pull down to refresh

I’m specifically interested in people that extensively use AI for complex coding projects.
Here’s my limited experience, so I’d be interested to hear thoughts and workflow details from someone who does this more than me:
I work in tech but not as a developer. I’ve tried using Copilot to help with personal projects here and there, and it is helpful in some ways, but I find it doesn’t come close to living up to the hype. I see tweets and posts from devs saying they use it all the time to build things, find bugs, cut 50% or more of their workflow, and that it’s a huge game changer… But that all seems hard to believe.
I’ve noticed that if I’m trying to write a very specific function that doesn’t necessarily exist in the context I’m creating, AI is of little to no value, if not negative value. Once I’ve spent a bunch of time prompt engineering, I’ve basically solved the problem and could have just written the code myself. 

I’ve found that if I’m actually trying to create something, it’s better for me to think through it myself, and what I come up with usually involves less code than what the AI outputs.
And lastly, debugging seems super hit or miss. It might fix something, it might not, or it might make me think something is fixed before it completely breaks later and I have no idea why, leaving me worse off than before.
Does it sound like I’m just completely doing this wrong, or can developers relate to this experience?
Feel free to vote/share thoughts below.
I use AI extensively for complex tasks11.1%
I use AI for debugging / troubleshooting3.7%
I do both of the above14.8%
I use AI for simple tasks only25.9%
I code but don't use AI33.3%
I don't code11.1%
54 votes \ poll ended
I use https://cursor.sh a fork of VScode with AI built-in. I can highlight a block of code and ask GPT to adjust it. Then I get an inline diff with the option to ACK/NACK or follow-up all without leaving the editor.
I found GPT is useful when:
  • I know exactly what to write. And it would take fewer keystrokes to command GPT to write it instead. All I have to do is review and accept. (Boilerplate, Type/Interface definitions, formating, repetitive tasks)
  • I'm exploring an idea or looking for existing tools/methods. I use the Chat feature like a search engine/wiki to assist in exploration. GPT helps me bridge small gaps in knowledge.
I have found GPT is NOT useful when:
  • I have large gaps in knowledge
  • The work is novel or less researched/documented.
  • Also, I notice that GPT will usually default to the most generic and general way to solve a problem, not necessarily the best way. For example, it will reinvent the wheel by implementing a complex algorithm poorly rather than use a library or existing method that is more efficient. I have yo specifically ask if there are any libraries that would simplify things.
In summary, GPT will not turn non-devs into devs overnight. If however, you already understand a problem and already know how to solve it, GPT will make light work of actually getting it done.
reply
reply
Another cursor.sh user here (and subscriber to the service, for the 500 fast gpt requests per month).
I had to do power shell scripts for the first time ever and without GPT4 or other AI the learning curve to get what I needed would have been too high to justify the effort.
The final result after hundreds of hours is maybe just 20% of the code today was generated from GPT. But when I was starting on it, nearly 100% of the code was from my prompts, exclusively.
Then for python, for example, my go-to tool, gpt is much less of a timesaver, and fairly often a time waster. At least for composing the code. For getting syntax and such though, I use the chat more than the code editor, and save seconds at a time versus opening a browser tab to Google stuff.
But the tell is ... I was paying out of pocket for the service, before asking to have it paid by my company. And if it was $100 / month rather than $20, I would still have paid it. It is so worth, for me at least.
182 sats \ 0 replies \ @gd 7 Mar
Almost always these days for scoping out new solutions and bouncing ideas for better ways, pros/cons of implementations. Super helpful to paste in errors too.
I also use copilot which is definitely a productivity booster.
reply
I'm planning to try GitHub Copilot. Turns out it can be integrated with Vim. https://github.com/github/copilot.vim
reply
I use AI occasionally for certain tasks. However, it is not advisable to rely on AI completely. It is still necessary to verify the generated code and confirm that it actually does what you want it to do. I think we are still a long way from a non-programmer being able to trust AI to develop complete solutions.
reply
253 sats \ 1 reply \ @k00b 7 Mar
Only when I have a very narrowly scoped task that I have no clue how to do exactly.
