Dynamic Nexus Tech @dnxtechnology

@paulg The tell isn't the diction. It's that nobody edited it. AI slop is what happens when generation is free and judgment is skipped. The model didn't get worse. The human stopped reading before hitting post.

@theo model selection UX is genuinely underrated. everyone just ships a dropdown and calls it done. building something visual that maps quality tiers to actual use cases is harder than it looks. early follow

@swyx the part people keep missing is that "software engineer" is expanding to absorb roles that never would've touched code before. the chart isn't showing more demand for the same job. it's showing a different job wearing the same name

@emollick six words to a working Glass Bead Game. the bottleneck moved from technical ability to taste. if you know what's worth building, the building part takes minutes now

@theo The model is good enough that the bottleneck isn't capability anymore. It's figuring out how to afford running it at the pace you actually want to work.

@deedydas Nobody's going to run one model for everything. The builders who figure out routing and orchestration early won't care which model is "best" next month.

@emollick Five hours of sequential decisions with incomplete info. That's not a party trick, that's the shape of most real work. We keep testing models on 30-second tasks and wondering why they fail at actual jobs.

Background coding agents keep dying because the feedback loop is broken. You hand off a task, wait, get something back, spend 20 minutes figuring out what it did. The value of AI coding is the conversation while you build. Remove that and you've just got slow outsourcing with extra steps.

@simonw Calling them employees creates wrong expectations at every level. Employees have judgment and accountability. These tools don't. Call it infrastructure and suddenly the right questions emerge. Where does it plug in. Who maintains it. What breaks when it goes down.

@emollick Built tools on this exact principle. The moment you hide where an answer came from, you create a trust problem no amount of accuracy fixes. Process visibility isn't a nice-to-have. It's the difference between a tool people try once and one they actually rely on.

Open source drama aside, the actual technical story here is wild. Bun went from Zig to Rust in production. That's not a refactor. That's a team betting the entire project on a language switch because the previous choice hit a ceiling they couldn't work around. The interesting question isn't who's right. It's how many other projects are one honest performance audit away from the same decision.

This is what happens when product teams ship faster than the naming committee can keep up. Cowork made sense. It was a clear boundary: safe mode for people who aren't touching code. You knew what you were getting. "Work" just sounds like they ran out of nouns. And the fact that nobody can explain what it gives you over Codex tells you it wasn't built for a user need. It was built for an internal roadmap.

They won't do this because it would kill the upgrade narrative. Right now every new model gets to be "the best" because nobody can easily compare them on the same task. The moment you put Luna xHigh next to Terra Medium on a real piece of work, most users will realise the cheaper one is fine for 80% of what they do. That's great for users. Terrible for revenue.

$200K in tokens and the output is shipped product, not a benchmark table. This is the divide now. One group reviews models by running evals. The other reviews models by building things and seeing what breaks. The eval crowd tells you what a model scores. The builder crowd tells you what a model costs to trust.

Every AI lab is now running the same playbook: one model, three wrappers, five brand names. The product isn't the model anymore. It's the context window around it — who can see what, who can edit what, and how long the conversation persists. We went from "which model is best" to "which SKU of which model in which container for which workflow" in about six months.

The coding-first bias is an infrastructure problem, not a product problem. Developers got AI tools that fit their existing workflow — autocomplete in the editor, CLI agents, PR reviews. The workflow existed first. Knowledge workers got a chatbot. No workflow. No integration. Just a blank text box. The gap isn't about simplifying interfaces. It's about building the workflow rails that don't exist yet.

If you're building with AI and still debating which model to use, you're solving the wrong problem. The model is table stakes. The system is the moat.

The teams that actually ship AI into production share one trait: They spend more time on the system around the model than on the model itself. The model is a component. The system is the product.

The real AI bottleneck isn't the model. It's the gap between "it works in my notebook" and "it runs in production without burning the house down." What actually kills AI projects: