# Retrospectiva #12 [![Portugal flags in the stands at Parken in Copenhagen](https://duarteocarmo.com/images/114/banner.webp)](https://duarteocarmo.com/images/114/banner.webp) First time at Parken - watching Denmark vs. Portugal We are getting dangerously close to the end of the year. 'Legra will soon be a full year old. She stands, has teeth, can almost walk. Slowly but steadily the household is getting into some sort of routine. Even though it's still a pretty chaotic and unpredictable one. I went to Lisbon for a couple of days and [spoke at Lisbon AI](https://www.youtube.com/watch?v=e7VYjMgMv2U) about some of my work with European Portuguese LLMs. Thanks to a great deal on a second-hand running stroller, Vitto and I are now back to running together. Didn't read as much long-form as I would like this month, stuck mostly with papers and articles. Using ----- **[Baby Jogger F.I.T.](https://web.archive.org/web/20151208002047/http://www.babyjogger.com/product/fit/)**: I never wanted to spend 500 EUR on a running stroller that I'd only use twice. I've heard numerous stories of babies that hate going for runs. So I found this one for ~80 EUR. I think this might be one of the best purchases we've made as parents. Allegra *tolerates* it. We've gone for multiple 10K+ runs with it. And now all three of us can go! **[Muesli](https://muesli.works/)**: We are all talking to our computers now. I've tested most of these apps. [Superwhisper](https://superwhisper.com/), [Handy](https://handy.computer/), [MacWhisper](https://www.macwhisper.com/). They all have pros and cons. Muesli has no cons. It's local. It's fast. It's customizable. It records meetings, it does speech-to-text. It's free, open source, and more importantly: it's 100% local. **[MoErgo Go60](https://www.moergo.com/pages/go60)**: I know I have a couple of reviews on my backlog for this blog. This is one of them. For now, all I can say is that the last time I felt I liked a keyboard this much was when I bought the [Happy Hacking Keyboard](https://duarteocarmo.com/blog/happy-hacking-keyboard-review.html). But this one is different. This one takes more getting used to. Slowly but steadily it's becoming my daily driver. Reading ------- **[Vision Language Models: Building VLMs with Hugging Face](https://www.oreilly.com/library/view/vision-language-models/9798341624047/)**: So many good names in the industry getting together to write a book can only be good news. Started this one this month and have high hopes. **[The Machinery of Life](https://doi.org/10.1007/978-0-387-84925-6) by David S. Goodsell**: Beautiful but extremely dense. It's not the longest so I'm trying to plow through this one just to tell myself I have finished it. LET ME LIVE MY LIFE. [![DeepSWE score versus RL cost and cost breakdowns for MiMo-V2.6 Pro and Flash](https://duarteocarmo.com/images/114/mimo-rl-costs.webp)](https://duarteocarmo.com/images/114/mimo-rl-costs.webp) The cost of large reinforcement learning runs. (From the Xiaomi MiMo technical report) **[Xiaomi MiMo-V2.6 technical report](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/MiMo_V2_6_technical_report.pdf)**: Read this one on the flight to Lisbon. A very well-written technical report that gets into the nitty-gritty. The RL part and distillation are the most interesting parts. Fun fact: the MiMo-V2.6-Distill-Qwen-9B distillation they made is my main local driver. **[Beyond Per-Token Pricing: A Concurrency-Aware Methodology for LLM Infrastructure Cost Estimation](https://arxiv.org/abs/2606.11690)**: Self-hosting LLMs on your hardware is not only about math. You can't just look at VRAM and token throughput. You need to (1) test your deployment and (2) get familiar with the use case. If you don't, you'd be better off with a cloud model. This paper tells you how much that oversight can cost you. **[Dear Congress: Here's What You Should Ask A.I. Leaders](https://www.nytimes.com/2026/09/24/opinion/congress-ai-sam-altman.html) (The New York Times)**: I agree with most of what Cal Newport says and writes. And I'm also pissed at all this fearmongering going around from Anthropic and OpenAI. In this op-ed, Cal tells the US Congress how to ask these companies the serious questions, not the stupid ones. Listening --------- **[How Will We Train Developers if AI Does the Routine Work? A Conversation with Scott Hanselman](https://www.infoq.com/podcasts/train-developers-ai-routine-work/) (The InfoQ Podcast)**: This is a topic that worries me. With increasingly powerful LLMs, seniors are more productive because they have scars from working before LLMs. What if you don't? **[AI researchers debate how close we are to recursive self-improvement](https://www.dwarkesh.com/p/john-beren-charlie) (Dwarkesh Podcast)**: A good conversation about the development of RSI (recursive self-improvement, i.e., AI making AI). I think some of these timelines are a bit too short. **[Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI](https://www.interconnects.ai/p/debating-rsi-the-us-china-gap-and) (Interconnects)**: A more grounded discussion to understand the other side of the RSI discussion. **[Better Know a Benchmark: ExploitGym](https://soundcloud.com/linear-digressions/better-know-a-benchmark-1) (Linear Digressions)**: Linear Digressions is a great technical podcast about AI/ML. I have absolutely loved their "Better Know a Benchmark" series. Each episode they deep-dive into a benchmark and explain all about how it works. Sometimes, I prefer to listen to an hour-long DJ mix rather than building the playlist myself. Some people just have better taste. Chris Luno and the people at Cosmic Dust Radio are just some of them. Watching -------- This video about the transition to becoming a father hit home for me: **[How Lauren Tan shipped 2,500 PRs in a month](https://x.com/poteto/status/2102050467505430555)**: Recommended watch on how to build "trust" with agents and how to parallelize more without losing quality. > here's how i shipped 2,500 PRs last month to production > > — lauren (@poteto) [September 21, 2026](https://twitter.com/poteto/status/2102050467505430555) I mean, who doesn't like a great backyard documentary?