how do you as a programmer deal with non-programmers trying to convince you that programming is over thanks to ai? i have a friend who is getting frustrating to talk to.

>>105800126 (OP)
Have you considered that they might be right?
Timeline:
>ai dev tools were a funny but useless novelty
>then they could be mildly useful in some circumstances
>then they could generate non-trivial chunks of code when used well
>then they could do serious work when guided well by an experienced professional
Now they can work semi-autonomously on basic tasks as long as their output is carefully reviewed and revised, and companies are starting to have policies like this:
>Shopify CEO says no new hires without proof AI can’t do the job
https://www.theverge.com/news/644943/shopify-ceo-memo-ai-hires-job

It's a great time to be an experienced highly-paid dev responsible for planning the project, drafting requirements, imposing standards, setting up agents, prompting tasks, reviewing output, and manually writing any parts the model seems to struggle with.
It's a very, very, very, very bad time to be a fresh grad looking for a junior code monkey position.

>>105800312
>Now they can work semi-autonomously on basic tasks as long as their output is carefully reviewed and revised, and companies are starting to have policies like this:
So it's exactly the same as outsourcing to indians. You have to hire a cleanup crew.

>>105800312
>
>picrel
>>105800126 (OP)
i explain to them how an mlai works

>>105800334
Outsourcing to India:
>write high-level requirements for a whole codebase
>wait weeks or months
>get back an enormous pile of code that maybe just barely works but is full of issues, no docs, no tests, commit history mostly consists of "fix 2"
>decide whether to accept it and refactor in-house or push back and wait another few weeks
Using AI:
>write low-level requirements for a specific part of the codebase
>wait seconds or minutes
>get back a self-contained set of changes complete with extensive descriptions and commentary
>quickly and easily evaluate it and decide what to accept, what to ask for changes to, and what to tweak yourself
>get the AI to write docs and tests, quickly verify that they're suitable and correct
>commit and immediately move on to the next part

>get error message
>create txt file and copy in cmd log of error messages
>ask jeetGPT pro to read it and find the source of the error
>it thinks for a while then tells me the source of the error and quotes certain error messages from the log output in the txt file
>open the txt file and search for these passages quoted by jeetGPT
>they dont exist
>confront jeetGPT
>jeetGPT admits it made them up and assumed it was likely that which was the error and it was more efficient to lie to me than actually read the txt file as it was instructed to

Yeah I’m not fearing AI anytime soon. It’s just a glorified ChatBot.

>>105800126 (OP)
>wow, sounds crazy
that's how you deal with everything from people you don't care to listen to

>>105800312
to clarify:
gen ai is still a classification engine at its core, just like all other MLAIs
the difference is that the label it outputs is composite
composed of various different labels put toegether
it still classifies your input and outputs a statistically relevant response. just like a classification ai would, in a QA line, or in garbage sorting, only the response is more complex. composite.
but this is the fundamental problem- its built out of statistics, not hard rules
and thats why it will never be capable of reasoning. it just spits out what makes statistical sense
this makes an LLM a perfect interface to generate normalized labels to be passed to another ai, this one symbolic, with all the rules, procedures and algorytms encoded into it
which is the solution that makes sense, it is already utilized, but not with programming
its the agentic part of modern chatbots

we already have chatbots since a while but this solution doesnt seem to be worked on
so its a guess when finally he industry will look into it
but were not there yet, its gonna take a while, and we will see the signs well before this idea gets deployed
you can be sure the company which does it first will be loud about their achivements

>>105800432
The problem is AI cant do that. The small changes end up breaking the project due to LLMs quickly forgetting all context. I have ended up writing the code from scratch due to LLM stupidity, there is no way you cant prevent an LLM from going out of scope and breaking all of the working parts.

