Initially, when I received an email notification, I was happy for him. Read the post in one go. It was clearly a huge effort, with failed attempts, restarts, and a lot of persistence. I reconized the same Dan we all met and loved in 2015. (React Europe 2015, what a lovely memory). But alongside that happiness, I felt something sad: I understood why this should amaze me, yet I didn’t feel as amazed as I expected.

I notice this more frequent when browsing Hacker News and stumbling upon these Great Hacks.

Let's get back a few years go. (Preferrably before Covid and War). These used to be some of my favorite things to read:

• JSLinux: booting Linux inside a browser - 2012
• 8088 MPH: We Break All Your Emulators
• Reverse engineering Apple’s GPU to make it work on Linux.
• Figuring out why GTA Online takes so long to load and cutting that time by 70%.

I would open those links with Vorfreude. They were like some sort of unfullfilled dream to me. I'm not capable of doing that myself (except maybe short period of time in 2015-2016), but man can dream. You read this articles and secretly believe: one day will accomplish such things, because there are so much to explore! Or maybe it's like buying a gym membership. Or just a hobby that you wish you had, but something else in life gets in a way, but at least you can peek.

Now there’s a growing collection of such stories with AI doing a substantial part of the work:

• Rewriting Bun from Zig to Rust with Claude.
• Rebuilding Playwright’s browser-control layer in Rust with coding agents.
• Using agents to decompile Modern Warfare 2, and the whole game deplatformization thing. Like https://playgta5.top/

These are the wonderfully specific personal projects. But now it's "my coding agent fixed my toaster" genre:

• A tool for controlling a coffee machine through its reverse-engineered Bluetooth protocol.
• Investigating the firmware of a monitor, webcam, microphone, and other gadgets within arm’s reach.
• Reverse engineering the drivers of a 2008 Microsoft Surface table so a coffee table can run modern Windows.

Imagine reading them 3-5 years ago? What a story: old hardware, a reason to care, going deep to make it useful, and care enough to share.

Now I find myself asking how much the agent did before I’ve even let myself appreciate the result. And I know by heart that agents right now (Opus 5.5 time of the year) are great in many things, they can reverse engineer a protocol, decompile and compile a game for another problem, but there's still lack of intention and you still

Another recent example was an interactive visualization of Calvino’s Invisible Cities, generated from one prompt. The author was mesmerized. I read it and felt very little. Another thing that would once have seemed extraordinary had become something I could scroll past.

I think partially my "wow" about such things used to come from imaginging the creator as a person. At past we had a privilege to "fix" things in our spare time, and this was making our whole profession more interesting. Learning something new while overcoming your struggle, Dan's post certainly has plenty of it. Achievement is still there, AI is just another tool, right? If some post got popular enough, enough people recognized it to care.

And in case of Dan and what I envy about Dan is the time he has devoted to math and learning it from scratch, long before AI.

I hope I will adapt, but right now, my excitement is somewhere else.

For the past few months, starting last December, anxiety has been my dominant feeling. Always in the background. Every few days another voice joins this movement and share: everything has changed because of AI, everything will change even more. Tools are obsolete, the way how organizations work is outdated, team dynamics, markets, workforce. As opposed to previous dramatic changes in our profession, it's no longer clear if it's possible to adapt for most of us.

Many people are trying to adapt, including me. In 2025 it was about final settlement: "it will be just another level of abstraction, problem solving is still on us". The question is do I understand enough. I used to learn by doing, but here, it's no longer work that way. Or at least not working enough to protect me from anxiety. Did I find the right word?

It's partially similar to being nervous before a competitive sport event, which is good. But partially a fear of loosing my job. I can't call these times comfortable, but it might get much worse.

I guess I'm just afraid to loose "experience" and start from scratch. But there's signs that it will matter.

AI Assistance all day long

A year ago, I was regularly consulting with LLMs about algorithms and APIs, used auto-complete heavily. Now it's all agentic long tasks. Business as usual, AI agents are no longer experimental thing for side projects. It's an essential part of most of workflows now. People may turned down a little on writing assistance, on generating slop prototypes, just because our attention is limited and trust is easy to lost. But it's hard to imagine not using AI Agent at all. I can even say, I envy those who can afford that.

Frontier models are still from the same companies

Let's remember this: In the beginning of 2026 it's time of Anthropic Claude 3.6 and GPT-5.3/5.4. They are gold standard and game changers. Gemini 3.1 family is good for everyday practical tasks, but rarely used in my bubble for agentic coding. Used though in many other places. Claude Code is a default agent. Codex is getting bigger, Pi.dev, which is behind OpenClaw getting traction. Antropic grows like crazy.

