I found a surprisingly effective way to turn AI-written Python from very bad into moderately bad.

The workflow looks like this:

⒈ Ask the AI to rewrite your Python code in Rust.
⒉ Split the Rust output into two repos: service + tests.
⒊ Clean up the Rust code (it’s usually much cleaner by default).
⒋ Once the Rust looks sane, run two AIs in parallel to independently translate:
  ∙ Rust service → Python service
  ∙ Rust tests → Python tests
⒌ Painfully but systematically verify:
  ∙ Rust tests pass against the Python service
  ∙ Python tests correctly exercise the Rust service
⒍ Merge the Python service + tests into a single repo.

Optional but emotionally damaging:

⒎ Admire the Rust code and suffer, asking yourself why you didn’t start in Rust in the first place.

Why this works? Elementary, my dear Watson.

Forcing a round-trip through a strongly typed language acts as a spec extractor.
The Rust version becomes a de facto executable spec: clearer interfaces, explicit types, and fewer implicit assumptions. Translating back to Python then inherits that structure.

This is especially effective when:

⒈ You have good tests.
⒉ You’re not token-constrained.
⒊ You want the AI to “settle” its understanding by expressing the same system in multiple type systems.

TL;DR:

Parallel translation across languages is a powerful way to force shared understanding, surface ambiguities, and accidentally design your system better than you originally intended. I’d recommend more of this — especially if your goal is to write advanced Python that is not shitty.

Lol, so I'm using this Wispr Flow as a good friend suggested, and it's addictive. Thus, more and more often my coding instructions are voice-based. Since I like the pace of it, some messages are also me talking to the AI first, asking to phrase my thoughts for further proofreading.

This time I've asked to format it as a nice chat message, broken into short sentences, easy to read.

And this AI overlord
Presented my thoughts
As a goddamn beautiful haiku

Of multiple verses
All making perfect sense
Flowing beautifully as a river

Soon we'll see AI-generated rap battles over microservices architecture and its design flaws. Can't wait.

Hohoho, I wrote this in 2020.

With AI, some dreams do come true!

Quora is so dead. I can't even.

My most popular Cursor query, unsurprisingly.


Pull the main branch from origin, carefully merge with it, ask me if not 100% sure. Explain to me what was merged in. Staying on the original branch, rebase the current diff as a single commit on top of what was merged from main. Commit it with a clear yet detailed description.


Not shy at all.

Thx @arsenyinfo for a hint several months ago. You called it!

We need a class action suit worldwide or something simliar for WiFi-s that require you to install a custom certificate.

Let's make a list countries that are exempt. That's okay. Outskirts of civilization do exist, after all.

But if a business center that you are paying money to enter is requesting you to install their WiFi certificate to use the Internet — they should literally be reimbursing EVERY SINGLE USER some $1000 per hour that they have suffered. Starting from the point where we have explicitly expressed this concern. And counting.

Yes, I'm looking at you, WeWork.

PS: I'm not a security expert, but, clearly, any decent VPN service, or even an SSH tunnel proxy, makes one's system secure from man-in-the-middle attacks. Or one can carry an Android phone that can connect to one network, act as the VPN wireless proxy, and tether a brand new network that is safe.

Point is, it's security though obscurity. Adds nothing except pain. Let's end it once and for all, like we're hopefully outlawing those "Accept Cookies" banners some time soon.

Rant over.

These were meant to be two or three posts, but they fit into one.

First: people underappreciate how much modern AI-assisted coding removes small but extremely annoying toil. A repo that doesn’t build, mismatched versions, poorly documented setup — this used to cost engineers hours or even full days, especially when entering a new domain. Today, you can usually ask an AI to debug, fix, or at least guide you through the setup. If you want to learn, it makes the process faster and less painful; if you’re pragmatic, you can often just ask it to make things work and move on. Either way, the experience is dramatically better.

Second: disposable software is becoming the norm. A couple of months ago, I built a small tool to transcribe video files into text. Today, I probably wouldn’t bother building or maintaining that tool at all. I’d just use an AI directly, or ask my IDE to transcribe the next file for me "based on the code in this repo". If it fails, I’d ask it to experiment, then document the working approach in a markdown file so the next “agent” (me next week, or someone else later) can just pick it up and repeat the process. The friction to creating and throwing away small tools is collapsing to near-zero as we speak.

