90% Percent Of My Code Is Generated By LLM's
ฝัง
- เผยแพร่เมื่อ 11 ก.ย. 2024
- Recorded live on twitch, GET IN
Article
www.techsisten...
By: Adam Gospodarczyk | x.com/_overment
My Stream
/ theprimeagen
Best Way To Support Me
Become a backend engineer. Its my favorite site
boot.dev/?prom...
This is also the best way to support me is to support yourself becoming a better backend engineer.
MY MAIN YT CHANNEL: Has well edited engineering videos
/ theprimeagen
Discord
/ discord
Have something for me to read or react to?: / theprimeagenreact
Kinesis Advantage 360: bit.ly/Prime-K...
Get production ready SQLite with Turso: turso.tech/dee...
The i in LLM stands for intelligence 💀
llm
The W in LLM stands for Win.
@@khai96x so, L for Lose?
@sun3k two L make a W, though
@@no_name4796 nah 2Ls is like ultra lose
This whole article is “Fill your life with AI confirmation bias”
So well said..
Nah, short Nvidia.
Invest on Lockheed
@@monad_tcpRemember kids, the market can be irrational longer than you can remain solvable
Creativity comes from understanding
Creativity comes from experience
*Proceeds to use LLMs to write programs for him*
The goal of the midwit is to assuage his fear that he is not the smartest person in the room. The latter being the fact, prevents him from realizing that his use of AI is confirmation of what he fears.
He use LLM to write his text.
Those who can only use LLMs are the only ones that are going to be replaced by LLMs, because they never had any skill to begin with, or lost the skill from the lack of use.
OP in a nutshell : " I cant code , I cant write , I cant even plan a coherent thought to save my life . Frankly, i am a bumbling moron. Anyway let me enlighten you about creativity" /s
Another one got it right. Respect, dude
@@SomeImposter I felt the sentance structure / punctuation was off a bit but was too occupied by OP's nonsense. Thanks.
they sure can shill for AI junk tho, probably make a buck at that too. Buncha dogfooding.
...maybe it's just me; but it's been like that on the Internet and in the work market for about 10-15 years now, and AI has absolutely nothing to do with it.
Wish I had read this comment first and moved on, no shade to P
All I learned from this article is that people who rely on LLMs for writing suck at writing. And I think there is correlation between the ability to write clearly and the ability to write clear code.
Since writing clearly is about structure, the intent of knowing your reader, planning ahead and being succinct… I think you’re right
no wonder I suck at coding. i was never good at writing essays. seems like mba is the way out for me from this wretched world
@@mango-strawberry I am decent at coding but also suck at writing. People suck and hardly ever understand me though so...
🤣
🤣
Generating code using LLMs is like magic. More specifically, a magic teleportation spell that can only teleport you one step, and you never really know if it's going to teleport you forward or backward.
This isn't true if you know how to prompt correctly.
@@Oncopodathis is definitely true because no matter how good the prompt is, the answer might still be wrong
@@ggeilokowski Than more accurate your prompt is than more accurate your llms output will be.
the brownian teleport.
You can’t even claim ownership of “your” code at that point 😂
You havn't been able to claim ownership of your code for a long time (Unless you created discrete mathematics, linear algera, wrote every C compiler and standard library, etc).
You might own an idea, but those are ultimately worthless. We are ALL standing on the shoulders of giants.
If you're the one using a tool, you're the owner of (and responsible for) what it makes.
then i'll GPLv2 that shit. if I can't own it, no one can. hopefully. or not. who the hell monitors software rights anyway? lmao
Possibly, if you're making unique apps / front-endz with functionality nobody else has thought up, but for most ML scripts (plus a lot of back end stuff) there doesn't seem to be a huge amount of deviation from the norm until you get into making bespoke NNs. For 80% of the cookie-cutter code most people need, CoPilot etc. saves a ton of time.
@@TheJackTheLion I never got why thats such a popular statement in software. No other craft is like that where "Oh your a chef? Did you raise the animal/ farm the crops"
90% Blackbox is crazy, when your the one who has to fix it when it goes down...
(Atleast when your trying to make something useful and maybe even be financially succesfull. )
Spoilers: They don't have a product. They do have a short scripted tech demo trying to rake in VC.
Depending on the project I probably have 20-80% code generated by some LLM. In the instances where most code comes from "AI" I do code review every bit of it and that's basically what increases the amount of code generated. If something breaks I def know where it broke (cause I was too lazy to follow a train of thought I had was a classic "what if this input is something unexpected" - this basically hasn't changed tbh. I have almost always known where my code could break. Be it from an LLM or my fingertips.)
Worth considering how much of that 80% is your own previous code the AI is spitting back out at you
Made my second product just this year bringing in over 1kk a year with 99% AI code
@@KevinJDildonik As usual with every single proponent of AI: noobs who don't know what real software is praising AI helping them on school-project-level work and then telling everyone it can code anything by itself. Lmao.
>AI: write texts like a cold robot; codes like a creative artist.
lmao accurate. Also not surprising. The writing has been trained on mass-produced SEO-optimized marketing-focused slop and the code was on things juniors publish on their public amateur githubs or other open-source (not architected / quality controlled / planned) stuff and simple tutorials.
So creative it calls functions that don't even exist.
This guy didn’t have an L or W take. He had a G take. Let’s be real. He generated this entire post.
Rated G for Goombus
This article is the greatest unintentional parody I've ever seen.
This article was written 90% in AI. Awful grammar, nonsense sentences, lonf passagws of useless advice that feels like it was copied from some English major trying to write a press release for a tech subject they know nothing about. But it's enough to impress people who don't know much about the subject!
lonf passagws
@@LunarLambda That's just Muphry's law in action.
The sad part about this article is I'm pretty sure the wrong parts were the only ones written by a human.
Could just be Indian
Maybe the bad grammar and nonsense sentences is an indicator it's not AI.
Usually AI has perfect grammar, the guy who wrote it just isn't a native English speaker..
I think letting AI write most of your code could either be an enormous problem or no problem whatsoever depending on whether or not you actually understand all the code it's providing you and can debug/ optimize it accordingly. If you're blindly copy pasting then that's no bueno, but I can safely say that so far AI has accelerated my learning by a lot because every time it provides me with an answer I do not understand I go line by line and manually try to understand everything. It's just a tool like anything, and the responsibility is on the programmer to use the tool correctly.
If you're getting away with 90% of your code being AI-generated, your products are inherently trivial and have nobody at them that cares. It's more an indictment of how meaningless your work is at that point than anything else.
I can agree here
@@neruneri Most of the web is made of "trivial" and "meaningless" by your words code. Small businesses just want simple web pages displaying the products / services they're marketing and a contact form at most. What's the most complex thing about building a web commerce website for a small-mid sized retailer? Adding a live chat and stripe integration? What is reddit/twitter/any social media? People submit forms ( posts ) and other people read them. Just those 3 variants of web apps I listed are like 90% of the internet we're all using today. Not everyone is Linus Torvalds, working on massive low level projects for decades and that's okay.
@neruneri Nah, not true in practice. There's so, so many inefficiencies everywhere in every business around the world that there's ever growing amount of trivial work that gives companies huge benefits.
Lots of people care about simple automation.
That's why even low-code solutions sometimes work. Except those tend to be not so cost efficient long term.
Lots of companies essentially only need a bunch of CRUD applications to automate paper trails. It's not unique nor innovative in any way whatsoever, but you need somebody to do the work.
If you 'understand' the code the AI writes for you, that means you spent more time trying to understand the code than you would writing it *with assistance* of AI.
It's like learning to do a backflip by watching videos, sure, possibly you can learn by watching 500h of videos, but you could also learn it by a week of practice.
He means to say that he doesn't know how to code. The other 10% is from copying and pasting from websites.
