Just added a bunch of visualizations for the data I track every day: GitHub contributions, workouts, steps, sleep, and more. Check it out: rgo(.)pt/activity Took a few hours to build, but now I'll have this data forever!
Rodrigo Dias
@rgo.pt
building https://openpost.social synthesizing https://edu.rgo.pt, working @ https://fe.up.pt, talking (web)dev, automation, linux, ai/llms, fitness
Thinking about repositioning OpenPost. Less "agentic social media scheduler", more "build your personal brand effortlessly". Would that be useful to you? What would it need to do beyond scheduling?
MCP going stateless is the right kind of boring. No hidden session ID, requests describe themselves, gateways can route by headers, and state becomes an explicit handle the model can see. Long overdue!
Coding agents are getting better at solving tasks, but the useful benchmark shouldn't be "can it solve this issue". It should be "can it keep a codebase healthy after the 5th fricking requirement change". That's where the current models still look subpar
UX is just the balance of trying to minimize both: - The number of clicks the user needs make to reach his goal; - The number of different things on his screen he needs to look at Get these two just right and you HAVE good UX.
DeepSeek is going to release a Kimi K3-level model that's 4x cheaper within the next two months. Then price won't be the issue anymore and the problem will be figuring out what to build with intelligence this cheap. Grab and ideas notebook ASAP!
The weird thing about reviewing AI code in a good repo is that you rarely find obviously bad code. You find code for features that should never have existed. The agent filled a hole in your prompt with an assumption that looked reasonable but didn't fit the product as a whole.
LLMs copy your codebase's habits. If the code is clean, they follow the patterns. If it's already a mess, they scale that mess frighteningly fast. We've gone full circle: maintain your codebase well, not just for humans, but because your coding agent is learning from it too.
Just got Codex to tune my NixOS config for better PC performance, and man, it is flying. Holy hell, Codex is so good. Yet another reason to use NixOS: every tweak is declarative and reproducible. These improvements are now part of my system, not changes I'll lose after the next reinstall.
A button that confirms your tap but ignores it is worse than a slow button. UI feedback is a contract. If you haptic-confirm input, honor the input. Users do not care that your animation had 300ms left.
Added a theme builder to Product Plate. You can tweak the colors, fonts, sizing etc yourself, or just keep hitting 🎲 until something looks good. Try it for your next project (SvelteKit!): https://productplate.pages.dev
DSpark is interesting because it is not “new model got smarter.” It is “same model feels faster because inference got better.” That matters. A lot of AI progress now is boring systems work: serving, batching, caching, routing, decoding.
Old hardware isn't automatically obsolete. A lot of it only feels* obsolete because the software got heavier. A 10-year-old laptop can still be a dev box, home server, writing machine, or dashboard. Most of the time the problem is not the hardware. It's the stack.
I open-sourced Product Plate. A SvelteKit starter for hackathons, product apps, and early SaaS ideas. The goal: stop rebuilding auth, billing, dashboards, forms, AI chat, deployment wiring, etc. and start at the interesting part.
The next big education product probably feels more like a game than a course. Courses deliver content. Games create momentum: clear goals, fast feedback, visible progress, rising difficulty. The future of learning is probably less “watch videos” and more “complete missions.”
SQLite is boring. That’s exactly why it’s great for mobile apps. When the network drops, the app backgrounds, or a sync gets interrupted, SQLite can keep state intact and the product feeling reliable. Offline support is only part of the story. Local-first thinking matters.
AI didn’t replace the hard part. It moved it. Writing code, drafts, plans, and docs is easier than ever. But knowing what to build, what to delete, what to trust, and what actually matters? That still takes judgment.
“Execution > ideas” is only half true. Execution compounds what the idea already is. A weak idea executed perfectly is still a weak idea. A sharp idea creates leverage before the first line of code or pitch deck. Ideas aren’t everything. Really. But they set the ceiling.
Tiny productivity win: I sync my laptop and desktop dev folders with Syncthing. This means I don’t have to commit half-finished changes just because I’m switching machines. Git is ONLY for meaningful checkpoints. Sync handles “I want to keep working from the couch”.
Vibecoding still isn’t a viable way to build genuinely well-engineered software. But most of the time, that’s not what matters. What matters is whether the software solves a real user problem, saves time, or removes friction. “Well-built” only matters after “useful” is proven.
AI is making your tech stack matter less. The best language or framework for a project should now be chosen solely based on the problem, not what the team already knows.
I prompted a custom TUI to manage my NixOS + nix-darwin machines into existence. Rebuilds, diffs, updates, formatting, linting, secrets, and cleanup - all in one place. Because "just run the right nix command" eventually becomes its own workflow 😅 Check it out @ github:rodrgds/nix-config
DeepSeek V4 Pro, Kimi K2.6, and GLM 5.1… Open-source AI is approaching frontier-level performance really, really fast. Not like I’m going to run any of these at home, but at least prices will be set by the market and not a single company.
Yes, configuring my NixOS machine wastes more time than it will ever earn me. Yes, I’m going to keep doing it anyway.
Learning is yet to take a complete revamp, TBH. With AI, you can learn so much more in so much less time. It's basically a teacher with infinite patience. I just don't think society has noticed the value of this yet, but it WILL happen.
Finally self-hosting my own FOSS social media scheduler. More updates are on the way; it’s still in the early alpha stage but already fully functional for all my basic needs.
Invest time in sharpening your tools and improving your workflow. Even if it doesn’t seem worth the effort at first, remember: it compounds over the months and years you do this.