Today I Learned

I want to keep up with AI as an engineer, not just as someone who uses it, so I've been learning with Matt Pocock's /teach skill. It turns a folder into a small course built on real sources. I plugged it into my markdown workbench and I'm working through evals, how transformers actually work, and what happens during inference. The back and forth and the little exercises are what make it stick. I guess I learn better when the AI quizzes me than when it just explains.

A few weeks ago, jev by typesafe.ai blew up the internet. I immediately picked it up and played around with it. Now that the dust has settled, I understand it much better and see so many great use cases that are obvious choices, especially for larger corporations. Jev feels like a great first step into AI for many companies.

In times when reading is more important than ever, but the signal-to-noise ratio is terrible, I keep coming back to these three reading modes. Awareness is my radar, retrieval is for looking things up, and connectome is the expensive one. This rewires how you think, and we simply cannot afford it for everything. I guess the real skill is the filter: deciding which few things deserve the expensive kind.

While setting up Omarchy, I also gave Herdr a quick spin. First impression: really nice and modern.

The idea is simple. Herdr runs as a server in the background, and my coding agents live inside it. Close the window and they keep working. A sidebar marks each one as working, blocked, or idle, so I don't have to go pane by pane hunting for the one that's waiting on me.

It feels a bit like Conductor, but for the terminal. I still would love to have something like Conductor for knowledge work.

Over the last few weeks I kept adding new tools to my life. Every single one made some task a bit easier. That was the whole point.

Then I looked at the pile as a whole, and it didn't feel simpler at all. More things to maintain, more little systems asking for attention, more mental load in the background.

What stuck: ease and simplicity are not the same thing. Ease is about one task, right now. Simplicity is about how much my life has to carry. A lot of complexity sneaks in through tools that each solved a real problem on day one.

So the question I want to ask before adding the next one: does this make something easier, or does it actually make things simpler?

Last week I put Omarchy on a bootable SSD and started turning my gaming PC into a little AI workbench for home.

The surprise was how good Linux feels right now. Not "good for Linux." Just good. Nice and modern.

What stuck, though, is the opinion. Omarchy is DHH's take on Arch, and it makes a lot of choices for you. That's exactly why it works for me. I love opinionated software. Somebody with taste already decided how the pieces should fit together, and I get to start from there instead of from a blank config.

For the past seven or eight years I've been a designer, and I've worked on software basically my whole life. Now I'm switching careers into AI, which pulls me back toward the informatics side that was always part of my studies.

It's a crazy time. And the thing I keep learning is how much design is coming with me.

It would be easy to pack all that knowledge away and start fresh. I'm learning the opposite: it's my angle. Design taught me to think in systems. How pieces fit together, how work flows, how a thing feels to the people inside it. None of that expires because the job title changes.

So I'm not starting over. I'm bringing years of design thinking into something new, and I'm really happy about that.

I keep sending people to "Attention is All You Need" and then watching the useful bits bounce right off. Fair. Dense paper, static diagrams, not much to grab onto.

Then I landed on Transformer Explainer. It's a real GPT-2 small running in the browser, with the whole stack laid out so you can type a prompt, drag temperature, and watch attention light up between tokens. Embedding, multi-head attention, MLP, next-token probs. Same architecture as the big models, just small enough to poke.

What stuck: for anyone curious how transformers actually work, this is way more useful than another abstract write-up. You learn by playing with it.

I came out of a long reading stretch this week and noticed something simple: my head had actually gone quiet. Mentally disengaged. Restored. That feeling shows up more reliably from a long reading session than from most of the habits I label as recovery.

And still, I almost never protect time for it. Reading keeps getting treated as leftover time, the thing that happens if the week has scraps left. I rarely put the stretch on the calendar on purpose.

So the reframe for me is pretty blunt. Reading isn't spare time. It's recovery. If I want that restored feeling again, the long block has to be protected like anything else that actually works.

Without a stop time, my days just keep eating themselves. The morning plan looks sharp, then work and personal projects quietly expand into every free hour. Suddenly it's late, I never chose to stop, and the day just took more.

What stuck: planning isn't only about what to do. It's also deciding when enough is enough. Mine is a bit of both. There's a clock time, and intentional planning still matters. Direction without a finish line isn't direction. It's just a longer to-do list in a nicer outfit.

The stop is what lets me actually enjoy the rest of the day.

Pokémon Champions pulled me in for strategy and mastery, then I noticed chess gives the same juice without a meta that keeps resetting. Competitive games land better for me when knowledge compounds instead of expiring. I played more than ten chess games in a few days and realized I enjoy the activity itself, even while I'm still bad at it.

Tried Superlist hard in August. Kept maybe 10% of it. The bit that stuck: quick capture, then every day I walk through active projects and pick 6-8 tasks that actually move something. Due dates got demoted from "organize my life" to "real deadline only." MCP is the surprise win, because agents can land in the same lists without me rebuilding a second task world. Once maintenance became the work, the app barely mattered. The ritual did.