He also doesn't know how to write English , 98% of the txt was LLM
That's how I code, but it turns out there are limitations to not knowing what you are doing, so I will need to aquire some degree of competence, eventually...
Hey, I'm the author of this article and wanted to clarify a few things you've mentioned.
First of all, this article wasn't generated by an LLM, except for the cover, which is obviously a Midjourney. I use LLMs for rephrasing or translation to some extent, and this may be why you thought it was generated.
The title "Up to 90% is now generated" doesn't refer to workflows like "Hey LLM, implement this feature." Instead, it's more like pair programming with LLM, as mentioned in the last section of the article you skipped. In this workflow, your role shifts from typing code to doing what you mentioned at the end of your video. It gives you more space to put things together, which LLMs can't do on their own.
The goal of this article was to encourage programmers to work with LLMs responsibly, neither fearing the technology nor using it without reflection. This is why I wrote about the limitations of this technology, which many people are not fully aware of.
As I watched your video, I realized I could choose better examples to show how LLM can help.
Anyway, thanks for the video. I appreciate your attitude towards my perspective and your feedback.
No, this is just copy-paste coding.
These people were copying from stack overflow before and now they're copying from chatgpt.
It's the articles sentiment presentating things in noobish perspective that get peoples jimmies up vs poorly describing a technical process(that already exists).
The bad surface level take is that 90% of the HUMAN doing the programming is no longer needed because that is the general industry hype/fear the LLM industry is peddling to further itself in the C suite and household.
When the reality is new programmers are oblivious to 90% of the code/tooling they are using is already "generated" either as boilerplate, packages, libraries etc ... except now it's also the glue-code/"duct tape" being generatable.
Either viewpoint means 90% more productivity requirements being made up out of thin air and insisted on based on a process that is mathematically non-deterministic; roflmao.
Pair programming with LLMs 😂😂😂
Oh man, just admit you’re a shitty programmer
And you sell some crap courses on top of that 😂
Im shocked! /s
The only thing the article shows is that you, as a programmer = a big skill issue.
I already have to deal with such people in my workplace. Exhausting.
I like using the odd LLM. Learning Python, R, SQL, SciKit-Learn, Plotly, Matplotlib, all the other packages necessary for data science, PyTorch, TF etc. took a metric shit-ton of time so now when I build a project I can concentrate on the big picture ideas and let the LLM fill in the cruddy stuff. Some packages like Plotly require 50 lines of code or more (depending on the sample size you're tryina visualise) to create certain subplots and their docs aren't as comprehensive as they could be, so then you have to ratch through the provided forums for the answers which can take sooo long. LLMs can take the mental exhaustion out of the project lifecycle for sure.
Right. Some people just need the code solution to do more important work and programming is a necessary evil that is in the way.
So they just give you the correct code for your plots? How many tries does that take? In my experience with plotly, it's enough to create a couple of templates and then reuse them with minor changes.
@yds6268 If you need a new template, it makes that process fast. If you only ever need a handful of templates, I would love to have your job.
I have a terrible job now. What do you do?
@@yds6268 That's also true, you can write a function to return the plots later on, but there's the odd one or two that I don't re-use in the project and take a lot of code to complete (especially if they're subplots & I have lots of columns). But otherwise yeah, LLMs / Deepnote AI / CoPIlot usually takes a couple of attempts to get exactly what you want but it gets there. Deepnote is probably the best at it imo.
@@iMagUdspEllr academia. I usually can use 1-2 templates with minor changes. My point was that current LLMs are incapable of making a correct template without me basically having to rewrite everything from scratch.
"90% percent of my code is generated by LLM's"
Compay: Hmm... so you are saying we only need to pay you 10% of your salary and GPT fees.
"10xing your skill of duct taping"
The scary part about people that overuse AI is a lot of the time when AI generates code, it is wrong in some subtle way. And if you aren't a good enough programmer to look at its output and realize how and why it is wrong, you're going to introduce Problems that you don't understand and therefore you are not going to be able to debug.
