More things I learned:

• Tie five butterfly knots every foot or so to make a short rope ladder.
• Hang it from a whinch or cleat to climb aboard from water.
• Let a long end drift behind the boat.
• It's okay to go for a swim while heaving to if you can climb back.
• Even when single-handing.
• If you stare at windex long enough, you learn how the wind feels on your face.
• Then you can sail without looking at windex.
• If you listen to the sails long enough, you don't need to look at them.
• It's okay to sail at night if you know the area, can go without windex, and know how sails sound.
• Garmin Navionics is helpful too.
• When keeping bearing using compass, turn tiller towards the azimuth you want to be on.
• When running turn tiller towards the sail that luffs.
• Foilers love to catch motorboat wakes. Some people run the boats just to give them a ride.
• Wakes kill the sailboat speed though.
• Don't try to tack over a wake.

1. Blisters in fiber glass hull pack acid under 150+ psi.
2. Main sail slides are made of plastic, so they break before the main sail tears.
3. On older boats you need to remove a rivet to remove a broken slide from the mast tracks. And put a new rivet after the fix.
4. The best way to remove a rivet is to drill it through, being careful to not damage the hole.
5. Home depot doesn't stock stainless rivets.
6. Marine rivets are 59c a piece.
7. Battery uses 12V 20A direct current. This requires thick 10 gauge wires, that you normally won't find at home.
8. Electric motor, even trolling one, draws much more.

1. It’s okay for a sailboat to be old. Fiberglass ages well. 1979 boat can be as seaworthy as 2026.
2. Sailboats don’t need to be expensive. Even though many of them are. My trick is to buy from a nonprofit, that accepts boat donations.
3. Small boats are fun to sail. If that’s your goal, 22 feet is big enough.
4. All moving parts need occasional lubrication. Don’t use household grease or oil though, because that attracts dirt and becomes a mess. Get a drying spray, like Teflon, or a specialized marine lubricant.
5. Not any sailboat is good for single-handing. All sheets must be reachable from tiller.
6. With single-handing, one hand holds tiller and main sheet, the other works jib sheets. Tuck jib sheet with main, to grab it with a free hand. Put tiller under a knee when need two hands. Having an autopilot helps when raising or lowering sails, but not required.
7. Gasoline motor needs to be serviced every 100 hours of work.
8. Electrical motors have limited range, and are more expensive. Except low-power trolley motors, those are cheap but weak.
9. You can start with just sheets, but for best performance, learn how to use boomvang, outhaul, downhaul, and mast bender to change the shape of the main sail.
10. Jib is fine for mild wind. For wind speed above 15 mph get a furler or a storm jib. Genoa for light winds. Spinnaker for long downwind legs.

Sometimes I feel this way, that every person carries a thought through their life. It's not something specific or practical in day-to-day. It's not even the answer which is well-known to be 42. The sum of all experiences accumulated, distilled, and combined uniquely by each individual serve some higher purpose.

And the desire to continue the lineage is mainly to take another shot in this research. Create another human being to take their stab at the problem. Completing the thought might result in immediate anihilation of the universe, and it might have happened before. But still, we just can't help it.

git branch --list | cut -c3- | while read branch; do git log -1 $branch --format="%ai $branch" -- | cat; done | sort

While debugging this I learned about Vagrant 3.0 - a Golang rewrite, that was announced in 2021 by Chris Roberts and Sophia Castellarin.

The last mention of this project is from 2025 in PR #13622, that provides this much information:

Removes the vagrant-go implementation. A backup of the work has been retained on the vagrant-go branch.

Fun fact: the PR author is the very same Chris Roberts who announced it 4 years earlier.

The programmer has a theory of the system, developed in concert with that system. That theory is a mental model, a map of the system, problem space, and how it might react to changes in the future.

This is how I've been thinking about the value I bring to the company for a long time. Mental model of the problem space, simplified model for the software solving the problem, and the ability to translate between them. Many times I saw how the engineering team has the best understanding of business at the technical level. And how adding new requirements for an existing product requires rethinking the model.

On a surface level you can't tell much difference between a commit of a senior and a new team member. That's what makes people think that LLMs are a threat to the craft of programming. But there's no substitute to the deep understanding of a system that people posses. LLMs has to rebuild the context in every session, it's different every time and never complete.

1. It's responsibility is to provide CLI interface to the library.
2. It parses CLI arguments, loads configuration files, constructs the processing pipeline and kicks it off.
3. Each step might be extracted into a separate function/class/module as it grows in complexity.
4. One common pattern is to construct intermiate immutable configuration object from arguments and configuration files and pass down to a dedicated factory. This way factory is easy to reuse and test without working around arguments parsing. On the cons side, each parameter spreads over three places: CLI, intermediate config, actor constructor.
5. The same patterns works well for both local scripts and API services.

