Tweet Disematkan
We did it!
Thank you to everyone who's supported me on this journey
💜
English
Dreams of Code
2.1K posts
Dreams of Code
@dreamsofcode_io
The most influential dev YouTuber you've never heard of. https://t.co/bFKErRq7p8 building @kiru_editor
Bergabung Şubat 2023
227 Mengikuti4.5K Pengikut
@Akbar501 I don’t disagree. I think it’s a classic case of “Theory of Constraints” coming into play.
English
@dreamsofcode_io He’s building enterprise class software, which requires specialized guardrails to maintain quality. We spent a lot of time and effort to get our guardrails right, and it’s honestly a non-trivial problem to solve.
English
I have no idea what I’m doing differently but all i see is massive productivity boost from using LLMs (properly).
Maybe it’s ADHD, maybe it’s knowing how to use these tools and work in parallel.
David Cramer@zeeg
im fully convinced that LLMs are not an actual net productivity boost (today) they remove the barrier to get started, but they create increasingly complex software which does not appear to be maintainable so far, in my situations, they appear to slow down long term velocity
English
@thefp4brain Yeah, I agree. It’s something that is very interesting because everyone has a different experience.
I’m currently landing on the idea that it’s “Theory of Constraints” which is causing this.
I’ve got a video idea from it at least!
English
What a lot of people get wrong is that it's not as simple as "if you dislike AI = you're just bad at it"
Imagine this being your first interaction with AI:
friend had an AI-pilled coworker who would generate 10k LoC PRs. And while I have my issues with it, this wasn't even the problem. The actual problem was they were bad at communicating and understanding what they did. So they would generate 10+ page pdf reports about it. So I'm sure you can see how that would slow down dev time for everyone (and, most importantly, disseminate the maximum LLMs=bad)
I love AI and use it daily, but interestingly enough, most of my bubble despises it. And I believe there is so, so much value in the bad stuff, the horror stories about AI and the productivity sink
These discussions are incredibly important because I have come to realize educating people about when to use AI and how to use it is a massive productivity boost for everyone. Because in the end, we all interact with people
English
It’s a learned skill!
When I started using agents I had little to no mental model at all, but over time I’ve gotten better at using the tools to fill that gap
Definitely requires taking an active roll in architecting and being proactive with understanding the changes.
I think the main challenge is that it goes from being a passive act through writing, to being an active action in order to regain it.
English
@dreamsofcode_io Agreed, my results are usually better the more I understand the codebase, otherwise it spirals fast.
It’s interesting that you can maintain a mental model of the codebase by reading diffs and chatting. I find it difficult to visualize the structure without working with the code
English
I also have ADHD and I find the dopamine hits to be pretty rewarding (although you’re correct it can be a challenge as novelty fades).
For me I think LLMs basically 10-100x throughput of code generation, but it doesn’t solve all of the other problems and bottlenecks that come in.
Perhaps that’s where we’re seeing the divergence, systems only move as fast as their slowest part.
English
the point of the quote'd tweet was that people are getting insane dopamine hits from the agents going wild, and in turn losing grip of their products and increasing complexity exponentially.
Re: adhd
i'm no psychologist, but I do have adhd so there's atleast some valid input from me on this. In theory, having adhd means I can context switch naturally accross projects which actually works out very well. The problem is if I lose grip of the work im doing -- not even opus 10 can get me intrested in doing something. I do not care for this cheap dopamine of watching myself get cucked by a clanker - I'm after that out-of-reach dopamine that comes from novel discoveries and overengineering things by myself.
neither of those are productive, so the conclusion is that the neurotypicals still win which is why people should stop calling adhd a superpower or whatever
English
@chikanma681 Yes!
I take an active roll in understanding and directing architecture / refactors still rather than just blindly accepting all changes.
My approach is, get it working, get it tested, get it refactored.
English
@dreamsofcode_io I think it depends on how you use. I handhold my agents a lot and works really well for me. However when I try to one shot everything like everyone on x talks about it fails
English
@Road_Kill11 I agree, but I don’t think this is unique to LLMs.
English
@dreamsofcode_io it's also size of project imo
smaller/newer projects you will be way more productive
English
Yes, 100%. Also more time to experiment with ideas and throw them away.
I agree. It’s exactly like having a small dev team. I used to be a tech lead / EM and agentic engineering is very similar but with some pros and some cons. The same rules apply in my opinion.
Maybe that’s what’s causing the divide, although I would expect OP to be have more of a manager role.
English
Do you have the feeling that you have more ideas than time/energy to implement?
I feel like I've been given my own small dev team. I'm the planner for them, and they do as I want, even if it's sometimes a silly request. My team kinda sucks sometimes, so it needs a bit of oversight and policies put in place, but it's such a net gain for me.