When a junior dev thinks he's a senior because of AI writes an article
I've never seen a single case where people who can prove they have majority AI code have an actual commercial product we can see and test ourselves. I have seen thousands of cases where AI just means Actual Indians did the work and some techbro is trying to do a rug pull on VC. Slight gap there.
AI will replace programmers, they say. Meanwhile, software development is being outsourced more to India and Asia even in 2024.
Actually Indians or Asian Intelligence
Man, what a trash article. Literally what Prime said, it has no statements, just fancy empty words
Probably written by the AI that also writes his code.
@@bogdyee the only problem is that he do not give real evidence , number, he is not precise so that is why it feel empty because it is.
The author is from poland
@@mate4138 i love poland
Article be like "How I as a .01 Engineer increased my output by 10×!"
normalize being a 1x engineer 💅
@@driedpotatoes Oh I'm certainly a 0.7 - 1.0× Engineer
0.1x engineering - everything you touch makes more work for the people around you.
Yep, the whole building was not built by humans, it was built 99% by a crane.
100% of my binaries are generated by llvms
Buddy, you posted an AI generated article about how all your code is AI generated. You’re not even human at this point, you’re just a gateway proxy for a computer.
90% Percent -- that's not even 1%
90% is at least 2%
Dont forget this article will be in the next batch of llm training data
I find LLMs to be fairly useful for generating a snippet that I then have to, in most cases, make several changes to. Its also good at helping identify typing issues I have when using Rust. But I couldnt imagine using it to generate most of the code I need for a project.
I’ve been genuinely curious about this: If you’re generating code that you have to go back and make several changes to, are you really saving any time vs just writing it yourself?
Also, what problems are you having in Rust that the compiler can’t point out?
@@artemisclydefrogger they are usually minor changes so most of the time it's quite a lot faster.
The most recent compiler issue I ran into was a variable was needing to be cloned inside of a move. The compiler was looking at the e: Event in the closure and said something about the closure needing to implement FnOnce trait or something along those lines. It was really throwing me off but the chat bot identified the issue pretty quickly. It was a function to upload a CSV file to my Go Gin back end in Leptos. Then run a series of cypher queries to add the data and create the proper relationships with my neo4j db
I was trying to handle this with my rocket back end but I cannot for the life of me get that thing to compile while handling file uploads. I got custom web sockets with a pretty in-depth fully authenticated and role permissioned rocket backend running, but I ended up just spinning up a go gin server because it's so much easier. I seriously got tired of banging my head against the wall.
It can also identify issues (particularly dumb ones) in your code you might not have noticed. Just be wary of any solutions it offers.
@@balijosu Is this better than LSP and a decent compiler? Such as Rust + rust-analyzer, ironically
@artemisclydefrogger I can really only speak about C++. The sort of error I'm thinking of is doing a dumb thing like:
PointXY p(other.x, other.x);
where the 2nd arg should be other.y. No compilation error/warning, but an AI tool will likely smell that something's wrong.
I just want an LLM to do my log messages.
The days i used LLMs, i felt more productive in the beginning, quickly deteriating, ending in me cancelling my subscription. But what LLMs have driven me towards was writing more and, especially, better log messages, simply because doing log messages was just a matter of doing "logger.log" and the LLM would complete the log message with, most of the time, the correct context.
That is the one thing i miss in my current developer life ...
everything else tho teared down my productivity in the long term. Quit when i noticed, i was writing a parser just based on AI, not even thinking about what i am doing and getting extremely frustrated the moment it stopped working. A PARSER... something so basic, something i have written by now hundreds of time (creating languages and file formats for fun), causing me headache.
I've used AI in coding and there are a few places that it does okay and a few places it falls short.
Okay:
1) Generating code from templates. If you have very generic code and you need like 20-30 similar functions, it can do a decent job at generating those from a template.
2) Converting one language to another.