I set up a few Grok Bots against my existing MCP stack in one session. Local news briefing, AI briefing, Monday Pixel Perfect Picks draft. Same jobs I already had, just scheduled.

What surprised me wasn't the output. It was how little the workflow actually changed. The agents did more of the work. I'm still in control. MCP just became the bridge: every piece of the stack keeps its own territory, and I can borrow the model intelligence without rebuilding everything.

Felt like real work, not another toy to poke at.

One week ago I set up my own memory layer for different AI models – inspired by Nate B. Jones OB1. The difference is already real. It's not magic, but it's genuinely useful. Having access to a shared memory across all models and all my projects is a real productivity booster.

Layers is a design tool, and more. I started using it for product work, then accidentally rebuilt the same layered thinking pattern for my personal life too. Turns out walking through seven structured layers is a really, really good way to actually understand a problem instead of just reacting to it. And with AI in the mix? My critical thinking has never felt sharper.

GPT 5.6 is genuinely different for knowledge work. It helps me build my LLM wiki in ways that feel new every single day. The tradeoff: I'm now officially one of the token bros, I guess 🤷‍♂️

I'm addicted to the choo-choo sound. I’m using Conductor heavily right now and agents now make this sound when they're done. I love that detail more than I should.

The bigger realization, though: context switching might be the hardest challenge we’re facing over the next year. Things move way, way faster now. Agents don't wait. You finish one thing, another is already done, and suddenly you're jumping between five different threads before breakfast. It's a lot.

And yet, I feel genuinely productive. More than I have in a while. I'm not sure those two things should go together, but right now, they do.

I've been nerd-sniped by personal knowledge management for a long time. Probably tried dozens of approaches in the last 10 years, and not a single one really stuck — the maintenance burden always killed it.

Then I stumbled across this concept by Karpathy. The core idea: you never write the wiki yourself. You drop sources in, the LLM reads them, extracts what matters, and integrates everything into a persistent, interlinked set of markdown files — updating pages, flagging contradictions, maintaining cross-references. You're just the reader.

That role reversal is what makes this feel approachable to me. Every PKM system I've tried eventually collapsed under the weight of keeping it current. This one offloads exactly that part. The LLM doesn't get bored, doesn't forget to update a cross-reference, and can touch 15 files in one pass.

I'm really excited to give this a shot and share my learnings here.

Yesterday at Config 2026, there was this tiny little announcement at the end of Skills for the Figma on Canvas agent. I think this is a huge thing since it opens up so many use cases that I know from other workflows. Imagine skills for a visual design critique, a preparation of the design rationale in slides, challenging questions, a hand-off checker that challenges the design from an engineering side of things. I'm really excited to try this out and build some superpowers for the agent!

Ideate → Brainstorm → Plan → Work → Review → Polish → Compound.

This is the workflow of the Compound Engineering Plugin from Every. I set up a small playground for it yesterday, and it has significantly boosted my productivity. I am ideating on things to build, brainstorming possible solutions, bringing them to the canvas, tweaking them, and planning the implementation. It not only feels productive; it truly is. However, the best part for me is the most mundane: the documentation.

I came across this video from Ridd. It showcases his current workflow for designing (or rather building) products using tools like Conductor, Paper, Claude Code, and others. You can clearly see a shift, with much more production work occurring in code rather than on the canvas. I'm currently trying out this type of workflow, and with my engineering background, it feels much more natural to me. I'm very happy that we have tools that support these processes.

Podcasts are one of my favorite types of media. Currently, I’m learning a lot about LLM architectures, model training, and the technology behind them. With NotebookLM, I can easily compile all my sources and generate targeted podcasts to listen to during my commute or while taking a stroll. It feels like I have superpowers nowadays.

I recently worked on a Figma plugin during an internal hackathon. Our goal was to optimize the hand-off process and incorporate AI into the workflow. However, we found it challenging to make a design "understandable" for AI due to the implicit knowledge embedded in it. We discovered that combining different layers was crucial. We serialized Figma layers, added screenshots, and provided some manual context. This approach significantly improved our understanding of gaps in the hand-off process and allowed us to deliver higher quality results

I just stumbled upon some open lectures from Stanford. It’s so crazy good that education like this is publicly available nowadays. One of them was about LLM training and the technology behind it. Transformers are the thing powering basically every modern AI model. You give them text, they break it into chunks, and then figure out how every word relates to every other word all at once. That context-awareness is the whole trick. It's why the model "gets" what you mean, not just what you typed. Stack enough of those layers on top of each other, train on enough text, and you get something that can predict the next word surprisingly well. That's it. That's the magic. At least to some degree.

