Darren Shepherd

Announcing: DISCOBOT🤖🎉 My personal coding agent session manager. I built this for myself so I can vibe code faster. Watch the video, link to the project thread below. I'll be live streaming today at 10AM PT to fully demo it and talk about why I built it.

@mcsetty I agree, competition is good. Have at it, have fun. Just don't have any expectation of beating GitHub at their own game.

@ibuildthecloud Seems like it would just be good in general to have competition

Oh my gosh stop it, people. We don't need a new GitHub. If your only complaint about GitHub is that GitHub sucks, GitHub can just fix that. We only need a new thing if it solves a new problem that GitHub can't do.

@ifeanyi_we Because I develop products for people and a lot of people use Windows and I have to support them. And I actually think Windows is a pretty good operating system.

@ibuildthecloud Okay Linux guru, why are you letting windows make you suffer then?

Ugh, windows terminal has a special exception for docker-desktop profile. They don't auto load WSL2 profile. I can't do the same thing as a user. If I want a "system wsl" distro it will show up in terminal. I guess I'll just have to exploit that as a feature now.

@ifeanyi_we Do you know I'm like a Linux guru? Actually created a couple of Linux distributions. Ask ChatGPT "What can you tell me about RancherOS and the creator of it?"

@ibuildthecloud lol you don’t want to listen, you have to install Linux..

I'm not really sure we need anything more than vscode as an IDE at this point. I care less about an IDE and vscode does all the things and is infinitely customizable and super easy to extend.

I'll tell you my problem. The PR-based flow is not useful to me anymore. Centering my development around the GitHub model is too slow. I'm developing a new model. And I'll probably still put my Git repos in GitHub. And as stupid as this sounds, even create PRs. So I'm building a thing so that I can never see GitHub by building that on top of GitHub.

@ibuildthecloud Define the problem

@loftyPuma That's a good reason. Basically we don't need everything github offers, just git hosting. That's a lot like code.storage

@ibuildthecloud GitHub is built on git. Git is what isn’t keeping up with agents. We need a new git, which results in a new GitHub

@grono_dev Nah, that won't work out. People might think it will, but rewrite with AI has all the same problems as rewrite in general. You just hit the problems faster.

@ibuildthecloud I call this trend "fast software" (as in fast food, not performance): 1. Nobody wants to maintain and fix existing stuff 2. Just throw everything away and start from scratch 3. Ship it, call it a victory 4. Go back to 1

@hirschibar Nah you don't understand how hard it is to displace number one. They can absolutely miserably suck for the next five years and then turn around and fix all the issues and still be number one.

@ibuildthecloud But if github could solve that "github sux" problem they would have instead of MAKING that problem. Let's just make a new github

Woo hoo. I'm not the only one thinking like this. These stupid crustaceans are the future.

Announcing a small project I've working on: Argus, an open-source workflow viewer for DBOS Transact apps. I tried DBOS Transact and was sold instantly. Three client projects later, I'm still hooked: durable workflows in a Postgres you already have. No dedicated clusters, no cloud-vendor state machines, no k8s sidecars. (@DBOS_Inc is co-founded by Postgres creator Mike Stonebraker, whose grander ambition is to one day replace Linux with a database-as-kernel OS. Another story.)

@CodexHere Maybe. Nobody's smart enough to have evil plans. It's only a bunch of people who are too focused on their tiny world.

@ibuildthecloud yeah, like 99.999% sure: learn.microsoft.com/en-us/windows-…

I actually want my OS to sync to a cloud account so it doesn't bug me that Windows requires one. But it really shouldn't require one. They need to undo that. From a pragmatic perspective I know it holds you back and there are certain solutions you can't deliver but that's fine. The user just opts out of those things. I just don't understand who is in charge at Microsoft. I want to believe Satya cares but actions speak louder than words and the actions show that revenue is more important than customer. We need CEOs right now that have a purpose. The CEOs that listen to Wall Street will destroy their company.

@Anatoliy197755 LLMs don't need requirements. They need input. It's slightly different. Requirements are formalized. LLMs are interactive tools primarily. Don't think of them as a functional component within a larger process.

@ibuildthecloud LLM need requirements, writing them is vibe coding. Complex logic needs either long description or compressed meaning aka code like math. Some compressed info way is needed to comprehend and pass to LLM. My bet is on data eng script

We're definitely going to get to the point where handwriting code doesn't make sense. Similar to handwriting assembly doesn't make sense. But the analogy breaks down. Developers don't really know assembly and they can't really review assembly. On top of it, the assembly that's produced by compilers is so optimized and confusing that you can't even understand it. For AI the code that's being produced, the developer needs to understand it could have written it themselves but it's just not as efficient or safe to do so. AI is not a new abstraction layer. It is a tool to create something you understand. We still need people to understand the mechanics of coding. We still need people to learn these skills and these are important skills. And even with the power of AI, you can't get away with not knowing these skills if you want to be effective. As always, there'll be people who try to take shortcuts and they'll get short-term gains and then it will fall apart.

Oh yeah I want to cache too. But caching implies you know somebody else owns the data. I'm also not very clear on how idempotent load is. Resume is a foot gun because if the agent has another interface to it, like all CLIs do, how do you even figure out if you need to re-load versus resume? There's an updated date somewhere, I believe. Highly skeptical that anyone manages it correctly. I love the idea of ACP but it makes me very uncomfortable about how difficult will it be to actually support it. I already tried ACP once about three months ago and it was pretty difficult but that was three months ago, which is a lifetime.

@ibuildthecloud I should clarify, the client doesn't own it. But many cache it themselves and wanted a resume that was faster without reloading the whole thread, which I felt was fair

If you make a protocol and everything's optional, it's so annoying. I'm looking at you ACP. I find the semantics around session load and session resume to be annoying. The frustrating thing is it's not extremely clear who owns the data. I want the agent to own the data not the client. But if session/load doesn't exist that's basically impossible.

@hirschibar Intel should thank nodejs and ridiculously inefficient code.

@ibuildthecloud Yeah, the real reason intel is flying. Every agent is a multi tread system. GPU inference is needed, but so is raw CPU

I keep telling people the need for compute is going to absolutely skyrocket ignoring our existing need for GPU compute. The sandbox providers are going to be the new cloud. It's distributed. We don't need one centralized solution. We need a lot of solutions.

Isn't this fascinating because the thought is that with agents we're going to need so much more GPU. Which is true. But you find out you need a lot more CPU also because once you start doing tool use you realize how ridiculously inefficient all of our applications are and suddenly I need to run 30 of them in parallel.

This man is on to something. We are entering a future where instead of having a one-to-many relationship for compute-to-user, we are having a one-to-one relationship. This doesn't replace the existing solutions or even need to work for the existing workloads. This is a new use case.