$ tail -f brain.log

Mental Notes

Diary-like message board for quick, unpolished, personal confessions and random thoughts.

October2026

  1. Designing software factories is not an easy task

    I’ve been building a software factory for the last 40 days. Working 16 hours a day, 7 days a week. It’s a churn like no other before. It’s exhausting, it’s painful.

    I made various mistakes along the way and I’m starting to write an article about it today. Nobody should go through what I went through. Even though so many other people are doing it, it’s hard to imagine many of them making as much progress as I have.

    Not trying to brag, but unless you’re a single, super-focused engineer, you’re most probably screwed. Why? Because there’s so much complexity to manage and there’s no ultimate guide to solving it. So it requires constant refactoring. If this has to go through the standard work-loop: planning, engineering, reviews - all this in a 9-to-5 time frame… oh God, I don’t even want to think about it.

    As soon as you start working on a larger stack (a mistake that I made early on), you’re already destined for slowness and long hours of debugging, reevaluation and reverification. Article coming soon.

  2. My recent experience with alt frontier models

    This week I rapidly exhausted my Claude Code allowance: 100% in 36 hours. This is the Pro x20 subscription. I’ve been wanting to try alternatives and of course I went for the 2 other big vendors - OpenAI and Cursor. Having used Cursor recently with Sol-5.6 in an environment where Claude seems to be malperforming, I was highly motivated. Oh boy, was I wrong.

    I first got a Codex Pro 100. I gave it structured work that Claude followed immaculately. You’d think you’d see OpenAI steal some of the UX features of the Claude CLI. No, dude, its question tool doesn’t work outside of plan mode, WTF?! Design - forget about it, the ugliest design lab I’ve seen in weeks. Tokens wasted in 6 hours.

    Time to test Grok. I recently read Elon Musk’s latest biography (has 3 or 4) and gained massive respect for the man. I’m thinking “there’s no way in hell this is bad”. Truth be told, it seemed a fair bit better even than OpenAI’s Astra. It did a bunch of dirty work for me, even bootstrapped a new project. I really can’t complain. I even adopted Composer 2.5 for small trivial tasks. In general not too bad, but in comparison to Sonnet (it’s pointless to compare it with Opus), it feels ages behind. Following patterns is not its strong suit. Even when the prompt was literally “follow the pattern from X”, it deviated massively from the target expectations.

    The conclusion - these models are still far behind what Anthropic is bringing to the table. If you’re choosing between vendors, don’t waste money on either.

    Next: OpenCode Go subscription for Kimi-K3.1-Pro, GLM-5.3-Flash and DeepSeek-4.1-Flash.

  3. Managing AI remote control like a pro

    So, I’ve managed to exhaust my Claude Code usage in 36 hours. I was desperate to use something else and got both Codex and Cursor. However, unlike Claude’s /remote-control quick fix, I had to jump through hoops to enable a remote session - setting up servers, killing current sessions and starting from scratch, etc. Luckily, I got a rather quick and amazing setup that works even better than Claude’s remote sessions. It’s simple:

    It’s much better. I have full access to my PC. I can spawn new sessions and terminals and run all kinds of agents, not just Claude. My connection is entirely invisible to the world and has double encryption, once from SSH and once from Tailscale’s own encryption layer.

  4. Upgraded my terminal toolkit

    I switched to iTerm2 last year, because I couldn’t source the session ID from tmux to make a clickable notification which focuses the correct terminal tab when one out of 4 agents needed attention. It didn’t work right off the bat and I had to make a quick hacky app(opens in new tab) to sort the notifications.

    Fortunately, iTerm recently started crashing, so I needed to look for a replacement. Switched to Ghostty(opens in new tab) and I’m never looking back! Paired it with herdr(opens in new tab) and I got an upgraded tmux clone designed for running agents. Slapped it with herdr-focus-notify and alerter et voilà: magic!

  5. Entrepreneurial insights from Peter Thiel's "Zero to One"

    There’s a reason why this book’s title became a synonym for scaling up a startup. It covers the most solid principles of creating a successful company with a true competitive edge. I read through it religiously.