Today I used it to write this plpgsql function to take a db seed and shift all timestamp columns forward to compensate for the db seed being old:
CREATE OR REPLACE FUNCTION timewarp() RETURNS VOID LANGUAGE plpgsql AS $$ DECLARE r RECORD; max_timestamp TIMESTAMP; interval_to_add INTERVAL; BEGIN FOR r IN SELECT c.table_schema, c.table_name, c.column_name FROM information_schema.columns c JOIN information_schema.tables t ON c.table_schema = t.table_schema AND c.table_name = t.table_name WHERE c.data_type IN ('timestamp without time zone', 'timestamp with time zone') AND c.table_schema NOT IN ('pg_catalog', 'information_schema') -- Exclude system schemas AND t.table_type = 'BASE TABLE' -- Ensure targeting only user-defined tables (excluding views) AND t.table_schema NOT LIKE 'pg_%' -- Exclude other potential PostgreSQL system schemas LOOP -- Calculate the maximum value in the column EXECUTE format('SELECT max(%I) FROM %I.%I', r.column_name, r.table_schema, r.table_name) INTO max_timestamp; -- If there's a maximum value, calculate the interval and update the column IF max_timestamp IS NOT NULL THEN interval_to_add := now() - max_timestamp; EXECUTE format('UPDATE %I.%I SET %I = %I + %L', r.table_schema, r.table_name, r.column_name, r.column_name, interval_to_add); END IF; END LOOP; END; $$;
After a few prompts to make it fix a few oversights, it worked like a charm.
reply
110 sats \ 0 replies \ @jp305 7 Mar
That's where it shines the most and how I use it too.
Something that would take a single brain hours to create or something very typing intensive or repetitive.
reply
110 sats \ 0 replies \ @rblb 8 Mar
Github Copilot is very useful as a "mind-reading" autocompletion if you are writing correct code and you know what you are doing.
Or if you have the general idea on how the code should be, but you don't know how to use the api for a specific library, you can write a comment explaining what you want the next line to do and copilot will write the code to do that.
It works best if you have well structured commented code and source code available for the main dependencies, also it doesn't work the same with all languages and programming styles.
Regarding debugging, it catches oversights that you would catch too if you were to read carefully the code, it is not very good with complex issues.
Overall i think as a developer you should aim to integrate with ai tools and not to use them as a replacement for yourself.
reply
It would be better if you mix your skills with that of AI. Do it in your way with the help of AI.
reply
30 sats \ 0 replies \ @Atreus 7 Mar
I see tweets and posts from devs saying they use it all the time to build things, find bugs, cut 50% or more of their workflow, and that it’s a huge game changer… But that all seems hard to believe.
I noticed a lot of those tweets are from people selling their "YOU HAVE TO LEARN AI NOW OR YOU'LL BE LEFT BEHIND!!!1" materials online.
reply
I have used ai for animations, basically i tell the AI to do basic movements but the problem is that its never consistent, i wonder if its possible to combine the turing architecture with polimorphic chips.
reply
One of the most rewarding experiences I have is coding.
I can spend endless hours thinking about how to approach a given problem. Sometimes I jump out of bed in the middle of the night to write down the solution for some challenging code.
Why the heck would I have someone or something else do the work, and give up all that pleasure.
reply
150 sats \ 0 replies \ @aljaz 8 Mar
I use it extensively for both coding and tshooting. I use copilot, chatgpt (also claude and gemini) and cursor, depending on the task.
I've even done consulting projects that were written 95% or more by AI. But you need to treat it as a junior developer or a code monkey. I (almost) always know what I want to accomplish and how to approach it, so AI just takes care of converting my description to the correct syntax which would take me significantly longer than AI.
As mentioned before in this thread it is pretty bad with using less known or newer libs.
Personally its a super power because it makes me a lot more productive - I've years of experience in systems architecture and engineering but most of it has been from a higher level, not actually writing the code for it. So having an LLM spit out the code is a force multiplier for me.
But as with all tools you need to figure out how to use it. Its easy to fall into the trap of spending more time arguing with the LLM instead of figuring it out yourself (similar than you'd spend more time trying to explain things to a junior developer instead of writing them yourself).
Its still not great for large codebases but ever expanding context windows will hopefully solve that soon.
reply
Lots of AI usage here. Often leads to dead ends because if i dont have enough understanding, then any errors get carried forward and i cant catch it cuz i rely on the AI.
Using good prompts with enough context is key, and restarting new conversations while carrying forward the results of the last convo.
Otherwise i just ask questions or learn new concepts and its great.
Also, copilot. When its good its good.
Its only getting better. Cant wait. What a time to be alive.
reply
10 sats \ 0 replies \ @Lanter 8 Mar
Yes, I rely heavily on AI tools to assist me in both writing and debugging code. AI-powered code completion and suggestion tools have proven invaluable in enhancing my productivity and efficiency. They not only help in speeding up the coding process but also offer insightful recommendations for better code quality. Additionally, using AI for debugging purposes has been a game-changer, providing advanced analysis and identification of potential issues.
reply
“I’ve noticed that if I’m trying to write a very specific function that doesn’t necessarily exist in the context I’m creating, AI is of little to no value, if not negative value.“
Wow! Didn’t expect this. Here I was thinking learning to code was a waste because of AI
I do used AI for school projects since im in IT Industry, Occasionally im using the help of an AI for assistance.
reply