Super fast inference

A glimpse of faster inference is here in various places. It's easy to be impressed by it. But you don't want to give up your best possible model. What if it miss something and introduce more bugs? For small safe tasks it's easier to edit a color of a button with them, than opening the file. Codex allows to run faster their models but it's twice as expensive. Still, worth it for many and I wonder what will change.

Black magic and art of satisfying agent needs

Claude Code shipping tens of features each month. You can now get certification for mastering it, I'm not kidding. People created a cheat sheet that takes full A4 of commands. People share their lifehacks. If you have a bad experience it's a skill issue, and jokes being made that it's literal now.

What is hard it is to scale these practices. Have you ever seen all ceremony and instructions, number of scripts and custom CLIs in a big project? Trying to describe it here would take more text then this post, but it's almost is how to answer to questions "what to do, so people using two leading agents can contribute in this codebase doing as low customization as possible".

Cursor decline

Cursor is not that fancy thing anymore, despite still being very impressive technically and product-wise. Hype was so strong a year ago. As IDE, there are better ones, and then again for agents I still use advanced terminal setup (occasionally trying new things, like Air).

In pursiut of autonomous agents

Anthropic, OpenAI and Cursor are experimenting with it. Days and weeks efforts where an agent does a very big project with as little of human intevention as possible. Competition can be replicated, so data and distribution is important as ever. As we develop more sophisticated frameworks for such agents, probably a completely new browser engine or operating system will be possible. And that brings anxiety again

Open Source is changing

Sending PR now is not limited by your familiarity of the project. It's still an effort, but not quite as before. And it doesn't make you invested into the project as much. I don't have high hopes sending PRs to random projects from the previous era at all. Maintainers probaly overwhelmed, why care.

Hangover after fake productivity

This is what I experienced after initial boost to "finish some apps I always wanted to finish". I got very little satification looking at the final result. I don't feel pride in them that much. It's some kind of depreceation of side projects, and it's somethin I need to get used to.

Indie Developer Dream: died or not?

The dillema of the last few months: on one hand it's never be that easy to have something published now, on the other hand nobody would need this. One can build a small service instead of paying you subscription. And this is devastation for many people, especially not builders. When you can only dream, previously you have an excuse: no time. Now you have another excuse: why bother.

So we'll see what exactly will be differentiators. Maybe it's time for influencers and socially active people, if that all that matters. People with taste and love of craft, will they have enough users?

Random notes from a single random working day

• Developing native macOS apps is incredibly frustrating. Can't get anything done that is nice to use without very detailed spec with a lot of follow-ups.
• Without oversight things can go in a wrong direction quite fast. You should either code review everything thoroughly or invent a sophisticated verification pipelines that takes as much time and constantly changing
• As everyone says: someone should decide what to build and what not to build, to keep a balance.
• You have bigger pace when doing something on your own. But only if you know youself what to build. Syncronization is pain. But only on code level. If you want to build the real product, the product and QA is even more important.

We still have work to do

It's important to decide what not to build

Until "please make this product possible and then keep adding features" prompt are viable, there are tons of micro descisions every day on every level.

Most of the projects appeared last months have not yet taken significant place in my life and my habits, except SOTA agents for software development. I still like polished user interface, familirity and predictability. I used to these small dopamine hits of hot paths when you interact with it almost instinctively, and reliazed I prefer it over talking or writing with some AI bot.

Constructor UI

So I need a special UI. First version looked like this:

It can be polished, thought-out, optimized.

But I quickly understood I need even more complicated engine. Do not want to implement it, will not want to use it.

Code editor

Why not instead adding a single huge text area, give some example for pseudo-language that can express condition, add validation and autocomplete?

A code editor. Something I have experience with. This I can deal with. Could be fun. Text-editor with parser, validator and auto-complete. It could have been an assignment task a 5 years back.

Natural Language input

In 2026 writing formal expressions is considered old-fashion and waste of keystrokes. People probably will never write regexes, Excel formulas and home brew DSL rules again. Almost certainly, tot at the site like this.

It's simple. Let's add a request that some LLM on the backend will transform into a rule? But I don't want to deal with LLM keys, and local ones that I'm able to run on my tiny server are bad. And what will happen with that LLM in two years?

Local LLM compiled in browser? That's much better. At the time of writing this, it's not yet there, but someday.

Just Ask Your AI

Okay, don't keen on embedding LLMs, but can't ignore it. Instead: just a copyable json schema and an input - past to UI and ask for writing. Last model, no complicated copyright.

This is what I ended up with. Requires copy pasting, and still ocassionaly break. Not a problem, and it's future proof in a way that agents can use it too.