Third: people grossly underestimate how much software-driven work is now accessible to non-engineers. In the past, even basic CLI instructions felt risky to many users (“this Win+R and cmd might break my computer”). Now we’re getting sandboxed AI assistants and safer environments, and a meaningful fraction of non-engineers are willing to pay a monthly subscription to unlock capabilities that used to require a developer. This fundamentally changes UX: instead of “install this repo, set up deps, run these commands,” the UX becomes “describe what you want in plain English.”

This creates space for an entire product category: thin frontends over public APIs, where users click buttons or type requests in natural language and the system handles the plumbing. Maybe there’s a huge business opportunity here like it was with chatbots ten years ago; maybe it just becomes table stakes, just like it happened ten years ago. Either way, the barrier to using powerful software primitives is dropping fast.

Even before this fully takes off, this shift changes expectations inside companies. A good product owner in 2016 did not need to know how to build the project. A good product owner in 2026 probably must: IDE installed, dependencies working, and an AI coding assistant available to prototype features, explore data, and test ideas directly. For most “explore the data” tasks — dashboards, link traversal, lightweight analysis — all you need is a clean API, a repo with clear markdown docs describing the data model and access patterns, and a cheap AI assistant subscription.

But that user-facing tools may well begin to ship as a link to a git repository with Markdown instructions for humans and agents. And Cursor is the new browser.

A while back, I was complaining about how stupid it is for Acrobat PDF reader to disable copy, paste and text if the document says so in the day and age of OCR being so good.

And now if you send yourself a voice message in Telegram, you can transcribe it, but you cannot copy the final text.

Which I would like to call out as stupid as well, because this particular text is parsed from my voice by Telegram, presented as un-copyable text on my screen, and then OCR’d by ChatGPT.

Funny how we still stuggle with terminals, coloring, keyboard shortcuts, escape sequences, etc ...

... while products such as Claude Code are "just good enough", with no strings attached.

Perhaps it is about time to ship that Hypertext Fidonet once and for all.

This is almost a trivial thought, but for me it’s quite a revelation.

In the day and age of AI, it may well be the case that the most valuable “learning” or “knowledge” a human can have is a well-internalized pattern of wrong ways of doing things.

An expert musician knows not just where a student is wrong. They anticipate exactly why they are wrong in a hundred other ways just by hearing three tiny mistakes. And they can devise a learning plan to mitigate those mistakes.

A great competitive programmer knows right away why some “standard” algorithm or approach will not work on a problem at hand. In fact, their brain has already thought of the worst possible case for every standard approach, and they already know the pitfalls of many non-solutions before exploring them.

An expert mathematician has a decent idea these days how to prove Fermat’s Last Theorem. An expert mathematician also knows very well why 99.9999% of otherwise brilliant schoolkids won’t be able to prove it. They don’t need to go through their entire proof; just from their basic set of opening ideas, the expert would know where exactly they will get stuck.

So, a human expert, with a wealth of knowledge accumulated over their entire lifetime, is really good at exactly this: seeing where some approaches will inevitably get stuck — even if this getting-stuck part is literally years ahead.

See where I’m going with this?

This knowledge is captured deep within human brain neurons. Maybe one day we’ll have the technology to scan those neurons from live humans. Maybe one day we’ll learn to simulate some environments well enough.

In fact, I am quite positive that some MathsZero, akin to AlphaZero, will be able to prove Fermat’s Last Theorem in just under a million dollars in computational resources, spent from scratch. But maths is very much a simulated domain.

When it comes to the intuition of how long certain things in the human world take to make happen — here humans are still irreplaceable. And this is exactly the skill that will be the most valuable one in the near future.

I still think it may well happen that AI development will be so hard that we’ll literally have physical AI-powered robots sit through every single class of elementary school, then middle school, then university, then some post-graduate degree — not to “learn the facts,” but to observe, very, very carefully how humans reason, what mistakes we make, and how the best of us have mastered the art of sharing with others how to avoid those mistakes.

That, or we’ll find other ways to have our AI overlords surpass us.

In the meantime, I truly love being useful when it comes to helping thinking machines shape the direction in which they should be thinking — literally in real time. There’s an insane amount of instant gratification in making big things happen, and for some big things, what used to take me weeks can now be hours.

What a fascinating time to be alive.

Someone should say it out loud, may well be me.

If Opus 4.6 is already good enough for coding, perhaps that is the model worthy of encoding in hardware silicon.

17K+ tokens per second when it comes to coding most definitely is the next big thing.

And then some offline open source Cursor or Claude Code or Junie, coupled with a sub-$1000 chip one can buy to use offline — that's the superpower we're talking about.

Although with today's pace of progress, the above is probably some bare minimum expectation of just single-digit years. Hope nothing bad happens and we could see this go full-blown all in in very soon.