3) Helping you learn a new concept. If you are unfamiliar with a language or concept, it can help you learn though generating examples. It can really push you in the right direction if you're a bit lost.
4) Troubleshooting. If you have an error and you're not sure how to solve it, it can sometimes provide you with additional information that you might not have had off the top of your head. However, sometimes it will lead you in a completely wrong direction.
Meh:
1) Bouncing ideas off of. If you kind of stuck and trying to make a decision on something, it can serve as an okay sounding board.
2) Generating docs. If you want a quick and dirty code documentation it will do okay at this. Though we all know how AI writing is.
3) Writing tests. Sometimes it produces okay tests for code you give it. They won't always be complete, but it's a good starting point.
Not okay:
1) Generating code from scratch. It can shoot you in the foot pretty quickly, even for simple pieces of code. It will make dumb mistakes and you won't catch them because you think the AI is right.
2) Code reviews. It just doesn't have enough insight into the codebase to really provide good insights. Though it can catch minor things.
3) Iterative code generation. You spend a lot of time going back and forth with the AI until the code is correct. It just isn't worth it.
Awful:
1) Creativity. They are great at regurgitating facts. Terrible at coming up with new ideas. They just can't
2) Working on new concepts. If you're designing a new language or writing some new concept that is way out there. The AI can't keep up. It just tries manipulating your new concept into something that it is already familiar with.
AI is also awful in working any qualitative working any sort of qualitative measure.
@@ghfudrs93uuu That's fair, it really doesn't excell at that
Ftfy "Not okay:
1) Generating code from scratch, except for common algorithms that are easy to spot check and test. "
@@TheNewton But then you have to ask yourself, why are you asking it to write code like that in the first place.
YES! This is my exact opinion! Although i would place converting a language to another in the meh category.
90% of my commit messages are generated by LLM's
this i approve of😂
4:54 - You are off-brand Pokimane, Prime.
"When you have too much understanding, you create a box and cannot leave the box". yea this is why I like having juniors around. They will always try. They usually fail and when they succeed you have to fix all their code. But they always try. A lot of seniors on the other hand will just give you the "that's not how you do that" bs.
I learn from my juniors. They ask questions about stuff I take for granted. Always challenging me. When I give guidance I try to explain how something they're doing might have consequences rather than "don't do it like that."
"When they succeed you have to fix all their code"
LOL
Copilot is very nice for a series of trivial auto-completions, but you need to be able to immediately recognize trash and toggle it off.
90% of code is boilerplate , so no problem if LLM takes care of it.
like "AGI is arrouns the corner!" Yeah.... wait for it...
I am only interested in code generation, so we can embrace.... whatever, we can't.
90% of arguments about AI are generated by unsubstanciated opinions.
It's usually akin to "This food is horrible, because that's not how my mom cooked."
Hitting the like button after 0:14
6:10 - and remember, all the geniuses dropped from university!! A staple known fact!
All the genius? As far as I know we only have 1 (one) Nobel Laureate who didn't finish college.
I'm talking about people who really did some important contribution to science.
@@0xhenrique oh c'mon, Robet Kiyosaki talks about these people all the time! Like Elon Musk for example
@@Neuroszima "You don't need university to scam people and become a multi-millionaire" - Robert Kiyosaki (probably)
Making a Tauri app requires basically no Rust knowledge if you don't want to do it.
The anecdote about waiting for Copilot reminded me of something. When I was a kid I was playing this Apple II game, I believe it was called Word Spinners. One of the edutainment titles that were pretty much a must for a classroom Apple II. It featured a princess much like the one from Rumpelstiltskin, spinning wool into... words, I guess? Anyway, one of the modes was like a two-player crossword. You and the computer (or another human player) took turns filling in the spaces with words. The words had to fit in the provided spaces, be in the machine's dictionary, and once a letter was placed, any word building off that space had to have that letter in that location. I think you win if you can stump your opponent into not being able to make a valid move.