English
I have a pretty good comprehension of the codebase even though I’m hardly ever in the code. I usually check the git diff and ask the agent about decisions as well as performing refactoring with the Agent to direct the code how I want.
Things can definitely spiral if you don’t take an active roll in architecting, but it’s easier than ever to refactor imho.
English
@dreamsofcode_io on the spectrum of “never read the code” to “spend most of my time in the editor”, how hands-on are you with the code when using LLMs? this seems to be the biggest difference between people in terms of results
English
@andrewcairns Part of me has nostalgia, but it’s so slow compared to using LLMs even with the need to refactor periodically.
English
@dreamsofcode_io Same. Never going back. Couldn’t if I tried
English
@levelsio @jeremyphoward Bruh, how you living like this.
Name your sessions.
English
Example of how it's complicated
How to attach to my 2nd tmux screen:
> tmux ls
0: 1 windows (created Mon Mar 16 10:23:41 2026) [220x50] (attached)
1: 2 windows (created Mon Mar 16 11:45:02 2026) [220x50]
2: 1 windows (created Mon Mar 16 14:12:55 2026) [220x50]
> tmux a -t 1
Sorry it's just not user friendly
Like tmux attach 1 would be user friendly
English
I hate tmux
It's so incredibly user unfriendly
The shortcuts make no sense
I wish someone would make a better tmux
Even just logging into tmux attaching the screen is an illogical hell to type
Again I hate tmux, it's so shit
Matthieu Richard@SpaceMatthieu
@levelsio Is there a good way to jump between tmux sessions on Termius? I find it quite hard to manage multiple codex/claude sessions on the go
English
@nerder_ Yep.
Concerning trend that more and more companies are adopting
English
Bro wtf is this.
Why are people complaining about forking open source software.
Ridiculous trend.
Malte Ubl@cramforce
English
@mattstellisoft I agree with 1, but not 0.
It’s impressive what I can achieve but compared to how much I still have to do, it’s got a long long way before it can remove me.
Maybe in some domains, but definitely not what I’m building.
English
@dreamsofcode_io AI is solving that problem by reducing the size of teams to 1 or 0.
English
I’m convinced A.I. makes writing code on a larger teams much more difficult.
I expect we’ll see that more code needs to be isolated and perhaps even single owner, with well defined contracts (and tests) to prevent things from breaking.
Basically, same problems that exist already but 10-100x
Jay@jayair
It's been an interesting week for us (it's only Tuesday) We are figuring out how to deal with the effects of AI on our engineering process And figuring out the impact it's having on our design process Funnily enough, it's been bad In both cases
English
@noself86 Yes. 100% this.
The moving slow was easier to manage, but as you said, the issues were already there.
That being said, I did find comprehensive testing and contracts helped with moving faster, and this seems to be working for me now.
English
tbh i think the problem is that AI exposes what was already true but hidden. large team codebases were always fragile, we just moved slow enough that the fragility was manageable. the answer isn't more contracts and tests to manage the chaos, it's smaller teams w/ simpler code that AI can actually see end to end. the coordination overhead was always the real cost
English
@nathanj815 That’s one solution, but does come with a lot of other issues.
I expect more module / feature based and languages adding better support for this in the future.
Rust workspaces are very good for this
English
@Arrghtv Agreed. Large project engineering skills are still very important.
This is why I’m still bullish on the need for good engineers.
English
@dreamsofcode_io This will only get better once teams stop utilizing AI as the primary means of solutioning their problems and instead see it as a single piece in a much larger problem solving process.
English
@jdegoes @opencode @forgecodehq I’ve not heard of @forgecodehq before but looks promising.
The pair programming paradigm is the one I believe has the most potential long term (rather than full automation)
English
Unfortunately, @opencode's write tool consistently fails for large writes, giving agents no useful information, so they enter a doomloop not of their own making.
Time to give @forgecodehq a try -- it's written in Rust, not TypeScript, which is a nice perk!
English
You’ve definitely piqued my interest.
I do have a couple of questions:
How does this work with code signing the application bundle? My understanding is that it would break the bundle hash due to modifying Current.
If it doesn’t affect launch, isn’t there a risk for a supply chain injection or MiTM attack?
English
@dreamsofcode_io @seatedro restart to update nags should be banished in 2026. bad UX. this is how to do it properly: x.com/joelcompiles/s…
joel@joelcompiles
Here's how we implemented non-intrusive, automatic updates on macOS with support for multiple versions: Place all app logic in a framework bundle, make the executable a thin trampoline that invokes the framework entrypoint. Ship with a separate update helper app that you register with launchctl to check for updates daily. Update helper lays down the new framework side by side and then updates the Current symlink to point to the new version. After a few sessions without crashes the old version is cleaned up. You can add a rollback to previous UI experience if necessary, update symlink and restart. Works perfectly. Our version uptake is ~100% in 2-3 days.
English