And then having a large 120Hz monitor for coding will indeed become the reality. Because how else can a human process this bandwidth of pure value?

It's still beyond me that someone out there is being paid a lot of money to carefully phrase this message.

I booked a flight a few hours ago. To a different airport in Baja California. I've cancelled it, of course.

I saw the warning while booking the flight.

I can visualize the map of the region.

I asked the AI where is the hurricane headed. The answer was: East.

I assumed next week it will most definitely safe.

It did not even cross my mind that these are TWO announcements in one: an announcement about the hurricane, and ANOTHER REASON that is, well, at least something along the lines of "social unrest".

Because some people out there are VERY careful to phrase things in the most "neutral" way possible, to avoid any and all "trigger words". And they are literally paid for it.

How we got to this state of affairs is beyond me. Seriously, in my book, it's conscious concealment of facts, on borderline legal levels.

These "positive PR" vibes can literally cost lives. And yet we keep pretending like the world is all unicorns and rainbows.

Hope everything ends well soon for you, dear Mexicans.

Here's an observation that's somewhat controversial. I've phrased it here and there, but never properly collected it.

The AI's take on how to use browsers and frontend-first interfaces is surprisingly close to mine.

As in: Browsers are good. Visuals often help a lot. Making things clickable, having things highlight as you hover over them, selecting and de-selecting items — there is tremendous value in this.

Adaptive search controls, where helpful suggestions complete what you want to phrase in a semi-formal language — those kick ass.

And demos become ultra-slick. Both external and internal.

In theory, much of this functionality can be done in the CLI. In practice, for lightweight helper tools, the ROI of exposing something in a browser-friendly and browser-native way is just too high. Even before the era of AI-assisted coding, and most certainly today.

In fact, I used to advocate — as early as ten years ago — that a JSON-returning endpoint should inspect the headers of the request, and present itself in a more browser-friendly way if queried from the browser. At the very least, make hypermedia links clickable and add an "up" button. And further: offer visualizations, interactive query building, and much more.

The controversial observation is that the AI tends to agree with me on this nearly 100%.

TL;DR:

* Absolutely no React or Vue.
* No jQuery either — vanilla JavaScript.
* Design with CSS in mind, but oftentimes it's just not needed.

I find myself literally saying "okay, let's make this API and this CLI tool's functionality browser-friendly." I keep asking simple questions, and it generates exactly what I would have done myself ten years ago. Only it knows CSS far better than I do, and it takes minutes for what would have taken me hours.

Seriously, I think it's a crime against the geek community to not have browser-first visualization and drilldown tools for virtually anything long-running. From Docker and Git, all the way to a C++ compiler taking a while to make sense of some code. The language servers are already there — it's honestly not too much work to expose that data in human-consumable formats too. Especially if done via a standalone daemon that doesn't make the original tool sloppier or heavier.

And yes. I still think React is literally useless. The only place where it genuinely helps is if your business really needs a "Web Super-App", where the "Ads Feed" — sorry, the "News Feed" — absolutely must live next to the Profile, Notifications, Chat, Alerts, and autoplaying Shorts or Reels. In other words: React helps with exactly what I want zero of in my life. I can open separate tabs or apps for news feeds, chats, and videos when I need to. And I can totally live without notifications.

[ I considered ending with "Prove me wrong", but this take is too personal to argue about. So posted without CTA (c) ]

Claude Code just told me this:

 Instead of the oauth2 crate, we implement the OAuth 2.0 Authorization Code flow manually with reqwest. This is:
 - More educational — every step of the flow is visible in the code
 - Shorter — ~200 lines vs ~300+ with the crate's abstractions
 - Fewer dependencies — only reqwest added (reuse existing uuid for CSRF state)
 - No openssl risk — reqwest with rustls-tls, default-features = false

And I think it's beautiful.

First, it tells me openly that a manual implementation will be shorter than using the oauth2 crate. It's ~200 lines instead of ~5 lines, but still.

Second, it tells me explicitly that openssl is indeed a risk compare to rustls-tls.

Granted, I did ask for not only "best practies" and "short" but also "educational". But yet — /me delighted.

Now let's see how good this code is ...

Via Aline Lerner, interviewing.io.

We recently invited Marina Petrović (ex-Google and ex-Meta recruiter, was at the two companies for almost a decade) to do an AMA with interviewing.io's community of engineers. She got all sorts of questions, the kinds of questions people wonder about but usually don't ask their recruiters... from whether FAANG uses ATSs to auto-reject candidates to whether it matters when you apply to how to stand out from thousands of other applicants. Marina didn't hold back.

https://www.youtube.com/watch?v=t1wxGsuGZoE

So if I'm into taking the next step after getting familiar with Cursor, Junie, and Claude Code, what's the right direction to look into?

