I agree with you in the sense that if you tried to take any model right now and cram it into an iphone, it wouldnt be a claude-level agent.
I run 32b agents locally on a big video card, and smaller ones in CPU, but the lack there isn't the logic or reasoning, it is the chain of tooling that Claude Code and other stacks have built in.
Doing a lot of testing recently with my own harness, you would not believe the quality improvement you can get from a smaller LLM with really good opening context.
Even Microsoft is working on 1-bit LLMs...it sucks right now, but what about in 5 years?
But the OP is correct -- everything will have an LLM on it eventually, much sooner than people who do not understand what is going on right now would ever believe is possible.
You do also have to worry about exec and other neat ways to probably get around stuff. You could also spin up YAD (yet another docker) and run Claude in there with your git cloned into it and beyond some state-level-actor escapes it should cover 99% of your most basic failures.
The thing I am working on is improving at the moment agentic tool usage success rates for my research and I use this as a proxy to access everything with the cookies I allow in the session.
I'm 42 years old and it has re-ignited mine. I've spent my career troubleshooting and being a generalist, not really interested in writing code outside of for systems and networking usage. It's boring to type out (lots of fun to plan though!) and outdated as soon as it is written.
I've made and continue to make things that I've been thinking about for a while, but the juice was never worth the squeeze. Bluetooth troubleshooting for example -- 5 or 6 different programs will log different parts of the stack independently. I've made an app calling all of these apps, and grouping all of their calls based on mac address' and system time of the calls to correlate and pinpoint the exact issue.
Now I heard the neckbeards crack their knuckles, getting ready to bear down on their split keyboards and start telling me how the program doesn't work because AI made it, it isn't artistic enough for their liking, or whatever the current lie they comfort themselves with is. But it does work, and I've used it already to determine some of my bad devices are really bad.
But there are bugs, you exclaim! Sure, but have you seen human written code?? I've made my career in understanding these systems, programming languages, and people using the systems -- troubleshooting is the fun part and I guess lucky for me is that my favorite part is the thing that will continue to exist.
But what about QA? Humans are better? No. Please guys, stop lying to yourselves. Even if there was a benefit that Humans bring over AI in this arena, that lead is evaporating fast or is already gone. I think a lot of people in our industry take their knowledge and ability to gatekeep by having that knowledge as some sort of a good thing. If that was the only thing you were good at, then maybe it is good that the AI is going to do the thing they excel at and leave those folks to theirs.
It can leave humans to figure out how to maybe be more human? It is funny to type that since I have been on a computer 12h a day since like 1997...but there is a reason why we let calculators crunch large sums, and manufacturing robots have multiple articulating points in their arms making incredible items at insane speeds. I guess there were probably people who like using slide rules and were really good at it, pissed because their job was taken by a device that can do it better and faster. Diddnt the slide rule users take the job from people who did not have a tool like that at first but still had to do the job?
Did THEY complain about that change as well? Regardless, all of these people were left behind if all they are going to do is complain. If you only built one skill in your career, and that is writing code and nothing else, that is not the programs fault.
The journey exists for those who desire to build the knowledge that they lack and use these new incredible tools.
For everyone else, there is Hacker News and an overwhelmingly significant crowd that are ready to talk about the good ole days instead of seeing the opportunities in expanding your talents with software that helps you do your thing better than you have ever dreamed of.
I recently wanted to monitor my vehicle batteries with a cheap BLE battery monitor from AliExpress (by getting the data into HomeAssistant). I could have spent days digging through BlueZ on a Raspberry Pi, or I could use AI and have a working solution an hour later.
Yes, I gave up the chance to manually learn every layer of the stack. I’m fine with that. The goal was not to become a Bluetooth archaeologist. The goal was to solve the problem. AI got me there faster - and let me move on to my next fun project.
> I could use AI and have a working solution an hour later.
That sounds really cool. You should share what you used.
> The goal was not to become a Bluetooth archaeologist. The goal was to solve the problem.
I'm sympathetic to this view. It seems very pragmatic. After all, the reason we write software is not to move characters around a repo, but to solve problems, right?
But here's my concern. Like a lot of people, I starting programming to solve little problems my friends and I had. Stuff like manipulating game map files and scripting ftp servers. That lead me to a career that's meant building big systems that people depend on.
If everything bite-sized and self-contained is automated with llms, are people still going to make the jump to be able to build and maintain larger things?
To use your example of the BLE battery monitor, the AI built some automation on top of bluez, a 20+ year-old project representing thousands of hours of labor. If AI can replace 100% of programming, no-big-deal it can maintain bluez going forward, but what if it can't? In that case we've failed to nurture the cognitive skills we need to maintain the world we've built.
It has also led me to a career in software development.
I find myself chatting through architectural problems with ChatGPT as I drive (using voice mode). I've continued to learn that way. I don't bother learning little things that I know won't do much for me, but I still do deep research and prototyping (which I can do 5x faster now) using AI as a supplement. I still provide AI significant guidance on the architecture/language/etc of what I want built, and that has come from my 20+ years in software.
This is is the project I was talking about. I prefer using codex day-to-day.
Same here!! 39, working through a backlog of side projects that never got built before. The velocity is insanely fun compared to even two years ago.
Best part has been building stuff with my kids. They come up with game ideas, we prototype them together, and they actually ship. Watching them go from "what if it did this" to a working thing they can show their friends has been incredibly enjoyable (instead of them asking why I'm behind my computer again)
I always liked coding but honestly liked the end result more.
reply