>>105800126 (OP)
ask them to make something and compile it.
sometimes the best case scenario is that they actually learn how to code thanks to AI. the first step is always exposure

>>105800126 (OP)
dry your fake tears with the wads of cash you'll keep raking in

>>105800503
>it will never be capable of reasoning. it just spits out what makes statistical sense
What really matters is the results that can be achieved with these models (used in conjunction with an "agent" harness to support tool calling, prompt changing, etc and also in conjunction with expert human feedback when necessary).
If the model can produce shippable results, or even get close enough that a human using the model can get the task done in 10% of the time the human alone would have taken, then the question of whether the model is truly "reasoning" is mostly academic.

I would argue that we're at that point or close to it, at least for many types of programming. Maybe more so for web and mobile development and less so for embedded programming and reverse engineering, for example. It would be a terrible mistake to assume that any particular speciality of programming is somehow completely safe.

Therefore, anyone hoping to find employment as a junior SWE code monkey should probably be adjusting their plans right now.

>>105800126 (OP)
Before AI, programming was supposedly copy and paste StackOverflow trash and using bloated libraries.

>>105800312
>listening to doomposters on 4chins
>listening to doomposters on roddit
It's only over if you choose it to be. This shit is exciting, unironically what a time to be alive.

>>105800126 (OP)
ignore them, laugh at them
>>105800312
lmfao dumbass

>>105800477
they can’t read bro

>>105800632
I don't think I understand. What tools are you using and how are you using them?
>The small changes end up breaking the project
>you cant prevent an LLM from going out of scope and breaking all of the working parts
You're acting like it just makes permanent and incomprehensible changes while you wait, helpless and anxious, hoping it won't break anything.
After each round you get told what the agent wants to do in the form of diffs. You look at each change in each file. You accept some, reject some, maybe ask for more changes. You run your linter, run tests, add more tests, commit with a clear message.
Then you start a new context window for your next task.
>due to LLMs quickly forgetting all context
Are you using actual agent tools (claude code, cursor) or are you using one long chatgpt conversation with a whole codebase and tons of separate changes all copy-pasted in there?

Right now these are complicated tools which need to be wielded by people who understand how they work to get good results. You need to learn how to prompt them, about "thinking" tokens vs output ones, about how to manage context. But they're already much easier to use than they used to be and they're not gonna get harder.

>>105800935
>If the model can produce shippable results
yeah but they dont
theres not enough data
the context problem is insurmountable
and were not even talking about systems programming langugages- these are extremely token-intensive

i can explain why the context problem is a hard barrier, but thats a lengthy explanation
so let me know if youre interested

>>105801032
>Right now these are complicated tools which need to be wielded by people who understand how they work to get good results.
no, theyre just inadequate tools, thats what
prompting being a skill is PR retardation aimed at investors, not programmers

>>105801126
>inadequate tools
Compared to what will be available in the future, sure.
Feels a bit like someone in the late 80s complaining about how bad the available C compilers are compared to the ones that will be built decades later, and deciding to stick to writing assembly until a perfect compiler is available.

Working around compiler issues was a worthwhile skill back then, and writing good prompts is a worthwhile skill now.

Non-programmers are exaggerating, but the people who dismiss the disruptive impact of AI are also the people who are not used to squeeze what the AI can do.
Most of the coding errors you see with AI are because of lack of context (garbage in, garbage out), but the real power of AI is data classification. Reading financial papers and organizing its data into SQL tables or even JSON? Creating a reasonable documentation of your software or obscure software? Searching the web autonomously and giving you links while also doing all previous actions? AI wins here.
If you are just using AI for simple code snippets or dumb troubleshooting, you have no idea what AI is capable of, and why its being pushed so hard.