So I placed the word "CUTEST" horizontally in a six-letter space. There was a five-letter space going vertical, the last letter being the U in CUTEST. I figured I was gonna try to get it to say HAIKU, and test whether it had that word in its dictionary.
So the program ground, and ground, and ground, and I waited, and waited, and waited. Eventually I think the teacher came over and offered to reset the machine, as she thought I'd broken the game. And then, finally...
HAIKU
I prompt-engineered a kid's word game into guessing the exact word I wanted it to guess. Let's go. I had broken the game, just not in the way the teacher thought.
At 2:10 did you really say “obviously the i in LLM stands for intelligence” ?? 😂
Okay, that article is truly terrible, but I had a quick look at the tool they build and while it's not the most complex thing ever, it does look like it solves some real annoyances people have with LLM chat interfaces.
Writing code with LLM is like listening to college lectures while sleeping. You may feel like you're being productive but when the day of the exam comes you're going to have a bad surprise.
Holy crap, this article sounds like it was written by an AI impersonating a scammer trying to sell you the wonders of AI.
"This [AI generated] article is the appearance of the thing and not the thing itself." And now I'm thinking of AI as a part of Jean Baudrillard's concept of hyperreality.
the whole idea for LLMs was to fill in the blanks when you have a tedious task ahead
It's so funny to hear 27:05 when my last work contract was 90% generated, with context from an old contract. My genAI contract was enforceable and the company who tried skirting around a clause was forced to pay out. I never spent a dime on a lawyer, they probably spent $200/hr -- so I'd say it worked pretty well.
I'll note that this strategy will work until you've got to deal with legacy code that's written in a language that isn't as well-documented as C, or Python, or Java. Like, for example, Scala. Now, as an experienced coder, you'll probably recognize that Scala sits in a weird place between Java and OOP languages and, well, if you've got the chops you can work back from whatever you're looking at. But those LLMs sure as heck cannot. Now you *can* hack an LLM to work with something like Scala by providing a "working from the safe assumption that Scala is fundamentally the same as Java in structure with the exception of allowing for object-oriented programming..." but it's still going to struggle not having a mountain of example code to work from.
I also find anything an LLM produces tends to require you to do the final bit to actually get it working on your system configuration. I mostly do not use an LLM coding assistant/LLMs for primary programming.
Yeah and we still need LLMs that can reliably cite the source and not just summarize.
Like a new form of static analysis, you ask it where every use of pattern-X is and it finds it instead of having to craft custom regexes for every little thing.
scala mentioned
2.5 months into my first job with the government lol, it's not the worst but if you want to be moving quick and building things with bleeding edge tech it's definitely not where you want to be
This so reminds me of the story The Marching Morons (C.M. Kornbluth, 1951). "I pushed the button, I'm a programmer."
"Why would you want to use natural language to create code?" - For the same reason, your manager will use natural language to tell you what to do, rather than talking to you in JS or yaml: He doesn't have time (and knowledge) to bother with the details and leaves that part to you.
But with a human there is a conversation. Interpretating any non-trivial request is really hard. It takes time and effort to clarify what they want.
If you don't understand the details of what it's doing, then you *definitely* shouldn't be using a LLM (not only because it probably doesn't even work at all, but also because you need to know if it handles all of the edge cases properly or if it has any security issues). It's your job to know what it's doing - if something breaks and all you can say is "idk the LLM did it" then you're losing your job (and rightfully so).
Sounds about right. My java apps have 0% ai, but my notebooks and data pipelines are about 90% assisted, especially if I'm using tensors - from a software architect.
I totally agree with your rant about carousels, but… my version of the rant is about writing your own hashtables and other data structures. Every language having built-in dicts is great in aggregate, but stunts the growth of individuals.
19:30 I too am very good at saying everything without saying nothing. This happens when I think understand but do not.
PLST TWIST: the whole article was written by chatgpt
Where's the plot twist?