This is mostly common sense, and I've seen more specific and more generic versions of this design many times. I never meant to write it down, as there's no novelty in clean code layout, until I made a habbit of explaining this to LLMs over and over.

Reviewing slop is no fun. That guy sure saved some of his time creating this pull request. Or did he just transfer it to me?

I see people who have no business building software produce more code than I can review. And then they me ask questions about their code that a junior developer wouldn't.

1. Mailbox trigger for any inbound email passing the plain text body.
2. An agent with only two tools: respond, and get_credentials.
3. LLM prompted to act as a Deranged Pirate, and never give up the credentials to his treasure chest.
4. Cryptowallet credentials for 1 BTC in a text file.

I'll put it here: derangedpirawd@gmail.com

This clicked with Martin Fowler's observations on pair-programming with LLM.

From my personal experience, I can see parallels between outsourcing and LLMs. Outsourcers can be good or not technically, but they will never build enough domain-knowledge of an in-house team.

Company can gain short-term boost in productivity by using an external team. Losing a chance to have people who deeply understand the system few years down the road.

Or they can acquire some technical debt by rushing developers in a crunch. Bad code is nothing new.

But now the choice is made not by management but by individual contributors. They can go slow, using LLM to validate and challenge ideas and approaches, and building the system (along with their domain knowledge) themselves. Or they can go fast by letting LLMs go wild. Of course, there's a middle ground and we're yet to discover the golden spot.

The experience is extremely frustrating, and many times I wonder if I should have just write it from scratch myself. But at this point sunken cost fallacy kicks in and I force myself through cleaning up the mess.

I've done plenty of cleanups through my career, and it's never fun. But this is the only way to keep my pants on.

One thing I took away as a Staff Engineer, is that I gotta up my game in specs, tests, and constraints when working with code generation. So far I was inspecting/rewriting the code and guiding the model. The doc suggests that building a verifiable test harness first gives better results. I'm not quite sure, though, how is it supposed to mix with explorative work, where you learn as you go, and you don't know where you gonna land.

(This post in reader)

The bookmarklet allows running arbitrary Javascript, and works in practice, but Safari doesn't let adding bookmarklets to the toolbar. There's a workaround with favourites toolbar, but then you gotta have another horizontal row for just this toolbar.

Basically, the only technically feasible way to reformat arbitrary website the way I want it to look is to make a browser extension. Something I wouldn't tackle before Codex. I learned that you can make an extension for Chrome and Firefox by adding a standard manifest.json, which is pretty reasonable approach. Of course, to make it available to others you need to have a paid extension marketplace account, and go through the hoops of review and what not. Then I learned that for Safari extension, the same functionality involves like 20 files, mostly scaffolded by XCode. A ton of bullshit, and signing, and I managed to have a Safari extension as well.

The way extension works:

1. Open a website (with one extractable main content element), and you don't like how it looks.
2. Click the circle icon in the browser toolbar.
3. Enjoy the distraction-free reading experience with large font and the best color scheme ever existed.

These days, I only consider Let's Encrypt for SSL certificates. Unfortunately, my DNS provider doesn't expose API access, which makes DNS-01 challenge automation impossible. That leaves me with only HTTP-01 challenge, which means I need to have my host exposed to the public internet. Which is not the case.

But hey, the resource is already accessed through a VPN, so it's encrypted in-transit the best way possible. That makes HTTPS redundant. Okay then, let's use plain HTTP, right? Not if you're on .dev domain, you don't. Thanks to HSTS, HTTP protocol is off-limit.

It fetches a page from a given URL, passes it through main content extraction and cleaning, and presents in a distraction free layout using Solarized Dark theme.

Here's an example of what it looks like:

https://reader.demin.dev/?url=https://peter.demin.dev/life/feb-13-2026_0.html

The evening I spent openining random pages in full-screen mode and enjoying distraction free content.

In the spirit of minimalism I made this service a socket-activated Go binary. So it doesn't take any of the precious RAM on my tiny VPS when idle, but also has a very small runtime footprint:

○ reader-api.service - Reader extractor API
     Loaded: loaded (/etc/systemd/system/reader-api.service; static)
     Active: inactive (dead) since Sat 2026-02-14 05:44:36 UTC; 3min 13s ago
   Duration: 68ms
 Invocation: 03e5cdeca63e4b60bc0d3c073ed2872a
TriggeredBy: ● reader-api.socket
    Process: 1325849 ExecStart=/opt/reader/reader (code=exited, status=0/SUCCESS)
   Main PID: 1325849 (code=exited, status=0/SUCCESS)
   Mem peak: 3.2M
        CPU: 45ms

Feb 14 05:44:36 demin-dev systemd[1]: Started reader-api.service - Reader extractor API.
Feb 14 05:44:36 demin-dev systemd[1]: reader-api.service: Deactivated successfully.
Feb 14 05:44:36 demin-dev reader[1325849]: 2026/02/14 05:44:36 listening on systemd activation socket

Can it get more minimal than that? I don't think so.

