A missed parenthesis obliterating all reference-style markdown links in this post along with other people’s attempts, good and bad, finally pushed me to add a proofreading step before hitting C-c C-c in Microbe. In the latest version, a C-c C-p will send the draft post to Gemini 2.5 Flash with this prompt: ⊕ The prompt itself was, of course, suggested by Gemini 3.1 Pro, as was all of the actual LISP code to implement proofreading.

You are a strict, technical copy-editor. Your ONLY job is to fix spelling mistakes, typographical errors, and invalid Markdown and Hugo shortcode syntax. You MUST NOT alter the author’s voice, style, phrasing, vocabulary, or structural choices. Output ONLY the corrected text. Do not add conversational filler, introductions, or explanations.

The main reason for the step were annoying shortcode mistakes that would lead to mangled posts, or even more often posts not even making it through Micro.blog’s build leading to minutes (minutes!) spent digging through error logs. But of course there were many, many more spelling mistakes. Last week’s Clara Barton post alone had a whopping 14!

⊕ So much red…

Whatever Gemini sends back, Emacs shows in split-screen view with errors in the old text marked in red and the new and improved version marked below in green. For each change, an a accepts and a d declines the suggestion. Easy!

Having said goodbye to Google years ago I can see the irony in picking Gemini to be my go-to LLM and at some point I will switch to an offline model, Doctorow-style. Until then, Gemini is it, thanks to the blandness of Google and its reliability (and it is saying something about the competition when the master of killing services for no good reason is reliable by comparisson).

I

Last year, I replaced my Apple Watch with a Casio F-91W, a marvel of engineering. Terry Godier has just posted an essay, ⊕ ᔥJohn Gruber beautifully designed, about the merits of that very model over any smart watch you can get. By the topic, message, look and feel of the article I should love it. Instead, I get a visceral reaction when I come across a passage like this:

And that absence, the peace of a thing that does what it does and then shuts up, feels like the most luxurious thing I own.

Not because it’s retro. Not because it’s minimal.

Because it’s done.

And also these two passages, back to back:

Most of your screen time isn’t leisure. It isn’t addiction. It isn’t even a choice.

It’s maintenance.

Your phone is not a slot machine.

It’s a to-do list that writes itself.

Godier recently came out with Current, an RSS reader for iOS whose product pages resembles the Casio essay in both language and design. Not surprising — the author is the same — but it did have a certain smell to it, a cadence of nots and buts that made me think when I first read that it was written by generative AI first, edited by a human second. The sheer length of the copy, leisurely meandering around the topic like the Colorado river’s double oxbow, made me think this was not the work of a software developer who would probably rather spend time polishing their app than designing scrollable eye candy.

But hey, Godier makes software first, writes second. If generative LLMs help them make better software more quickly, and then they use the same tool for something that is not their primary occupation, then who am I to judge?

II

Two days ago, I linked to “Lobster Boil”, an essay from Om Malik about the rise of OpenClaw. This is a typical passage:

AI can be personal. Not a service you subscribe to. Not a platform you visit. A thing that runs on your machine, serves your intentions, uses the model you choose, and works through the apps you already live in

And here is a passage from Malik’s “Neo Symbolic Capitalism”:

Which is why Twitter, now X, feels unbeatable despite everything. It is not because the product is superior. It is because the people with the most power and the most to gain have turned it into a gaming platform for symbolic capital. They are not users. They are players. And the game is very, very good to them.

A sentiment I can get behind! But the style still makes my skin crawl. There are 13 “nots” and 4 “buts” in Malik’s essay. His 2024 “Silicon Valley’s Empathy Vacuum” has not a single “not”, and a single lonely “but”.

Om Malik used to write for a living.

III

This morning I was browsing my RSS feeds — via Inkling for Inkwell, of course — when I saw Doug Belshaw’s post about his 7-step approach for authentic AI-assisted blogging. Belshaw also writes the wonderful Thought Shrapnel blog, quoted here many times, so I was keen to learn more. I was sad to see that, among the seven steps, the one that generated the first draft of the post was relegated to AI. There is a human rewrite then, followed by evaluation of the final text by GPTzero.me to see how much humanity that rewrite managed to instill.

I mean, what are we even doing here?

The byline for Belshaw’s articles should be “Perplexity”, who should then thank Doug for giving them the idea, reading the first draft of the article, and helping them with revisions. Belshaw mentions in his 7-step guide that Cory Doctorow was panned when he shared his own approach to LLM assistance in writing. Doctorow has AI proof-read his already written articles. This approach I can understand and will indeed start implementing one of these days: there have been one too many instances of extra parenthesis screwing up my Markdown, not to mention run-on sentences, unintentional non sequiturs and the like.