Browser or OS-level standard in the future

If LLM progress were frozen, we would have eventually something like this:

<input type="ai" validation="schema.json">

A special input, that browser recognizes and use AI that is tailored to you, making it very simple to create a rule.

But really, no need to bother. All browsers soon if not already will be able to do it anyway, without any hints from developers and just help with any kind of input everywhere.

And even then, only useful we have traditional interfaces. If OpenClaw shown us anything, we may rethink the whole idea and people will just interact with the single entity. It'll solve prloblems by implementing Weirdstats under the hood, but you'll never even know about it.

You are not “checking your old projects” the way you used to check old photos. It is more like opening a drawer of odd objects you once built by hand: a tiny wooden box, a key you no longer own the lock for, a perfectly shaped stone you picked up because it felt important at the time.

Your projects are like that now. Some still open with a satisfying click. Some fight you. Some are just labels pointing at places that do not exist anymore. You start with the easy ones.

A directory called potomushto.com is mostly text files. Markdown. A few HTML pages you once generated and never touched again. You open one and it just appears. Instantly. Instead of “restore environment,” “dependency mismatch,” “this plugin can’t be verified.”

It feels the way it feels to find an old printed photo. HTML still differs. The rendering is different than it was in 2025. Fonts drift, spacing changes a little, but the words are the words. Links still point somewhere or they do not. Either way the text stays. Plain content does not need an ecosystem to survive.

Then you open the harder drawer.

A little macOS app you were proud of. Sleek, minimal, “native.” It used to feel permanent because it lived inside the operating system, not on some website. But the OS has evolved the way cities evolve. It did not “break” your old apartment, it just stopped building roads to that neighborhood. Code signing rules changed. CPU architecure changed. Sandboxing got stricter. A security model shifted after some global incident you barely remember now. The app still exists as a binary like a fossil exists as a shape. Recognizable, but not quite alive.

You can revive it. Of course you can! Everything can be revived now with AI. But reviving is never free. It costs attention, and attention in 2050 is expensive in the way your free evenings used to be expensive.

Finally, the games.

Some run in a compatibility bubble and feel right in a weird way, like hearing an old song on new speakers. Others load but the magic is not there: the physics are off, the controller mapping is wrong, the timing feels like someone retold a joke with different pauses.

You remember, suddenly, that games are performance more than just software. Timing. Latency. Input. A particular kind of friction. And time changes friction.

You scroll through a neat list of services that were once cheap because each one felt like nothing: a small site here, a webhook there, a little worker, a domain redirect, an analytics script, some uptime monitor, a storage bucket with “just a couple files.” Each one was fine. Together they are a constellation with a monthly price tag that looks more like a utility bill than a hobby.

Every project that stays alive demands a little ongoing payment.

So you consider the other route: the home machine.

Not romantic, just practical. A quiet box that you replace every so often, and the maintenance cost spreads out across years like buying a good tool instead of renting one forever. It is not free, but it is predictable. It feels like owning a shelf in your own basement rather than renting drawers in a hundred different buildings.

One machine in the corner of your future home, humming gently. Some projects hosted there, most archived there, all of them yours in a literal sense.

And then you hit the real graveyard: the integration projects. The ones that talked to other services like WeirdStats. Most of them are gone, not because your code rotted, but because the other side disappeared. APIs closed. Companies pivoted. Services merged, rebranded, died, got regulated, got acquired, got sunset. The internet became different.

Those projects feel like old screenshots of conversations with people you no longer know. Not sad exactly, just un-runnable.

2075

You are older now, older enough that you stop trusting anything that depends on a company staying interested.

By 2075 it is harder to imagine that a hosted blog still sits quietly on some free tier, unchanged, faithfully served from the same place. Even if the page still exists, it is a miracle you do not want to budget for.

But the repository feels different. It may survive, not nesseceraly as a git, but as some source control. Filesystems still exist in 2075, one way or another.

A hosted site can disappear like a storefront. A repository is more like a box of papers. It gets copied. It gets mirrored. It gets bundled into backups and exports and “takeout” archives. It ends up inside someone else’s cold storage the way family photo albums end up in attics. Even if the original platform is gone, a bundle often survives somewhere because it was small, easy to copy, and once copied, hard to fully erase.

Maybe the GitHub Pages version is gone by then. Probably a long ago. Maybe your domain points to nothing. But the git history, zipped into some archive you forgot you made, still has all the posts. Markdown is still there. HTML is still there.

By then, much earlier even, AI have eaten so much of the internet that “the past” is accessible in a new way. Not as exact files, but as a kind of memory. Slightly compressed. Slightly vague. Full of smooth edges where details used to be sharp. But ready to be regenerated almost instantly into something you can touch.