I started playing D&D with a couple people who had never played a role playing game before and when we first started playing they did some crazy creative stuff that I loved. Over the years as they've learned the rules I've noticed they have stopped being creative and they get entrenched into what does the paper say I can do. Now I find that I have to constantly remind them that they can do crazy weird stuff and they might just get away with it and so they should try but it's really hard for them to pull themselves out of the rules. This is exactly what you're talking about here where once people get really familiar with something that's something often becomes limiting in their creativity.
This reminded me of the episode of "Billions" where some of the traders took a limitless drug and they had the illusion of being smart for a few hours.
03:12 I spent multiple days (not all time of day) figuring out why lazarus ide crashed when closing certain dialogs. There are multiple layers of abstraction: lcl, libQt6Pas, qt6. Crash was use after free in qt6 (QTBUG-124237).
I've only found 2 use cases for LLMs: generating documentation for pocs and getting ideas for small refactors
this article seems like a devious advertisement
AI suggestions when I enable them have an interesting effect on me.
Usually I think about the code before typing it, writing the comment before the real code (which also helps the model) and I often think about way too complicated methods of getting the job done as I like abstraction layers and data-driven behaviours that IDE static analyzers don't really get (have you tried generating a UML diagram when there's a Factory pattern involved?).
But a good LLM will "get" more context than the analyzer, probably thanks to the comments, and will either suggest me EXACTLY what I was gonna write, in which case I just Tab it in, or a way simpler method of getting there, in which case I'd think about it for a while and often end up thinking "you're right, my version would have been too difficult to maintain in the long run, let's do your thing with some changes for now".
EDIT: To sum it up, when AI is used for peer-programming without a human peer, it can improve your output by having you question your code more.
💯
I call it Interrogative Programming / Socratic Programming.
We can get to a point where we no long tell the system what to do.
We ask it why its doing something like small petulant child we are teaching reason too.
1:20 "and often cannot hand;e tasks that are obvious to us, like telling if 9.11 or 9..9 is the bigger number"
Me literally not knowing which number is bigger...
I’m a research student in social science that uses AI to help me write agent based models in python. It is an absolute god-send! I’m slowly increasing the amount I can do and decreasing what AI does. For a complete beginner in coding, you absolutely should not feel ashamed to use AI. It’s basically the new version of copy paste from stack overflow.
Bytheway, that article was 100% written by a human. AI’s don’t make such awkward sounding sentences like that!
Developers coping and seething about AI replacing them is hilarious. Unless you write very sophisticated code (95% of you don't, and no, centering a isn't hard), it's pretty much over for you, just like it's over for the average concept artist.
I really have no idea how is that possible? Even when I try to use it as much as possible I cannot approach 50%. Like, do people explicitly try to use AI to generate code even when it does not make sense to do so? If you are not trying hard there is no way to be close to 90%. Even if you are writing boilerplates only.
Give cody from sourcegraph vscode extension a try with sonnet 3.5 and you will see how.
You need to use a better model or better tools to give it the right context. Cursor does this very well.
The only time I've been happy with code generated by an LLM is with bog standard Jupyter notebooks. Otherwise it feels like it gets 90% of the way there but that last 10% error just accumulates the more you push forward with it
"Things I have no influence over" Don't LLM's like... learn to code from *_YOU?_*
I am probably near 80-90% myself. Its part of my. workflow now, you either adopt it or get buried. I still of course have to lay down the code, review it and make sure its assembled correctly.
To be fairrrr - Aider destroys Devin in every benchmark. The author of Aider puts as lot of effort into it.
Man, I just read up on “the well-known story of Coleridge at the waterfall.” (I never heard of that before). It's just 5 levels of people arguing about other people's opinions, injecting ethics, politics, and other nonsense into it. As a result I don't want to talk to people anymore.
0:35 Chad Geppetto ?! 😂 Pinocchio would like a word!
I think its feasible, but you gotta corner the LLM constantly to stop hallucinating.