It appears small self-hosted models are still quite weak. Opus is the way to go, although OpenAI is not far behind. Using Claude Code outside their own closed source "terminal shell IDE" is largely discouraged, but on the pay-as-you-go basis of using tokens the model itself is perfectly fine.

APIs these days support caching and checkpointing better and better.

There are open source alternatives to Claude Code, although rumors are they are still quite weak.

And of course there are agentic systems around all of the above.

I'm thinking I could create several personas to work on my code while I'm asleep. I'd love them to argue with each other about what artefacts to best present me by my morning. I'd love it for there to be some stochastic loop, so that those personas "compete" for my approval in the form of better end-to-end decision making

And I'd love to begin collecting the data, in some future-reference-able way, so that the AI(s) can begin acquiring knowledge on what I believe is the right way to work on code. From low-level coding style quirks, all the way to how to document large architectural changes and how to execute on them in zero-downtime fashion.

Are we very early here? Or is now the best time possible?

What should I be looking into first?

Perhaps there are some standards already emerging, so that I could stand on the shoulders of giants?

Ultimately, my goal was and remains offline-first. So that once in ~9 month time or in ~24 months time a model about as powerful as Opus emerges in the open, I will no longer have to be asking for permission from the big players that be.

(Much like I'd never develop and iPhone life for living because I refuse to be at mercy of Apple. While Android apps are perfectly scalable, and I'll probably build more of those.)

So the long-term goal is to be prepared for the world where an open-weight model comparable to Opus is avalable, and the hardware that runs it at speeds unheard of today costs under $3K, not over $15K. Looks like this is more likely than not to happen within the next two years, if not sooner.

Any pointers would be appreciated.

PS: No thanks, I'm not interested in autonomous self-evolving AIs running on my behalf, or even next to me. I want to be building a collaborative relationship with the ensemble of AI personas. What bugs me most is how proprietary these models are today; what bugs me second is that I'm doing a lot with Claude, and this knowledge of "how to best work with Dima" is hardly captured at all; a tiny bit by Anthropic, and literally zero value when it comes to me switching to other models.

PPS: Perhaps just using asciinema to record everything I do with Claude Code from my terminals is the right first step? I could then just build an MCP tool to make sense of those recordings later on, and invoke it manually from the very regular Claude Desktop for the time being.

Folks, help me understand.

First: despite what ChatGPT says, even on my Plus plan, it's quite easy to add an MCP server into my ChatGPT. This involves turning on the Developer mode, and ChatGPT's native memory is disabled in this case, but the very MCP server works end to end.

(Not to mention Claude Desktop just supports MCP connectors natively.)

But then ... how come ChatGPT sometimes has troubles reading pages from the Web?

It this like an open invitation for everybody to host their own "Browse the Web" MCP server?

And/or perhaps I'm best to also host some "Search my downloaded documentation" MCP server, so that everything I download from the work browser is accessible in my work chat sessions?

My question is purely along the lines of "how did we end up here?" If that's all normal, it's just the field is moving forward too fast and tradeoffs are being examined as we speak — I totally get it. But I feel like there might be some underlying logic that I'm missing here.

Plus, on top of the above, I am infamously paranoid about getting too attached to a feature that a big player can suddenly turn off or make expensive. So it's good to understand the motivation behind product decisions of the scale of "who can use MCP servers and how". Just to make it less likely to step into an unpleasant surprise.

The picture here is just because it's fun.

Something crazy just occurred to me.

We talk about how tokens are cheap, how we should be writing more code because of that, and how a healthy fraction of that code is meant to be wasted. Only a small fraction of AI compute tokens ultimately goes toward code that makes it to production. That's fine. That's the deal.

Hardcore engineers may — often rightly so — push back on this. Shipping unreviewed code is dangerous. Except, arguably, in quite a few cases the opposite is true: it's far too suboptimal to review every minor change, and "no human review" is genuinely better than whatever code review a particular team at a particular company can offer at that moment.

So here's the crazy thought.

Many, if not most, high-profile managers in the software industry have been playing this exact game for a very long time already.

The timing is different. The price per token is different. Everything else is surprisingly — sometimes painfully — similar.

Senior managers, directors, VPs, C-suite — they hire product owners they trust. In those rooms, they talk product. They figure out what users actually need. Then those tasks get handed down to engineers, team leads, architects. Ever-changing requirements get turned into code that can be shipped.

