Retrospectiva #6

Well, now would you look at that. Article 100 on this website. Ten years of writing on this small corner of the web. I'm writing this month's newsletter from my favourite place: the airplane.

Happy Easter if you celebrate. With clients spread across time zones, I didn't really get the time. Next week I'll be at AI Engineer London. Email me if you're coming too!

I'm experimenting with a slightly different format this week.

Using

Pi.dev: After watching this talk on how some people use Pi, I decided to (re-)double down. It's become my main driver. I've created a nice collection of skills (some private), extensions, and even ported my favorite theme to it. An agent harness is a bit like an editor: it needs to be yours. I want to /preview a longer agent plan in a nice web page. I want to answer questions in a nice interface with /answer, I want the theme to adapt to my macOS system theme. And, most of all, I want to be able to use it with ANY model. Even local ones.

Codex: Some things need my full attention. Others don't. When it's a quick fix, a breaking test, a simpler analysis, or something I can run in parallel, Codex has been my go-to. You can preview diffs, open in a terminal, create PRs, all from the same place. I tend to throw at it the smaller parallel things that are not "high risk".

Hevy: I continue to run. Not as much as I'd like. As you get older, weight lifting and strength training become increasingly important. After freestyling my gym routine for the past years, I wanted something a bit more sophisticated. After testing 3-4 apps, I landed on Hevy. It's paid but not expensive.

Snapseed: My inspiration to take photos comes and goes. But I need an editing app that makes the ugly Danish sky just a tad prettier. The previous version of Snapseed looked largely abandoned so I resorted to Darkroom. Now that everything is a subscription everywhere I started looking again. The new Snapseed app is excellent.

Reading

Empire of AI - Karen Hao: I tested a new reading routine this month: One chapter a day. Ended up finishing this one. The book has some interesting tidbits about the history of OpenAI and how it all came to be. It also has some slightly deeper commentary on the ethics of AI models and infrastructure build outs. Unfortunately, it had a bit too much TMZ-style tech drama I don't really find interesting.

Inference Engineering - Philip Kiely: I picked this one up on my Palma. Even though I just started, I can already recommend it. As I work with larger models, datasets, and infrastructure - it's a great field guide.

"If the cluster is not on fire, you are not using it enough"

An unsolicited guide to good research - Eugene Vinitsky: A great deck on how to do great ML research, and how to find interesting problems. I liked the framing: pragmatic and to the point.

How Kimi, Cursor, and Chrome Train Agentic Models with RL - Phil Schmid: An incredible roundup of the reinforcement learning techniques some of the larger players are using to train models.

Listening

The Roving - Bonny Light Horseman: What can I say? After clicking "don't recommend" on those Ruff songs on my Discover Weekly, sometimes Spotify eventually makes you discover something new and interesting.

Reverie - Jule: I managed to tone down meetings and put some nice focused stretches of work this month. And a good SoundCloud mix is essential to keep me company.

Confronting the CEO of the AI Company that impersonated me - Decoder: Long-term fan of the Vergecast. I love Nilay's interview style. I do think Shishir held his ground pretty well. An interesting discussion, if we set aside all the drama.

The Handyman of High Art - Rich Roll: I'm a big fan of Tom Sachs. The NY-based artist that inspired much of Casey Neistat's brutalist aesthetic. And I'm a sucker for brutalism. An interesting look at his life and work.

Watching

Everything We Got Wrong About Research-Plan-Implement: Loved this one. On how to tackle codebases with agents. Some interesting mentions of: "Hey - we let the agent rip, and actually regretted it. Then we went back, and thought about how to do it properly". Agents are likely to eat software, but we are still learning how to deal with it.

Living in the Matrix - Casey Neistat: From the master, extremely relatable. I also feel like chasing the green quadrant every time. But maybe the other 3 are a necessary evil.

The flight attendant is about to come by and scream at me for not closing my laptop. See you next month!