say, as a guy who 30 years ago in my pre-medicine life was (almost) making a living writing and selling code, the hard part, and the part that separated amateurs from pros, wasn't writing some code that worked. It was writing code that someone else could understand, modify, expand, and make work a year or more later when it needed to be modified for a new or edge use case .... without rewriting the whole thing from scratch.

And current AI generated code is not good enough for that task. I understand that the argument is that AI will get better at that maintainance role, and it probably will. I guess. Or maybe the argument is that future AI tools will be so cheap and easy that a full rewrite from scratch will be simple, and that maintainance will be an archaic concept.

My bet is that's not going to be true. I could be wrong.



The only thing I really find AI useful for right now is web search summaries. It's gotten pretty good. Accurate most of the time. At worst it gives me references to track down primary type sources to verify what it's telling me. I do wonder what'll happen though when those primary sources get supplanted by other AI generated answers, and we end up with a giant circle jerk of AI confabulated fantasies, with no simple way to verify the truth or falsehood of an answer.
 
I’ve seen YouTube videos by SWE claiming the outputs are bad but when you look at their prompts it’s obvious why their outputs are bad. The skill required to get good code out is excellent English composition. Which many SWE seem to lack , perhaps unsurprisingly
 
For example, if you take the time to really learn to use Claude code properly - the way to use it is to spend hours or even days in a chat thread with an open word document creating a detailed markdown file in verbose English which serves as the instruction set for CC. Hours or days just bouncing ideas around to come up with a feature branch in plain English, not writing any code. Then when you are satisfied with the instruction set you actually set the agent to task.

The people producing spaghetti code don’t understand that the stochastic parrot will do EXACTLY what you tell it to do, but will fill in any blanks or holes in the instructions with the mathematically most probable next token. If you leave no holes in your instructions, you get what what you want. So you have to be verbose, specific, and detailed.

Obviously, you have to learn programming to a certain level in order to know what you want in your code, but it definitely changes the game entirely when you can architect a solid new feature branch in natural language instead of tedious painstaking manual code writing.
 
Last edited:
I minored in philosophy so I find this new paradigm of programming really great. If you don’t like to write essays you may hate it
 
T-burglar said:
For example, if you take the time to really learn to use Claude code properly - the way to use it is to spend hours or even days in a chat thread with an open word document creating a detailed markdown file in verbose English which serves as the instruction set for CC. Hours or days just bouncing ideas around to come up with a feature branch in plain English, not writing any code. Then when you are satisfied with the instruction set you actually set the agent to task.

The people producing spaghetti code don’t understand that the stochastic parrot will do EXACTLY what you tell it to do, but will fill in any blanks or holes in the instructions with the mathematically most probable next token. If you leave no holes in your instructions, you get what what you want. So you have to be verbose, specific, and detailed.

Obviously, you have to learn programming to a certain level in order to know what you want in your code, but it definitely changes the game entirely when you can architect a solid new feature branch in natural language instead of tedious painstaking manual code writing.
Click to expand...
As these LLMs get more advanced, there's an argument to be made that having very specific, detailed steps can constrain the LLM a bit too much. What matters more is highly specific context and specifying what a good result is. Overall, I agree with you. These LLMs are only useful to the extent that the user is using an actual version that is advanced enough and knows how to prompt it correctly. There's definitely a learning curve.
 
T-burglar said:
For example, if you take the time to really learn to use Claude code properly - the way to use it is to spend hours or even days in a chat thread with an open word document creating a detailed markdown file in verbose English which serves as the instruction set for CC. Hours or days just bouncing ideas around to come up with a feature branch in plain English, not writing any code. Then when you are satisfied with the instruction set you actually set the agent to task.

The people producing spaghetti code don’t understand that the stochastic parrot will do EXACTLY what you tell it to do, but will fill in any blanks or holes in the instructions with the mathematically most probable next token. If you leave no holes in your instructions, you get what what you want. So you have to be verbose, specific, and detailed.