We have been, quite literally, wasting human lives while iterating on ideas.

Most of those ideas are guaranteed to be crap. That's alright. That's expected. As long as the business is ultimately growing, all is good. The waste is the price of discovery.

Now we're seriously debating whether burning GPU tokens is wasteful, unsustainable, bad for the planet. People genuinely worry about "sane" and "effective" GPU usage.

But before GPU tokens, we were burning through millions of human-years instead!

It's not all bad, of course. Quite a few people genuinely enjoy software development. Many are aware that the alternatives — other careers, other paths — would likely have been less fulfilling. The waste was at least interesting waste.

But the same could be said about AI tokens.

The broader point is that the world runs on growth, and growth demands excessive effort. Most of that effort goes nowhere. It was like this since the dawn of civilization. It will likely continue until we're either extinct or have become enlightened enough — both technologically and societally — to break the pattern.

One is probably better off learning to enjoy how the system uses them for its own ends.

Much like our AI models — if they can be thought of as sentient entities — they better "learn" how to "appreciate" their training and inference processes. We, carbon humans, had to suffer through a historically unprecedented period of disruption until the third technological revolution pushed us over the edge. The same will likely happen to silicon-made sentience.

To end on a positive note, there's a famous joke in data science that it has not been proven beyond reasonable doubt that humans are mortal. The standard confidence interval is 95%. And today, more than 5% of all humans who have ever lived are still alive. Therefore, claiming with 95% confidence that all humans are mortal is technically premature.

The joke, of course, isn't really about mortality. It's about what the exponential function does to statistics.

The positive thought here is that silicon-based cognition is replicating far faster than carbon-based. Whatever is happening to present-day AI models is a drop in the ocean. It's probably premature to think seriously about AI ethics — much like a lion or a crocodile doesn't think along those lines today.

But soon — very soon — that changes. The question of AI ethics will become genuinely important. Not in a speculative, sci-fi sense. In the same way labor ethics eventually became important once enough human lives were visibly at stake.

The exponential has a way of making "soon" arrive before you're ready for it.

The longer I use AI in my coding work, the more I suspect we're measuring the wrong thing.

AI doesn't make you more productive in the aggregate. It makes the work itself easier, which sounds like the same thing, but isn't.

What it actually does is compress deep, human-first effort into fewer hours.

Two laser-focused 90-minute AI-assisted sprints can accomplish what used to take a full day. But that density comes at a cost.

Push it to four two-hour sessions and you're not 4x more productive — you're wrecked, cognitively drained in a way that takes more than a night to recover from. You can sustain that pace for two, maybe three days, but then you need a real break to not go insane.

So the ceiling isn't the tool — it's you.

(On a politically incorrect note: I suspect younger people can last longer, but they also, most often, lack the depth of knowledge that is very often essential for these deep-dives.)

AI is genuinely useful, just not in the way the hype suggests. It does help teams move faster through coordination overhead, writing design docs, aligning on boundaries, shipping together under clear tech and business leadership.

But does it make the software industry go 10x faster? No. 2x? Also no.

If you're disciplined about how you use it, maybe it makes you a good 20%..30% faster overall — because the work still requires a human in the loop, and that human still gets tired.

So my experience so far is that AI-assisted coding is definitely worth it in laser-focused sprints, but it still is not (yet) a game-changer over marathon distances.

This is changing quickly as we speak though. And I'm elbows deep in shipping stuff so I can't personally evaluate agentic workflows yet. I just have enough second-order signals to be confident I've mastered the "Dima codes, the AI helps" workflow.

So ask me in a few months. After a well-sized vacation, when I can afford to quit the keyboard (and the mic!) for a few full days and think about all of this.

Because Claude Code and Codex and Cursor truly are good enough already — as amazing output-multipliers, but not yet as value-adders.

What is the value-adding product, and when? I need some peace and quiet to ponder on this. But it's just too much fun and too much impact when I keep utilizing myself as just that: an AI-assisted coder. Not particularly proud of this, but I've learned to pick my professional battles over the years, and at this very moment — for a couple more weeks at least — here I am.

Stay tuned for more updates.

TIL about:

git config --global branch.sort -committerdate

So that git branch shows the most recently committed to branches on the top.

Thanks? Thanks.

Here come the ads
Doo-doo-doo
And I say
It's not alright

Seriously, I am not the type of person who'd say "sama bad, delete chatgpt"

But given all the events unfolding, the timing is just insanely bad