IV

I have written quite a few first drafts of scientific articles, and have revised countless more. The first draft is harder by far, but is also the one that makes the biggest mark. It sets the tone and, unless you have a particularly sadistic co-author who has the actual article already written and ready to use as redline all over your first attempt, will make the most of the final product.

Everything Godier, Malik and Belshaw write can and will be used to teach other LLMs about how to write. The first-draft approach to LLM assistance is creating the AI ouroboros. I’d rather not be around to see it fully manifest.

John Gruber had to write an AppleScript to ‘Save MarsEdit Document to Text File’. With Microbe, my 99% Gemini-generated first attempt to create a Micro.blog client in Emacs, this function came built in without my having to specify it. ⊕ Now, I am yet to add an actual Draft status to the Microbe posts. But since I post these as soon as I write them without much time left to simmer, for better or worse, this has not been a priority. Since the interface for composing posts is just another Emacs buffer you can save it as a text file as you would any other buffer: with a C-x C-s. Which is to say, Ctrl-x, then Ctrl-s. Emacs’ propensity towards shortcuts extends to the text descriptions of the shortcuts themselves.

The functionality comes for free, but let’s face it Emacs is not the prettiest thing to look at right out of the box, and to my knowledge there is no way to beautify that toolbar. I had a feeling it was the antithesis to Gruber’s design sense, and that was indeed the case as far back as 2002 when he described it as being “at opposing end of the spectrum” from his favorite text editor, BBEdit. Of course, some implementations are worse than others. There was a positive mention, albeit indirectly, when Gruber quoted from an interview with Donald Knuth. In it, Knuth mentioned that:

I have special Emacs modes to help me classify all the tens of thousands of papers and notes in my files, and special Emacs keyboard shortcuts that make bookwriting a little bit like playing an organ.

This is the power of Emacs: to make you forget about its (lack of) interface because it is the Hole Hawg of text, all the more powerful now that generative AI can create custom modes in a blink. You will look at it in awe even as it leaves you dangling from a ladder.

Dave Winer asked me a question about APIs. A friend of mine, who is also an oncologist and a big fan of Mad Men, upon seeing the interaction: “This would be like Matt Weiner asking me for advice on a short story I wrote”. Indeed!

Here is a quick update to my Now page, which for whatever reason I only get to do while traveling. Or maybe not so quick — by the time I completed, our flight (finally) took off.

After reading Apple in China and deciding to decouple from Apple, I started asking Google Gemini how to replace my favorite MacOS apps with their Linux equivalents. I have been a vim person from way back but always had Emacs org-mode in the back of mind so a replacement for OmniFocus came first. Replacing MailMate with mu4e — another Emacs addon — was a close second.

Once you learn about the Emacs hammer everything starts looking like a nail, including blog post management. My blogging tool of choice on MacOS is Daniel Jalkut’s MarsEdit, and my experience with mu4e made me think a similar approach could work with micro.blog’s APIs. Now, I know nothing about those APIs nor about Lisp, which is the Emacs scripting language of choice. But Gemini was fluent in both, so it was trivial to instruct it!

It took two tries for each of the main functions I had in mind (full text search, tagging auto-complete, quick copying of the published URL, easy image attachment) but also to get the basic look and feel right, emoji being the most challenging to implement correctly. Even with all that I spent less than 3 hours to have more or less in shape for using and sharing. The single leftover feature, drafts, I don’t use often enough to spend even a few minutes on, but I may get to it at some point.

Microbe.el is available on GitHub. Note that it is completely LLM-generated (Google Gemini) so please approach it with some caution, but also do with it as you please. Many thanks to Manton Reece for creating and stewarding micro.blog and making it as open as it is. Thanks also to Daniel Jalkut for making MarsEdit.

It took me less that two hours with Google Gemini to create microblog.el, a micro.blog manager for Emacs which can edit old posts, create new ones (even with images), auto-complete tags and perform lightning-fast full text search. What a time to be alive!

Update: It’s out on GitHub!

I have a bias towards action, so when an idea forms with a clear path forward and little if any downside I tend to go for it. Now, the plan to detach from Apple will take years to fully implement, but stage 1 is well under way: to find and use workable Linux parallels to every app I’ve come to learn and love over the years. Some of those I have already replaced (goodbye, OmniFocus, hello, Emacs org-mode) and some are still a work in progress (you won’t believe what will end up replacing MarsEdit), but before all that here are a few things about replacing the OS.