I just stumbled on Codepen 2.0, which is currently in beta and it's finally a very modern way to build short little prototypes. I just love this way of working and I already picked it up around university time to build front-end focused experiences or little component snippets and make them easily shareable. Still looking for a good way how to combine that with an agentic workflow, but for now Codepen 2.0 feels pretty epic.

I started playing tennis last year in September. I'm playing somewhat regularly right now and had my first ranked match yesterday. I lost it very closely in the champions tiebreak. I'm quite happy with my performance, but when you watch Roland Garros in the evening, this felt like a completely different sport. I love tennis, but I also learned that it takes a long time and a lot of discipline just to be somewhat average. Keep on grinding!

During my vacation in Japan this month, I learned how deeply the Edo (1603–1867) and Meiji (1868–1912) eras still define the nation. This contrast really becomes incredibly present as you travel through different regions. Moving from the preserved samurai districts and quiet temples of Kyoto or Kanazawa (the isolated, traditional Edo era) to the westernized red-brick architecture and bustling port vibes of Yokohama or Kobe (the rapid modernization of the Meiji era) shows how beautifully Japan blends its secluded past with its industrial leap forward.

I cannot write prompts when I’m tired; the output is poor. For my side projects, I have found a different workflow that better suits my needs. I plan throughout the week, refine the plan from Monday to Friday, and then implement it with intent. This approach provides me with greater clarity and, more importantly, time to consider whether the project is truly worth building. Much of this thinking time does not occur at my desk!

I just stumbled over the fact that Command + Shift + 1 is actually a Macintosh shortcut to eject floppy disks. Well, I guess in 2026, nobody's using floppy disks anymore. That's why it's pretty safe to reassign that shortcut. I set it to capture text on my screen with CleanShot because this is one of my most favorite use cases. And hell yeah, this workflow feels good.

I love the convenience of applications like Conductor and the user interface on top of agent orchestration. However, I've recently found myself returning to the files, tweaking CSS, and adjusting pixels with small increments through many iterations. This workflow often disrupts my context in agent windows. I appreciate the new agent window from Cursor, and the transition to the editor window has become a workflow I truly enjoy. I start and plan with the agent, then iterate in the editor. It's the best of both worlds.

Copernicus is a space program from ESA and is focused on Earth explorations. The cool thing is that there is a newsletter that shares an image from Copernicus with you every day and it always has a nice connection to current happenings in the world. And it's a really nice newsletter that I love to enjoy every morning. Copernicus image of the day

I recently heard somebody sharing a very interesting prompt: Let's step back and think really hard: how can we make this simpler and dumber while still achieving our goals? This is such a good addition to my workflow and I'm using it 20 times a day.

In one of the recent "How I AI" podcasts, I came across a slide that intrigued me and sparked my thoughts. Now, I constantly wonder if this is truly worth my attention and whether I will improve tenfold if it genuinely delivers ten times the impact.

I always thought, based on the name, that it was a Japanese creation. However, it was actually developed by an American architect in 1979 and called "Number Place." Japan simply rebranded it and made it popular. Playing it regular recently and it’s such a fun little thing to do!

Figma keeps track of the last typography style you interacted with, which might seem simple but is quite useful. This means that by clicking on an existing text field that uses the desired style, I can speed up my workflow significantly. I just hit T and start writing, rather than writing first and then switching the style afterward. Little things like this add up.

I’m a sucker for keyboard shortcuts. One of my favorite new ones is the hyper key (✦), which basically combines ⌃⌥⌘⇧. Yep, that’s impossible to press. I combined it with Raycast, mapped it to my Caps Lock key, and assigned it to a bunch of commands to open apps. ✦F = Figma, ✦S = Slack, etc. Timesaver and endless possibilities!

I just stumbled over life-os by Gabriel Vadivia. It’s a super interesting, markdown based knowledge base, journal and diary and looks very interesting to me. I struggled a lot with creating something like this because I really overthought the folder structure. Just letting Claude run and maintain that in the background sounds so good.

I'm currently building this TIL section within Notion because it's just the easiest and most convenient for me. But every new entry requires a redeploy. I now set up a custom script with a redeploy hook on Vercel so I can redeploy my website just with a Raycast command. Big time saver.

I’ve been quite enthusiastic about keyboard shortcuts lately. I looked for a way to reactivate the mini player in the Dia browser while watching a video with another tab active. It turns out that pressing "I" on your keyboard in YouTube activates their mini player. This also resolved my Dia issue!

I stumbled upon the Chrome DevTools MCP and set it up with Claude Code. It's especially useful for making small updates or running experiments on an existing page. This feels great. Currently, my biggest challenge is making quick iterations on existing pages with AI when I don't have direct access to the code. This remains a significant issue in large companies, where there's a clear separation between designers and developers, or in unconventional agency setups. Overall, this is a pretty cool solution. Now I just have to make it persistent somehow.