Obviously, you have to learn programming to a certain level in order to know what you want in your code, but it definitely changes the game entirely when you can architect a solid new feature branch in natural language instead of tedious painstaking manual code writing.
Click to expand...
Basically AI is another compiler in the coding process in your example. Only thing different from what was before is that the person doesn’t need to learn a more complicated syntax to follow. Instead, they can theoretically use their native language.
 
Last edited:
T-burglar said:
For example, if you take the time to really learn to use Claude code properly - the way to use it is to spend hours or even days in a chat thread with an open word document creating a detailed markdown file in verbose English which serves as the instruction set for CC. Hours or days just bouncing ideas around to come up with a feature branch in plain English, not writing any code. Then when you are satisfied with the instruction set you actually set the agent to task.

The people producing spaghetti code don’t understand that the stochastic parrot will do EXACTLY what you tell it to do, but will fill in any blanks or holes in the instructions with the mathematically most probable next token. If you leave no holes in your instructions, you get what what you want. So you have to be verbose, specific, and detailed.

Obviously, you have to learn programming to a certain level in order to know what you want in your code, but it definitely changes the game entirely when you can architect a solid new feature branch in natural language instead of tedious painstaking manual code writing.
Click to expand...

So what you're saying is, to use these tools effectively, you need a firm grasp of data structures, algorithms, user interfaces, security, and project management, as well as a nuanced feel for the precise language used to communicate your vision to the IDE, er, I mean Claude.

...

Is that not what writing a program in a traditional language is?

🙂

Programming (competently) is 80% planning - designing and specifying what does when and where and how - and about 5% typing code, and about 15% testing and verification of function.

(to conjure some numbers out of thin air)

So why do all the hard work and then rely on one of these AI tools to save the 5% effort of clicking the keyboard a bit?

When you write code you know how it works and what third party libraries it depends upon (and you know why you chose them), and the only hallucinations and assumptions and fantasies you need to worry about are your own. You build it and document it in a manner that another person can come along and read it, understand it, and maintain it.

I'm not totally ****ting on AI in this field. There are great use cases for it as a tool to assist someone who's competent. A good example would be automated tools to examine code for memory leaks, buffer overruns, and that sort of thing. But for actually conjuring the project out of thin air, call me a dinosaur if you will, but **** that, it'll all end in tears, and forgive me but the naive people and corporations buying into this are 100% going to deserve exactly what they get.

Whew

I suppose none of this matters for trivial tasks or hobby projects. And there's certainly no shortage of badly written programs out there that work, mostly, well enough, that maybe AI could've done as well (or as poorly). I'm probably overreacting and maybe time will prove me wrong.
 
Advertisement - Members don't see this ad
CopaceticOne said:
Basically AI is another compiler in the coding process in your example. Only thing different from what was before is that the person doesn’t need to learn a more complicated syntax to follow. Instead, they can theoretically use their native language.
Click to expand...
That's the crux of the problem. Native languages spoken by people are inherently vague. Computers don't manage uncertainty well.
 
pgg said:
So what you're saying is, to use these tools effectively, you need a firm grasp of data structures, algorithms, user interfaces, security, and project management, as well as a nuanced feel for the precise language used to communicate your vision to the IDE, er, I mean Claude.

...

Is that not what writing a program in a traditional language is?

🙂

Programming (competently) is 80% planning - designing and specifying what does when and where and how - and about 5% typing code, and about 15% testing and verification of function.

(to conjure some numbers out of thin air)

So why do all the hard work and then rely on one of these AI tools to save the 5% effort of clicking the keyboard a bit?

When you write code you know how it works and what third party libraries it depends upon (and you know why you chose them), and the only hallucinations and assumptions and fantasies you need to worry about are your own. You build it and document it in a manner that another person can come along and read it, understand it, and maintain it.

I'm not totally ****ting on AI in this field. There are great use cases for it as a tool to assist someone who's competent. A good example would be automated tools to examine code for memory leaks, buffer overruns, and that sort of thing. But for actually conjuring the project out of thin air, call me a dinosaur if you will, but **** that, it'll all end in tears, and forgive me but the naive people and corporations buying into this are 100% going to deserve exactly what they get.

Whew

I suppose none of this matters for trivial tasks or hobby projects. And there's certainly no shortage of badly written programs out there that work, mostly, well enough, that maybe AI could've done as well (or as poorly). I'm probably overreacting and maybe time will prove me wrong.
Click to expand...
Well yes, you need to know how to write the code yourself (mostly) and understand concepts and vocabulary to know what to ask for.

But if you’ve ever had to bang your head against a wall for days before you found the silent bug in your algorithm because the IDE doesn’t flag it and it was essentially a typo that still compiles, you’d appreciate how the LLM eliminates certain pain points.

Like I said, you’re within your rights to dislike the new paradigm but if used correctly it is absolutely a productivity booster and removes the kinds of pain points that I hate.
 
I use claude a lot for golf. I taught it how to “read” a snap shot of my scorecard that contains data for each hole from my rounds and tell it to add it to a season long spread sheet and summarize how the round went. Then Claude and I go back and forth on what it got right and wrong and areas of improvement. Took a half dozen rounds to iron out its reading but it’s been spot on since June.

I think it helps and it’s easier and cheaper (and probably safer from a data breach perspective) than one of the many golf apps out there.

Also been using Claude code for a golf gambling app when I play with buddies. That takes a little longer to plug in the games/points/strokes some people get, but hopefully will be able to do the same thing of taking a photo at the end of the round and make sure the bets are settled correctly.

Chat is better for image creation.

I like using both.
 
BoogieAndGroove said:
I use claude a lot for golf. I taught it how to “read” a snap shot of my scorecard that contains data for each hole from my rounds and tell it to add it to a season long spread sheet and summarize how the round went. Then Claude and I go back and forth on what it got right and wrong and areas of improvement. Took a half dozen rounds to iron out its reading but it’s been spot on since June.

I think it helps and it’s easier and cheaper (and probably safer from a data breach perspective) than one of the many golf apps out there.

Also been using Claude code for a golf gambling app when I play with buddies. That takes a little longer to plug in the games/points/strokes some people get, but hopefully will be able to do the same thing of taking a photo at the end of the round and make sure the bets are settled correctly.

Chat is better for image creation.

I like using both.
Click to expand...
Interested in this, Boogie! Tell me your secrets on the scorecard part.
 
pgg said:
That's the crux of the problem. Native languages spoken by people are inherently vague. Computers don't manage uncertainty well.
Click to expand...
Look into retrieval augmented generation for one of multiple answers to this issue. With enough precision RAG , reinforcement learning, etc even a SLM can feel like it knows what you want better than you do / is reading your mind.

I am one of the most natural born skeptics you’ll ever meet. And I was skeptical about these systems being much more than gimmicks at first. They are not. Of course to get real value out of them you have to find a use case where they shine and learn how to actually use them, which is a time and effort investment beyond just playing around
 
Shimmy8 said:
I use it a lot for investment, tax planning, and prep for my CPA meetings. It is incredible for summarizing data and spitting out info.
Click to expand...

Same. I bounce a lot of ideas off of it in terms of budgeting for fixed costs/variable expenses and how much to expect to save/spend monthly/annually. I used to do it by hand, but using Claude is so much faster to see/visualize. It frees up a lot of my time.

I also use it for non-medicine/non-financial applications that are outside of my comfort level (Linux/CLI). I can troubleshoot issues much, much faster than browsing forums/Discord on my own.

I understand the environmental concerns, but the biggest scam is corporations convincing you that you are the problem. It’s a drop in the bucket comparatively. That being said, I use it maybe every other week when it’s time to budget or troubleshoot tech stuff and not as a daily Google alternative.
 
You must log in or register to reply here.