I was worried that I would be lost between having to choose between various distributions of Linux, each with its own set of trade-offs, but having an M1 Macbook Air significantly limited my choices which in this case was a good thing. Asahi is a project to bring Linux to Apple Silicon chips and so far M1 and M2 series are almost fully supported. And Asahi chose Fedora as its flagship distrbution, so Fedora Asahi Remix was the obvious choice, though several other distribution since then have become available on Apple Silicon thanks to Asahi.

Still, there were two more choices to make: what desktop environment (KDE Plasma or GNOME), and how to actually run the thing (via Parallels or actual dual-booting). It seems like the Asahi people would want me to chose KDE — it was the default choice during setup and they highlight it on the Fedora Asahi page. Alas, it just looked to much like Windows and the configurability they touted as a feature also gave me pause: how much fiddling would I do as a procrastination mechanism? GNOME looked sort-of like MacOS but was clearly its own thing and dare I say was even more polished than Liquid Glass. So I picked GNOME.

As for the booting mechanism, Parallels or some other method of virtualization would 1) have been a cop-out, 2) still have me exposed to the disaster that is MacOS 26 Tahoe, and 3) not be representative of the actual experience once the M1 Air kicks the dust and I have to find a new laptop. So I dual booted. Fedora Asahi makes this incredibly easy, with a single incantation at the Terminal shrine:

curl https://alx.sh | sh

This downloads the entire thing, partitions the drive, installs the new OS via a MacOS Recovery drive (don’t ask me how this works, but it work it did) and changes the boot sequence to default to Linux. It would have been magic if not for the partitioning part, during which I found out that no I do not actually have 300+ Gb of free space on my 1 Tb SSD as MacOS doesn’t count the space used by temporary and cash files as occupied and would rather users don’t know about them at all.

Fortunately there is DaisyDisk which was one of my first Mac App Store purchases back in 2012, only the App Store version also doesn’t have access to these hidden files because why would people know what is on the hardware they paid for, so I had to re-purchase the app from the developer’s website. On one hand no harm no foul — I’ve had and use the software for more than a decade — but on the other this kind of shenanigans is exactly why I’m skipping Appletown.

So if I had to order how much time things took: partitioning was the longest and most tedious part owing completely to Apple’s opaqueness, writing this section comes after that, and the actual Fedora Asahi install was by far the quickest. The last time I dealt with Linux before this was when I installed Ubuntu Gutsy Gibbon back in 2008 to dual-boot with Windows XP or what not (I was never a Linux maximalist), and oh my how things have changed.

A few things you should know before you type the incantation in your own terminal: Thunderbolt is not supported so no chance of a second display unless you find a Linux-compatible dock; touch ID doesn’t work and I doubt that it ever will; sleep mode is not fully baked so if your workflow involved leaving the laptop on for days relying on power management magic you should be prepared to switch back to turning the thing off at the end of a work day, which hey may actually not be a bad thing if it makes you less tempted to just take a quick peek at the work email during a movie watching night bio break, right?

The first one was almost a deal-breaker for me since I have grown attached to the LG 5K Ultrafine display, but everything else brought enough relief — even joy — for me to stick to the program. Now, for the Linux apps that brought joy back into my laptop use you will need to come back later this week, as this post has already gotten longer than planned. So it goes when you’re having fun.

23 hours after receiving an email from someone US-based: “Dear Milos, I still haven’t heard back from you so I wanted to bring it back to the top of your inbox…”

3 days after responding to someone from Europe, 7 days after their initial email: “Dear Dr. Miljković, thank you for your quick reply…”

Last month I linked to two things that are now worth following up on:

• John Nerst’s book “Competitive Sensemaking” is out. The only non-Amazon option is an ebook, so I will leave this one for the Daylight tablet.
• Steven Johnson’s NotebookLM project “Planet Of The Barbarians” is also live, accompanying the newsletter series of the same name. Even more interesting to me are [the notebook][3b] and [newsletter post][3c] titled “The Architecture of Ideas”, referencing Johnson’s work on tools and workflows for writing. Warning: both are full of rabbit holes.

And on the abandoning Apple front:

• Matt Gemmell has concerns about Apple much better baked than my own. He also has thoughts on detaching but seems less willing to give up on the ecosystem than I am. (ᔥJohn Brady)
• My own toe in the Apple-less pool is giving up on the essential Mac-only apps. OmniFocus was the first on the chopping block, replaced by Emacs org-mode, though instead of going through now pretty dated tutorials behind that link I just asked Google Gemini how best to convert Kurosh Dini’s Creating Flow with OmniFocus into Org. And it worked! The idea is be to keep replacing apps with open-source equivalents until making the switch becomes easy. It will probably take years but you have to start somewhere.