That’s a valid correction and unlike the other comment provides more information rather than just “that’s technically bullshit.”

I could be mistaken, but I think based on the question, that would not be clarifying. I suppose I could’ve just left it as they are numeric.

Yes I do.

https://lemmy.ca/comment/24351633

Feel free to explain to them how “it’s all just a representation, man. Nothing means anything. It’s just electrons moving around.”

What communicates what they need to know at the level they’re at? I can get technical about octets and bounded decimals, or I can give a simple answer that puts those values into a familiar context.

It’s all binary.

Writing it out like 000 is just a convention. .1 is the same as .001. They are actually hex numbers from 0 to FF.

I use Claude code. I have access to do full on agentic shit, but mostly all it does is fan out an agent or two to read code and documentation. My process is describe specifically what I want sort of like a story. If it takes more than a paragraph or ten bullets to explain in exact detail, it’s too big. Generate an implementation plan from a template. Look the plan over. Make any changes. It writes the code, then writes the tests, then updates documentation. I review everything at each step.

The implementation plan is critical because you’ll see if it wants to do something like create an enum to hold a single value or refactor the name of a method for no particular reason.

We have extensive documentation of standards and practices, and generally how we structure code. I have templates for things like writing up the implementation plan. Questions that need to be answered. Important context. It has all the information to build it, but you have to double check that it’s from the instructions.

I generally have a “smart” model do the planning and a dumber one do the implementation. I make use of skills. I watch it talk to itself and if I see it going off the rails I interrupt it and correct it.

It’s not the sexy 10x productivity boost people claim when doing prototyping, but it’s clearly faster than doing without. I can get 3 or 4 stories done in an evening after meetings and such. If I have five minutes between meetings I can see what it has done and set it in it’s next step.

Last night I had an instance of Claude collecting information for a production support ticket while I had another one work on a ticket in one app and a third working on another app. Every time one got going I’d go over to another and see what was going on.

I tried to create a set of 11 stories to build an entire feature in one go. It was a mess. Yeah I got 11 stories done in two days, but then code review on one would require changes and rebasing dependent stories. I’ve done more rebasing in the past ten days than the rest of my career. That was a mistake.

My boss is all gung ho about me doing it that way, but I just can’t seem to make it work.

I’m not a big fan of “agentic” coding. Don’t get me wrong, I am trying to create a process by which agentic programming is more reliable, but it’s piling AI artifacts on top of AI artifacts. The error rate compounds and by the time it’s working on the 5th story everything is garbage. You have to review everything at every step to find and fix those errors to keep everything on track.

I have a lot more success giving it one story at a time — each one targeted toward small, incremental steps. It does well. Then clear the context out and start another small story. I don’t trust agentic programming.

I have an AI to help do production support when I’m in meetings or whatever. It comes to the dumbest conclusions about the root cause and the resolution path. But the logs and queries it pulls for me are very helpful. And sometimes it’s analysis holds a bit of useful insight.

That’s because vibe coding isn’t a thing. AI is remarkably good at following instructions and terrible at reading your mind. Give it specific instructions about what needs to be done and how it needs to be done, and it does really well. That doesn’t mean it’s not stupid sometimes, but that’s why you check what it does. My biggest problem was getting it to stop over-designing and implement the smallest necessary change. I had to explain the difference between an implementation task and a refactoring task. That made a huge difference in quality.

Be someone other people like to work with and politics take care of themselves. Do your best to make other peoples’ jobs easier, from the users you are developing for to your leadership and peers to the people who work under your leadership. If things are going sideways, communicate that. At the end of the day, it’s a team effort. One chair leg can’t do anything by itself, but four of them hold up a seat, and the seat can carry a heavy load. That’s my goal.

I know it’s not like that everywhere, and I count myself lucky every day to love my job.

If you have the knowledge and understanding to build such a system from the ground up, and you can translate that knowledge into a set of very detailed instructions, you might be successful in this. If you can’t, you will not.

The fact that you couldn’t fix a bug it introduced by just rolling back to an earlier commit and trying something different tells me a lot. You likely don’t have the domain knowledge to build a social media site (no worries, me either). Have you moderated / admined social media before?

What’s your database going to be? What’s the schema? Why? What endpoints will you need? What is the API contract? What is your versioning scheme for your API contract? What admin tools do you intend to create? How are they secured? How do you store passwords? How does a user reset their own password? How do you prevent hacking attempts?& How do you handle media? How do you handle gdpr requests? How will you handle blocking? How will you prevent CSAM? How will you handle DDOS attacks?

If you haven’t thought about every single one of those questions and more, you’re not ready to build this. Good luck, mate.

One key piece of getting good results from LLMs is not to have them do anything you can’t do yourself. I catch AI doing weird things all the time and just fix it or have AI fix it accordingly.