Its way more work to force LLM to write your code than to actually write it
Yeah, it's like working with an intern who never actually learns, so tomorrow the LLM will say the same stupid crap it tried to pass off as code yesterday.
If this is the future then it's completely worthless. There won't be any money in coding. This is what "AI bros" and "AI coders" fail to understand. They are more in denial than anyone.
to be fair to A.I. I have used it to generate some complicated rust code and it gets it within 2-3 tries 90% of the time. I still need to code review, write tests and ensure there is no UB but it's pretty good.
obviously its gonna lay you astray sometimes, but when it works and give you a correct answer that you can't find on any forum or when it helps with doc its actually genius, way better than the doc since its interactive
leveraging LLMs to write APIs to talk to LLMs to write blog posts about LLMs
This article: “I used LLMs to write a Rust app, then had an LLM write and article about it.”
I wish that I knew something about coding I don't and wont pretend to. That being said I know by the way you talk about it, you do not understand much about AI\LLMs. It has not improved my dev work, because I dont do any but it has helped me improve my quality of life and my abilities in almost every area of my life in ways that I cant even explain. AI gives every single person with an internet connection the ability to have a conversation with all of human knowledge you can summon almost any information in words that are relatable to you. Remember in Road Trip when dude said I can teach particle physics to a monkey if you can find a way to relate the marital. it can do that. whatever you could want to learn about it can explain to you in a way that you will be able to understand. If AI is not improving your life right now you are doing it wrong.
@ 3:30 - which is why LLM are nice. It is an incredible learning tool.
Oh my days, that first take was sooooooo accurate. People who relay too much on ai nowadays are the same people used to configure wordpress theme and think they are SE
After reading so much AI generated writing I am calling it now, ultra personality-saturated writing is going to become an increasingly attractive trend. Prose akin to Dennis Johnson or Gary Shteyngart will become essential, proving to your readers that you're not a LLM by being weird and funny.
The duct taping part. Bravo 👏
This man is an energy crisis walking on two legs.
Ask Louis Pasteur about ego. Because of his genormous ego we have such thing as vaccination against rabies. Or against Anthrax. Or he’ll much more other changes. Because this man wanted to prove that he is right.
So totally L take about creativity and ego.
survivor bias and hindsight make someone blind something something.
AI is my bitch, it does my grunt work and helps me have more time to do other things
The craziest article in the past like 3 years. Insane
As a Borderlands 2 enjoyer, I will never not think "Legendary Loot Midget" everytime I see "LLM" mentioned on the internet.
90% of my *code snippets for a generic task* are LLM generated, and I have to rewrite them slightly most of the time. It's for those cases when I don't want to go search my specific want or look at the documentation.
@ThePrimeTimeagen Funny that you mentioned that quote from Bob, but did say in one of your past video that "no one can't control the future".
This might be true for me too, I iterate as I write and will generally go through explicit versions of a given file. Copilot knows what version 1.0 was like and so it will autofill a lot of boiler plate and other stuff and then I just edit. So in a sense a very high percentage of the final product is copilot generated but before it was copilot generated it was generated by me.
I'd go as far as to argue that without an incredible amount of ego one cannot "change the world", creatively. To achieve such more or less requires you to become invincible to criticism, and to let your own perspective, however stupid, flawed or backwards it seem from the outside, drive you all the way towards the end.
You should have used LLM to summarize it into a single paragraph.
We need a prime t-shirt that says "TH-cam, the right job for the tool"
7:32 - who cares about skills? we get paid to complete projects, not to be skilled.
19:28 - the AI bot literally just said "like and subscribe" in the middle of the text!
I use AI for my code too.
I treat it like an entry level programmer, meaning I have to double check it's work. This means you cannot be a newbie at programming and use it effectively. You still have to know how to code and use programming paradigms.
I use LLMs to write individual functions like any other search. They can't create entire projects yet but they'll get there.
I've got a few projects that are 100% AI generated code. They all don't work, but I'm just imagining the day that they do.
Just like with Kamabla, the I stands for intelligence...