# Retrospectiva #9 This month's Retrospectiva comes a little later than usual. The whole house got sick battling the flu the past couple weeks. Mixed with a small heatwave here in Copenhagen, means summer hasn't blessed us like we expected. At least not yet. Now that I'm feeling a bit better, I finally managed to get some writing in. Let's get to it. Using ----- **[GLM-5.2](https://huggingface.co/zai-org/GLM-5.2)**: Finally. An open-weights model that competes with the big frontier models. For the past weeks I've been using GLM-5.2 extensively. Through [OpenCode Go](https://opencode.ai/go) it starts at 5 USD/month. It's different than using your GPT 5.5 or Fable, but it feels good in the open-source way. I urge you to take it for a spin. It costs about one-fourth of what Fable costs. And it deals with 99% of issues Fable does. [![Artificial Analysis chart showing GLM-5.2 cost per intelligence index task](https://duarteocarmo.com/images/107/glm-52.webp)](https://duarteocarmo.com/images/107/glm-52.webp) GLM-5.2 catching up - [Artificial Analysis](https://artificialanalysis.ai/) **[Pi Stats](https://github.com/phun333/pi-infobar)**: You probably know by now that I'm a big fan of the [Pi](https://pi.dev/) harness. If you like Claude Code better, that's fine - enjoy yourself. One issue with Pi is that it's hard to know exactly how much you're spending. I discovered this little application called [Pi Stats](https://github.com/phun333/pi-infobar), which gives me a widget in my menu bar that shows me exactly how much I've spent and on what. Even though I still pay for a fixed plan every month, it's still important to know (in terms of real value) what's going where. [![Pi Stats overview screen](https://duarteocarmo.com/images/107/pistats-overview.webp)](https://duarteocarmo.com/images/107/pistats-overview.webp) [![Pi Stats languages screen](https://duarteocarmo.com/images/107/pistats-languages.webp)](https://duarteocarmo.com/images/107/pistats-languages.webp) Pi Stats in action (not my usage) **[Lazygit](https://github.com/jesseduffield/lazygit)**: We write less code, but we are accountable for more. Some of it is not worth looking at — but the important bits are. For the past few months I've experimented with a lot of "code diff viewers". I need somewhere I can review the changes the agent has made, and direct it in case anything goes wrong or weird. I've tried [Hunk](https://github.com/modem-dev/hunk) (and even contributed a couple of PRs: [#310](https://github.com/modem-dev/hunk/pull/310), [#347](https://github.com/modem-dev/hunk/pull/347)), and [codediff.nvim](https://github.com/esmuellert/codediff.nvim) in Neovim, which is what my `d+d` mapping calls. But lazygit paired with [difftastic](https://github.com/Wilfred/difftastic) has been hard to beat. Fast to iterate and great diff highlighting. **[Sofascore](https://www.sofascore.com/)**: I like watching sports. Not all sports, not all the time. But when I do (surprise), I like looking at the data. Sofascore is the app that gives me all of this - in a way I love. Who plays where? Who used to play where? When you're not 15 and playing FIFA extensively - it's hard to keep up. This team out of Croatia has built an incredible app! Reading ------- **[Deep Learning for Biology](https://www.oreilly.com/library/view/deep-learning-for/9781098168025/)**: I finished this one over 2-3 weeks. I recommend it. A great read in an area where there's much left to do! **[Reconstructing the Mind's Eye](https://medarc-ai.github.io/mindeye/)**: Some wild research out of Princeton that shows how they reconstruct images from brain activity. Yes. You heard that right. **[Scaling Laws, Carefully - Lilian Weng](https://lilianweng.github.io/posts/2026-06-24-scaling-laws/)**: What is the ideal amount of data given a certain model size? Given a certain compute budget? And vice versa? These are just some of the questions researchers have been asking themselves. Lilian writes about them beautifully. [![Chinchilla scaling laws comparison](https://duarteocarmo.com/images/107/chinchilla-2.webp)](https://duarteocarmo.com/images/107/chinchilla-2.webp) Image source: [Hoffmann et al. 2022](https://arxiv.org/abs/2203.15556), as shown in Lilian Weng's [Scaling Laws, Carefully](https://lilianweng.github.io/posts/2026-06-24-scaling-laws/). **[Inference cost at scale with napkin math](https://injuly.in/blog/napkin-inference-cost/index.html)**: A fast and short reference on how to think about serving models on GPUs. If you wanted to serve GLM-5.2 to multiple users, what would that cost? **[Investing with Agents - Lon Riesberg](https://lonriesberg.com/posts/investing-with-agents/)**: A very fun read from Lon Riesberg (creator of the awesome [Data Elixir newsletter](https://dataelixir.com/)) about using LLMs to make investments in the stock market. Listening --------- Not the most inspiring month musically. Being sick means less running, which means less listening. **[Machine Learning Street Talk](https://www.mlst.ai/)**: One of my favourite podcasts about Machine Learning/Deep Learning/AI. You should give it a listen. **[Skate Muzik: Listen to SADE - Worldwide FM](https://soundcloud.com/worldwide-fm/skate-muzik-listen-to-sade-14)**: I mean - Sade. Do I need to expand? Watching -------- **[World Cup 2026](https://www.sofascore.com/football/tournament/world/world-championship/16#id:58210)**: Norway just eliminated Brazil. Morocco eliminated Canada. Portugal drew with Congo and is playing Spain today. I'm not a sports fanatic, but I come from a place where football is very close to religion. **[Western States](https://www.wser.org/)**: When I can't run, I watch other people run. This year's Western States 100 miler was an exciting one to watch. Even though I was rooting for [Hans Troyer](https://utmb.world/runner/5883300.hans.troyer) to win this one, [Vincent Bouillard](https://www.wser.org/2026/06/29/2026-race-recap/) made an amazing run! Record broken, again!