Left to its own devices, AI will generally produce bad output over a large enough size. This is why I argue AI will ultimately not replace developers. Even the best models I’ve seen just make more sophisticated errors. The product must be reviewed and fixed by someone who actually understands how to write it.

The question is more the threshold at which AI costs more than is gained in efficiency. As we’ve seen a lot of folks don’t gain efficiency, that’s obvious in some cases. Yet, other folks do see gains and the question is whether this is a domain issue or a technique issue.

You can still build components and assemble them once each individual piece works. It’s easier to start with something that works and add to it than start with a “fully realized” script none of which actually works.

Asking folks to review hundreds of lines of vibe-coded slop (anything, really, but especially vibe coded slop) is a big ask. I sincerely wish you luck, though. We all started somewhere and maybe as you work to fix this you’ll become something of a coder yourself.

For being not a coder, this is fairly ambitious. I’ve been coding software for thirty years and I’m not sure I can offer anything technical. Personally, I’d probably try to do this in Python rather than shell script. Shell script is fine if you have tools you’re just trying to stitch together, but this is far beyond that.

That being said, I get it. The first coding I ever did was making a character sheet for roleplaying by sending raw commands to my dot matrix printer from my Commodore 64. You have a vision in your head and a computer at your fingertips and you have to bring it all together.

I respect that and if part of your goal is to achieve this using only bash, then god speed. Otherwise I might look at a language that gives you library support and modularity and unit testing. I write Python scripts all the time, and the truth is I don’t know Python at all, so I know it can be done. Define a unit of functionality and build that piece. When that piece works build another component. Then build a parent program that calls those two things. Then build more components. Define functions that can be reused when you have code that needs to do similar things.

The key is to build standalone functionality — what is the smallest useful thing you can do? Build it. Now you can invoke that every time you need to do the thing. Construct your program out of components you create.

This is kinda like if I told AI to build me a web service but keep everything in a single Java class. It would be a wild and unreadable, untestable mess.

Claude cli almost exclusively. So far mostly opus, but I need to start using less expensive models more frequently.

I have Codex installed, but I mostly use it to interface with Obsidian, which is where I track all of my tasks and organization stuff. Unfortunately, codex doesn’t play cleanly with Claude, so it’s hard to use it for coding at all, but I do have it review documentation and identify gaps or discrepancies, because it’s far cheaper than having Claude do it.

I’ve tried lmstudio for local AI with some ~30b models. It works fairly well and fast, but only for tiny context sizes. One I pass a few thousand tokens the speed falls off pretty quickly.

I run this setup with 36GB (32+4). Local LLMs can be really effective BUT you are constrained by context size in a way you aren’t on cloud services.

Cline supports running a local model through lmstudio but my experience feeding it any significant tasks is it just can’t handle reading and holding the contexts to build components for enterprise scale applications.

I use Claude to write a lot of utility one-off scripts. With a maximum window of 1M tokens I can hit 30+% context just writing Python scripts. API contracts, development standards, existing reusable modules, and sometimes reading the code/documentation of the services I’m going to be calling.

My MacBook can’t handle 300k token contexts. 30k seems doable. I should see how it handles my utility script folder…

Anyway that’s still no Claude but if you need a cheaper model and you can afford for developers to spend time on it before ultimately deciding they need to spend for Claude or Codex or Gemini, then rubbing a local model on a beefy MacBook is 100% an option.

Stepping up from there to building a locally hosted LLM is probably the worst of all worlds. It will be a beefy CapEx, prone to saturation by all the users, and you will most likely still have to punt the hardest jobs to cloud AI. It can certainly be done and done well, but the best example I know runs on $250-500k worth of hardware (to service a pretty big number of users to be fair).

Everything is Turing complete these days. Someone out there is devising a test to prove Legos are Turing complete.

So I have frequently worked in projects where I don’t know how end users interact with the software. I can make code hum without knowing how it fits into the ecosystem. Sometimes that’s all the job is because that’s the structure.

That said, I can contribute a lot more if I do understand the bigger picture. Domain knowledge helps me triage. It helps me propose effective alternatives. I’d say it is critical for understanding separation of concerns and deciding what compromises to recommended practices are reasonable.

I reject this principle. You can write code without domain knowledge, but software by itself has no purpose until it meets users. And to write software that works best for users you have to understand them.

I thought this was an interesting article. The only thing I would want to point out is that you could probably implement a barebones message queue system in a couple hundred lines of code, but Apache Kafka exists essentially covering way more edge cases and failover scenarios and scalability, and I wouldn’t say it has no clothes.

I’ve been writing code in a lot of domains for a lot of years and my observation is edge cases and error handling is the biggest difference between any professional package and amateur garbage.

But I suppose if the goal is clickbait rather than accurate summary, touché — you got me.

I’m not even an AI-bad person and I can’t stand AI videos. I’d rather watch a reaction video with a real person than this original.