You ask yourself: if a model can reconstruct the feeling of your old app from something by accessing your memory directly, does it matter whether the binary still runs?

If a model can recompile your game into whatever the current modern computer is, does it matter that the original engine died? If a model can take your old blogposts and lay them out in any style you want, does it matter which generator you used? It is tempting to say no. It is tempting to accept the new kind of preservation, the kind that keeps the gist but not the grain. But then you look again at the printed photos.

A printed photo does not try to help. It does not summarize your childhood. It does not improve the framing. It does not infer what you meant. It just sits there with all its awkwardness intact. It preserves the exact light that happened to fall on the scene that day.

A model gives you a version that is usable. A file gives you a version that is true. Maybe nobody needs it except you. Maybe this museum is private.

Maybe in 2075 it is not about future generations at all. It is about having a few artifacts that do not negotiate with you. A few things that remain stubbornly themselves, even when everything else in the world prefers to be regenerated.

This lets you meet your earlier self.

1. stand-alone personal blog
2. do not care just give me engagement
3. stand-alone personal blog

I’ve had several phases in my life when I tried blogging. The current version of potomushto.com is probably the 4th reincarnation of it. At one point, I created a separate Tumblr just for a seven-day trip and I still love that blog. Going back and rereading it once every five years gives me chills.

Another time, I used a throwaway name on Blogspot or WordPress. Something that cost me nothing and required zero hassle. And it’s still there. Still working. Still openable after all these years. Still close to zero readers and visitors.

Twenty years ago, I loved writing so much that I did it almost every day. After a while, I decided that what I was producing didn’t look serious enough, and that I needed a more professional blog. Then I decided I wanted to write essays. Exquisitely bad pieces. I guess I just loved the act of typing on a keyboard. They were so bad that I came to an inevitable conclusion: not just to stop, but to shut the whole site down.

And then, ten years later, at 2015 it became obvious that social media gave you a much bigger boost. It felt great riding that wave: you could easily discover new, interesting people, and they could find you just as easily.

That went on until I started my new full-time job. I had less and less time to write. Ny energy for “extra” writing was slowly running out. I felt no need, and that was a mistake. No side experiments, no social networks.

Compare these two images:

Now, in 2025, when most of my social network circles have dissolved and scattered across multiple platforms, I feel a bit of regret. I feel like I missed out having personal web-site empty all these years. So now I care much more about having a permanent place — something that can still be referenced in the future.

And technologically, all pieces have always been right here. It turns out RSS is great. Owning your content feels right.

The only thing left to solve is me having insufficient writing skill.

What makes blog good

A blog can be good as a private intellectual artifact whatever that means, and better if it resonates outward meaning: RSS subscribers would make me happy. I want to be proud of it.

For improving the writing skill I need consistency. Consistency in writing attempts but keeping the same core, angle, principles. Consistency not in producing reliable mediocrity but in thinking. Keeping posts in drafts for years, never publishing them because they are not good enough is something inevitable.

This is the most naive but hard truth type of phrase that everyone hears constantly during any kind of German classes. Another version is that to perfect a new thing, you'll need to do it 10 thousand times. Of course the devil is in the detail. You need Tight Feedback Loop, Deliberate Difficulty (Not Comfort) and Reflection + Correction. A similar take is here on Veritasium.

My plans to try out different things for feedback loop:

• Reading the draft out loud and fixing awkward English
• Rewriting the same paragraph 3 times in one sitting
• A trusted peer review
• A review from AI but not a rewrite

So it's muscle memory for writing. And I love to train muscle memory, which usually gained by drills.

You need to have a taste for being self critical and edit in place, so reading and noticing what works for others, a long-term work I'm doing frankly not sufficient for English. Not reading nearly as much books.

Deliberate difficulty is just inishing any post that will pass my internal quality gate. That is deliberate difficulty. Partially because writing in a non-native language increases cognitive load, partially because of an internal quality gate. And of course I want to touch difficult topics. Exposing this definetly makes it risk and increases demand.

Reflection is most confusing to me. Writing and reflecting on your past work instead of touching and silently editing it, is not something I'm used to. I should embrace and track posts I'm embarassed about.

New Year Resolution

Having said that, my goal for next year is simple: to write regularly and finish posts. Not to optimize for reach, and not to wait for confidence, but to treat writing as practice. Let's say 12 posts a year.

Most of the time there will be no feedback. No comments, no replies, no visible signal that anything landed. That silence is part of the process.

So when it’s time to share, I will do it anyway. Publish, step out of my comfort zone, and repeat, trusting that consistency in thinking, difficulty, and reflection compounds, even when no one is watching. How naive.