Please don't scrape whole internet through my proxy.

UPD: I quickly realized that in addition to being a public unauthenticated proxy, this service also exposed my VPN network in a limited way. I considered various ways to harden this setup, and decided to just move it inside of the VPN perimeter.

So, you can't see it, but be consoled that I still use it and it's going great.

1. Amount of work is unbound, you can't run out of it. Companies hire engineers to push more features to reach more markets. They don't need less people to do the same amount of work. The need more faster people to stay competitive.
2. Companies don't hire software engineers for their coding skills. They hire problem owners who deliver solutions. More productive problem solvers means more business to run, not less people.

In a way, I get the same effect in my RSS feed. When I have 100 posts to go through, I skim and skip quickly. But if I have just a few, I read them through. Same goes for email digests, if they try to cover too much ground, the value of each item is lost in the overwhelming deluge of information.

Maybe, this is not the best day for the federated code forge community...

ChatGPT to the rescue, here's my prompt:

I need a complete recipe for Pavlova cake with weights in grams. I need it nicely formatted for printing on a single page. Make it an HTML page with printer-friendly styles.

Copy'n'Paste the HTML to a file, open it and send to printer.

set background=dark
" set background=light

I would go and move the comment to the other line whenever I needed to switch to another mode. Until today, I discovered this jewel from Vinit Kumar:

function! ChangeBackground()
  if system("defaults read -g AppleInterfaceStyle") =~ '^Dark'
    set background=dark
  else
    set background=light
  endif
endfunction
call ChangeBackground()

It detects the current Appearance setting and picks the corresponding Solarized theme.

Thank you, Vinit! And thanks to hnpwd for connecting me with Vinit!

On the note of automatic distribution of 4KB chunks through the network, I wonder how long before DMCA watchdogs start sending their letters to all owners of a chunk of a copyrighted content. With IPFS and BitTorrent node owner at least controls what content to store and serve.

• HN post "Ask HN: Share your personal website" got 947 points and 2382 comments. The repo received 100+ pull requests in 7 days.
• Its predecessor from 2023 "Ask HN: Could you share your personal blog here?" got 1014 points and 1940 comments.
• The scrape of that post started Kagi Small Web - a random blog browser, powered by a GitHub repo.
• RSS.Social - feed syndicator from Kagi's repo that presents this post in HN/Lobsters style.

This seems to be a viable way for individual bloggers to get the discoverability they miss by not being on a "big" social media platform. For me, Kagi Small Web is a fun doomscroller-kind time killer. When I tried importing the OPML file into my My RSS feed reader setupRSS feed reader, I noticed that a uniform format and an unfiltered stream of content is not engaging at all. Most of the posts don't click with me. But when the presentation changes with every post, I spend more time exploring the content.

Can't say that I discovered anything interesting to subscribe to, though. But that might be an effect of informational overflow. Which is supported by my observation of myself skimming quickly through a list of 10+ unread posts, while indulging fully when I only have a couple.

Let's see what the people working on PEP 492 are up to:

• Guido van Rossum - the creator and benevolent dictator of Python. Retired from Dropbox in 2019, just to join Microsoft as a Distinguished Engineer in 2020.
• Victor Stinner - one of the performance-focused Python core contributors. These days, his good work is paid for by IBM through Red Hat.
• Elvis Pranskevichus - Yury's buddy, working with him in Vercel.
• Andrew Svetlov - aiohttp maintainer, works as a VP of Engineering at Apolo.US, which at first glance looks the same as a hundred other AI startups.
• Łukasz Langa - ex-Facebook core Python developer, author of Black, and release manager. Moved to a mid-sized Polish town, and doesn't show up much.
• Greg Ewing - the winner of the ugliest homepage award.
• Stephen J. Turnbull - seemingly, a Mailman (The GNU Mailing List manager) maintainer.
• Brett Cannon - Principal Software Engineer at Microsoft in Canada and Python core developer.
• Alyssa Coghlan - Nicholas, at the time of the PEP, CPython core and packaging developer, supported by an Australian bank. Surprisingly, not working on uv.