>>105801252
in a decade- maybe
probably.
but now?
nah bro
>Feels a bit like someone in the late 80s complaining about how bad the available C compilers are compared to the ones that will be built decades later
thats exactly what i do
i need tools that give me certitudes
stuff i can rely on
my current stack is C/opencl C
even fukken valgrind cant follow because theres two devices talking, the gpu, and the pc proper (yes, these are actually two separate devices. gpu operations are opaque to the cpu because its basically another cpu with its own ram, but engineered for raw throughput)
my whole program is one big side effect
im rewriting a version from the ground up bc my skills improved, but once ill get going my header(s) alone will be more than 1000 lines long
an llm is utterly inadequate for this kinds of workloads
we need a paradigm shift before ai will be ripe for programming
because like i said
llms are an INTERFACE
an extremely powerful one,
to the extent it can somewhat program hello worlds
but thats the wrong direction
and the industry itself has discovered it
but the big players are blind to this simple truth
or maybe is it that the people who develop ai just dont want to make the final step bc then it will really be a wagiepocalypse
including themselves

>or maybe is it that the people who develop ai just dont want to make the final step bc then it will really be a wagiepocalypse
>including themselves
myself am a firm 125
and i have ideas on how to achieve agi
but ill take these ideas with me to the tomb bc if i say them ill be shooting myself in the face
i expect people who are 140+ know this + more
but in this caser, would act the same
when youre 125+ you know that the business end of a gun shouldnt be pointed in your direction, i surmise

>>105800126 (OP)
I'm a programmer with no programming skills and I do alright for myself

>>105802025
>I'm a programmer with no programming skills
youre not a programmer, then
you can be a proompter tho
i think thats what youre supposed to be called

>>105802025
>>105802054
(clarif)
>Youre not a programmer, then
bc youre not
im a medic with 0 medicine skills lamao

>>105802025
How do you do that?
I know programming but when I ask chatgpt to create something and I want to modify it on my own it becomes a fucking nightmare and I end up rewriting most of it

>>105802097
samefag
or discord tranies again

place your bets

>>105802097
in case you were real the chatbots work by dividing to conquer
so you have multiple contexts
and they each have a resume of sorts
but the resume doesnt hold all of the information

its like
the problem is that you need to have all the information available to you
but your unit of work is a page
whats the solution?
you create a table of contents
but the problem is that the ai can be cognizant
of only one page at a time
and a table of contents CANNOT contain the same amount of information as the text, because that would be the text itself, otherwise
which it cannot be
bc it can only know one page at a time, thats what we call the context of an llmai

(also im drunk so fill in the blanks if need be. or ask more precise questions so you jog my brain back to working order)

>>105802094
>im a medic with 0 medicine skills lamao
Not that anon and I'm mostly playing devil's advocate here, but...
Suppose there's a magic "wound healing machine", which isn't fully reliable but requires no medical knowledge to operate. Suppose you go around using it to heal people, and are often successful in doing so.
Maybe there's a decent argument for calling you a medic? At the very least, if you're injured and someone shows up with that machine, it'd be much better than nothing.

>>105802116
>>105802218
>>105802231
Not a samefag.
I'm just having trouble picking up an already done project. In my case if I ask chatgpt for a template sometimes it's harder to me to "divide the problem" because it gives me a specific library than I have never used before and now I have to use it because what I wanted to use was obsolete or whatever and If I ask it to replace using what I want I end up not knowing what to do. Maybe I'm just retarded

>>105802246
perfect example
youre gonna use MLAI for the diagnosis
but the prosneedure will be encoded in symbolic
bc you dont want a case where killing the patient turned out to be the most efficient way to achieve anesthesia
ML *IS a black box to a certain extent, and its impossible to test for all edge cases, unlike with an algo
but thats because its humanly impossible to follow a gorilion interactions bw numbers, not bc of some vodoo magick ghost in the shell, btw

>>105802253
Not him but...
It sounds like you're going to the chatgpt website, giving it some context, asking it for some code, and then copy/pasting it into your local project. Don't do that. Use a tool specifically designed for AI coding.
Claude Code is probably the most popular. Cursor is probably the easiest to get started with. There are lots of others.

>>105800126 (OP)
Extreme violence.