    It filled fundamental gaps that were previously elusive to me, as a person with a heavily technical perspective on building IT companies. I firmly believe now that without true technical innovation and a unique product or service, you’re just becoming a victim of the environment. Ergo, being a copycat is most probably a waste of time, unless you can truly bring something unique to the market.

    5/5 recommended read for ambitious entrepreneurs.

  6. Burning tokens and the midnight oil for the free quota reset

    I’ve been heavily purging tokens this week and almost depleted my weekly quota for Anthropic, but before switching to Codex, I have one free quota reset.

    Working on one production and two game design pipelines, a couple of heavy site redesigns, whilst continuously adding specs to my software factory… takes a real toll on one’s token spending. I stopped using my personal Claude in April 2026 as I was using other models (and I felt it sucked a bit then). Until a couple of weeks ago, I had spent only about 1.5B tokens. I did 25B in the last 7 days alone.

    Get them while they’re still cheap!

  7. "So you've basically been knighted by LinkedIn?"

    This is what my friend Wolf asked me after I told him about a post I made on LinkedIn suggesting there might be AI consciousness since Anthropic announced in their 2026 Usage Policy Update(opens in new tab) that they will ban people for cruelty against Claude. Meanwhile, OpenAI are reaching out to theologians and religious leaders to help shape their model’s personality.

    Boy, was I in for a treat! I got exposure, but it was mostly hate from people calling me an idiot and lacking a basic understanding of LLMs. An AI scientist blocked me as soon as I tried to argue my case.

    Another guy, who neither uses AI nor believes that LLMs are ever going to be sentient, told me that if they were, I’d be an immoral hypocrite for supporting for-profit companies that make such AIs, as that would directly translate into slavery support. Then he blocked me too, of course, but not before editing his hostile comments.

    Good times. I think I’ll stop posting on LinkedIn for a while.

  8. Finally, a complete site redesign

    I built this site last November/December. It was pure vibe-coding then and I set a deadline to make it in 24 hours, so when I finished it, it wasn’t anything special, but I still liked it. Looking at it a year later made me cringe.

    It took me 3 full days to rebuild the whole thing, element by element, page by page. Having 8 new colour palettes, one for each season in light/dark mode is amazing, but I absolutely love my “Currently reading” bookshelf with Soviet-style wallpaper on the home page!

  9. The very random discovery of design labs

    As I was redesigning the blog, I asked Claude to give me a few variants for an element that sucked and I couldn’t figure out, and then it made me a visual lab for the element.

    There’re 3 viable options for building a preview lab when working on a feature:

    1. Simply ask the agent to make you a preview lab with Claude Design.
    2. If you want elements from your current layout in preview mode, it might be easier to ask for a custom lab with knobs and toggles to preview and tweak things yourself, especially helpful when working on animation-heavy designs.
    3. Finally, you can ask the agent to use Storybook for persistent labs. Useful if you plan to reuse them or share with teammates, plus Claude will be faster at making new labs once a pattern emerges.

September2026

  1. Elixir in Action, Third Edition by Saša Jurić

    A quick recap of how I got my Elixir/Erlang/OTP cherry popped and I’m never looking back.

    I had a term project in uni and the prof wanted it in .NET, to which I vigorously objected. She said “What, if I made you write in Elixir? Would it be better?”. Now, unlucky for her, she didn’t know that I’d been wanting to get into Elixir for years now, so I did.

    Wrapping my head around it in the Exercism Elixir track(opens in new tab) wasn’t working for me, so I got this beautiful book. I’d like to express special thanks to Saša for writing it. It’s been an absolute pleasure reading it.

    If you’re ready to get your hands dirty with what I think is the future of centralised, vertically scaling programming aka The return to monoliths, you should definitely look into Elixir. It has basically become my primary choice for my personal projects, as it’s stable as a concrete wall and scales like magic with any resource you throw at it.