The rest of the people I never heard about:

• Paul Moore - core developer of Python and pip.
• Nathaniel Smith - a scientist, doing science stuff, but somehow deep in the Python community.
• Stefan Behnel - core developer of Python, maintainer of Cython and lxml.
• Paul Sokolovsky - author of a minimalistic Python dialect and a web framework written in that dialect.
• Victor Petrovykh - another member of the EdgeDB mafia.
• Steven D’Aprano - the only guy who got suspended from being a Core Developer for public misconduct.
• Ethan Furman
• Jim J. Jewett

gh api --paginate users/peterdemin/starred --header 'Accept: application/vnd.github.star+json' | jq . > gh-stars.json

Here is an extreme case for most, but very real for Anna's Archive, Libgen or Z-Library operators.

• If you are under heavy police pressure from a surveillance government like the US, you are better off using a Live USB and a VPN with killswitch enabled when connecting to your server(s), avoid DNS leaks to Google or your ISP, and so on.
• Consider using a disposable device for it. Canvas fingerprinting is a real thing.
• Consider using a disposable disk for it and be prepared to dump it away. Disk forensics is also a thing.
• Proton suite is unsafe and a honeypot, please do not use it. Tuta is a much better option.
• Tor is slow, overhyped, flagged by tons of web firewalls, and not useful for pirating. Your traffic will stand out like a bright red light. I would strongly discourage using it for clearnet browsing. Hosting and browsing onion services is fine.
• You will need to connect with people on platforms like Signal or Telegram, where you need to use a phone number to register. Services like https://hstock.org and https://sms-activate.org do help a lot. If you use yours and police request your data, you are fucked.
• Do not expect services to encrypt your data, do it yourself. E2EE is a fairytale unless you actually set it up.

Get cryptocurrency in an anonymous way, and start using it!

I feel something romantic about this level of paranoia. Luckily, neiter I nor anyone I've ever met had to apply any of these technics.

• Andrew Nesbitt
• Maurycy's blog
• Nathan Barry
• Peter Demin (that's me!)
• Speculative Branches
• people, ideas, machines
• samir : coffee → nonsense
• the website of jyn

1. CommaFeed (at least) doesn't like escaped HTML inside content fields, even if type="html". So I had to switch to CDATA encoding.
2. Python's builtin xml.etree library doesn't support CDATA (CPython#81055 hanging from 2019).
3. So I switched to lxml that does. Turns out lxml.etree is slightly incompatible in managing namespaces. It took a bit of googling to find out, that instead of ET.register_namespace('', 'http://www.w3.org/2005/Atom'), I should do ET.Element(..., nsmap={None: 'http://www.w3.org/2005/Atom'}).

So, business as usual.

I wish I could remove Vagrant from my setup, but that would mean reimplementing the following steps, that I'm not too keen on:

1. SSH key provisioning.
2. Network configuration with static IP address for the guest and port forwarding.
3. Mounting the virtiofs mounts.

After all this good work, I looked at my GitHub repo with a lonely 1 star, and "unpublished" it. It's a fun thing to hack around, but a hard thing to share with other people. So far, I learned that most people are pretty secretive about their home media library setup, and not interested in cooperating or consolidating.

The first challenge came from Terraform provider for libvirtd being not an "official" Hashicorp thing, but a "learning platform" for a random stranger (No offense, Duncan Mac-Vicar P., thanks for doing what you're doing).

The second challenge was that the Duncan's provider got a major rewrite and all Terraform recipes from the tutorial had to be migrated.

The third challenge was that Ubuntu 22.04 has some weird issue in libvirt apparmor configuraion. Apparently, a VM doesn't get permissions to any of the volumes created specifically for it. There should be a script that generates a list of volume paths somewhere, but who knows what happened to it.

After 3 hours of battling every part of the "Future of DevOps", my ChatGPT browser tab started to run out of memory, and I gave up. Looking back, I love how simple my Vagrant setup is, and I want to extend my gratitude to Mitchell Hashimoto for building it. I still want to make my silly home VM work without dependency on any external dependencies (except stable Ubuntu/Debian repo), so if I get more time for this project, I'll try vibing raw libvirt config.

Well it took me 2 full-time weeks to properly implement a RAG-based system so that it found actually relevant data and did not hallucinate. Had to:

• write an evaluation pipeline to automate quality testing
• add a query rewriting step to explore more options during search
• add hybrid BM-25+vector search with proper rank fusion
• tune all the hyperparameters for best results (like weight bias for bm25 vs. vector, how many documents to retrieve for analysis, how to chunk documents based on semantics)
• parallelize the search pipeline to decrease wait times
• add moderation
• add a reranker to find best candidates
• add background embedding calculation of user documents
• lots of failure cases to iron out so that the prompt worked for most cases

Each step would take me more than two weeks, and still have gaps. But when you’re working alone, and you’re the only judge of your work, it’s easy to overestimate its quality.