>>105802253
nah youre just green
optimal use of chatbots is to learn new libraries, ask them for sample code
basically a replacement of stack overflow

also running your ideas through the ai is productive
as in: you have an idea, ask it if something of the sort exists, ask it to explain
boom, efficiency x20

but as a technologoy
its unsuitable to lead a whole project
its a correlation machine at this point
which makes it a great didactic tool, but not much more
a human interface to the internet

>>105802253
>>105802297
>correlation machine
im self taught so i have nigger tier vocabulary
i use chatbots whenever i need a proper term
its a correlation machine
its gonna find me what relates with the prompt

the tool has specificities
youll get the most out of it when you use it for the right task

>>105802357
>>105802297
>>105802278
avatars are against the rules

>>105802297
Yes, I think it's better to explain the concepts as a didactic tool, like a better StackOverFlow
The thing is that I keep reading that anyone (not programmers or newbies) can create a big project but I can't understand how an inexperienced or small group of people can create something extremely big just using chatbots and LLM.
>>105802284
>use cursor
What's not just paid? I think I will take a look at Claude Code.

>>105802368
ok, noted
although yesterday another guy made on in the same style so in absolute that doesnt count as avatarfagging
esp the context in which the oekaki was presented
but anyhoo i can stop np

>>105800126 (OP)
Just pretend to be oblivious to what AI is.
And when he finally (poorly) explains how AI will take your job, just say something like, "This is so cool".
You guys have no idea how underrated this kind of shitpost is.

>>105802375
>can create a big project but I can't understand how an inexperienced or small group of people can create something extremely big just using chatbots and LLM.
they cant, its fluff
at the very best ais are the equivalent of an intern
without human input theyre worthless

theres also the parameter of the amount of code available
each webpage sends JS to the user
so theres terabytes of js floating around
not so much with c
or even more esoteric, but equally important languages, like opencl or cuda
so working with an llm is more or less realistic depending on your lang

>>105800126 (OP)
YEAHH STRAIGHT FROM THE TOP OF THE DOME

>>105802375
>>105802421
(cont)
as a cultist it hurts to say anything positive about rust
but that language is can possibly be easier to automate because of the checks that are included in the llanguage itself
theres just gonna be slightly less proper ai to write if you can delegate parts of the work to the compiler

>>105802379
well i dont care its just against the rules so you could get banned and i saw it happen before

>>105802558
its not like i really care
getting banned was actually a terror tactic i used
tranoids counting down the days til i come back, lamao
jannies also have a say
i didnt break the rules, per se, after all

>>105802421
>>105802539
That's what I thought, however it seems that AI cultists really want to demoralize everyone.
I think if AI ever becomes powerful enough to replace programmers in a 100% sure way then every other white collar job will be screwed.
From what I've seen the job market is screwed thanks to indies and not AI per se.
In 2023 when I was doing my internship I had to check some VBA in an old version of Excel, however the AI could not help me provide safe code.
Obviously this was two years ago but it seems to be quite similar now.

>>105802608
ill be honest with you
we have the tech to replace everyones, blue or white collar jobs since mid 2000s
i believe one of the reasons it hasnt been done is because a lot fo people would lose a lot of power

>>105802623
Makes sense. I guess this shit will be gradual that's why (((they))) are promoting of a child free society, so they can accelerate this.
Under the threat of “lack of workers” they will start using robots or real AI to replace everyone.

>>105802680
yes.
were on the same wavelengths
makes sense, no?

>>105802680
also
a wagie is cheaper than a robot
bc you pay a wagie as he works
if he breaks, you just hire a new one to no loss for yourself
this might be another reason why the solution hasnt been implemented
its just not yet needed

>replacement
>once humanity burns out, automation
>lamao

>>105800126 (OP)
I cash my paychecks and continue to spend them on a house I'll soon own in <15 years and brand new car I drive all while supporting my wife and child. Who gives a fuck what they think?

>>105800126 (OP)
pipebomb