dodothebird

Let's face it: we need brain-computer interfaces.

best ive read this month. mariozechner.at/posts/2026-03-… > And I would like to suggest that slowing the fuck down is the way to go. Give yourself time to think about what you're actually building and why. Give yourself an opportunity to say, fuck no, we don't need this.

As an adult, you're going to get the urge to return to hobbies, like reading, that you used to enjoy as a kid... it is very important that you do so, and start doing things that bring you joy again.

if i ever try to write code by hand again

No-one has figured out how an eng team should work with agents yet. Be wary of anyone telling you they know how to do it. Keep exploring. blog.exe.dev/bones-of-the-s…

Full Moon by Maromaro 👌

When I worked in the Kubernetes ecosystem, I had the interesting perspective in that I had tried to write so many orchestration systems historically. The majority of them were terrible, but I understood the domain very well by the time Kubernetes popped up. Unknowingly I've done a similar thing with agents. OpenClaw is the bleeding-edge architecture of agents. But I have built so many agent frameworks. That I know this crap so well. All the dumb nuances. On the one hand agents are very simple. It's just a loop. But it's very similar to Kubernetes orchestration. It's just a reconciliation loop. But these loops in practice end up being sort of complex. I'm honestly really excited to work in this agent realm. I don't care so much about models. I take that for granted. But how do you build an agent on top of a model? It's almost like I have 20 years of experience that make me perfectly suited for this domain. And the agent is the new unit. We went from servers to virtual machines to containers. The new unit is Agent. Which is weird; that doesn't necessarily make sense but it's because AI is different. The new unit is not Sandbox; it's Agent. That is our new unit of compute. Infrastructure always had three tiers. Storage, networking, compute. We've added a fourth tier, which is models. The Agent is the unit that ties together storage, networking, compute, and models. With every unit there's some corresponding asset. A server is obviously a physical thing. A virtual machine had VMDKs, AMIs, OVF. Containers have Docker images, Docker files. An Agent is just a file system. It can be stored in git. It can be a zip. It's really just a collection of files, largely markdown files.

@pdrmnvd x.com/dodothebird/st…

my ability to produce code is exceeding my ability to read and understand it and i have no idea what to do about it.

Opus 4.6 and Codex 5.3 were released on February 5. They already feel ancient.

One unproductive AI discourse pattern keeps to be how individual workflows preferences are talked as the universal hallmark of software engineering. Group 1: A solo builder with agents, their preferred stack, and a pile of markdown files, working on their own apps, is the right way to build and everyone else ngmi Group 2: A much larger group building with agents at scale inside companies, where coordination, reliability, shared systems, and organizational complexity create a very different set of problems which most people don't hear about. tbh individual workflows can still be directionally useful to show new ideas, but they can also be not stable, and enterprises might have very different problems that individuals don't ever have. It's like why small startups don't need to or shouldn't operate like Google, but Google kind of has to operate more or less like company of Google's scale. All ideas are good but much of the AI narrative still very confidently comes from group 1 too little from group 2 (with few notable exceptions).

As agents become more capable, especially with models like GPT-5.4-High, writing code becomes less significant. I often create issues for a monorepo-wide project backlog, and even large code refactors turn into backlog items. I take time to reevaluate these tasks. If I still find them worthwhile, implementation is straightforward in a repository designed for agentic orchestration. This design focuses on writing clean code, documenting it, and adding feedback tools such as various tests, linters, and formatters. These are mostly practices we already know. The "agentic" aspect primarily involves skills, memory, and how we manage these in a persistent and deterministic manner within a repository.

Why Codex CLI defaults to GPT-5.4-Medium on Plan mode?

So it has come to this...

@leerob @mark_k One reason I consider Cursor for personal projects would be a polished desktop app, like Codex App, rather than VSCode slop.

@mark_k But wait, there's one more thing...

Are you going to give Cursor another chance, now that Composer 2 is here?

Before: a SWE at a bank could move to Google. Same skills: write code, review PRs, debug. Now? Skills at each tier are becoming GENUINELY DIFFERENT. Tier 1: Review AI code at massive scale, understand distributed systems, and contribute upstream Tier 2: Use platforms with guardrails, bring in fractional senior expertise for judgment Tier 3: Understand business problems, ship solutions fast, no deep expterise needed, but plenty of human interaction

With today's high bitrate update, the PS Portal is a good remote player device that does justice to the PlayStation brand. Although cloud streaming is not available in Turkey, the device is now quite usable on my home network.

@rohanvarma x.com/dodothebird/st…

Some interesting data we pulled today showed that ~40% of Codex users use multiple surfaces, between the App, CLI, and IDE extensions. Everyone seems to have a primary preference, but a bigger-than-expected chunk of users launch codex agents outside of their primary interface. If you use multiple surfaces, I'm curious why? And what could we do to improve the experience?

devs confuse "agent orchestration" with running multiple agents in terminal, let me explain this difficult subject: Orchestration: spawn, schedule, track, steer, reorder, queue execution Run 10 terminals: not orchestration

@jorgemanru Yes, I understand. It's indeed strange. I find OpenAI models to be superior, but perhaps Anthropic is A/B testing some new models😀

@dodothebird No, my perception had heavily shifted towards Claude again last week, before the new context window announcement. Highly anecdotal, I haven't done any kind of formal comparison here.

My very subjective perception from the last weeks: - Claude was ahead of Codex. - Codex suddenly became as good as Claude, sometimes better and faster. - Overnight, Claude is substantially ahead of Codex again, both in speed and performance.