A one-trillion-dollar valuation is no joke. Anthropic and OpenAI are both at this level, and they are both preparing for IPOs.

Of course, expectations are high. Many people will bet insane amounts of money on these two companies. But are they even in a good position to win the AI race?

I try to find the answers in the video above. Here’s a brief description of what’s in it.

Both Anthropic and OpenAI started as chatbots, then quickly pivoted to coding. LLMs are extremely performant pattern matching machines, so coding, which is a general language domain with very small vocabulary and easy grammar, was a low-hanging fruit. Both companies are relying extensively on this and the majority of their users range from vibe coders to enterprises looking to cut their costs.

But there’s life outside the tech / digital vertical. Too many people get stuck in their own opinions and end up in echo chambers. While outside these echo chambers, established companies like Google, Meta or SpaceXAI are already finding use cases outside coding. Google is embedding AI snippets in search results. Meta uses the Ask Meta megabot to increase engaging, and SpaceXAI is tweaking its algorithm using Grok, while at the same time uses Grok as an affordable conversation partner. All these 3 companies are quiet, but there’s a fundamental difference between them and the 2 main contenders.

The difference is in the users, and how much it costs the neolabs to acquire them. Go ahead and see the rest of my argument in the video.

If you like it, just like and share, and if you have any other topics you want to me touch on, just live a comment on YouTube, I read them all.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about Grok Bot, the agent released by SpaceXAI, having its iPhone moment, and, if you’re premium, a sneak preview of my latest web app: unslopifymy.ai. If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

Grok Bot it’s the “iPhone Moment” for AI Agents

There’s nothing revolutionary about any of the parts it’s made, but the way everything is put together makes Grok Bot stand out as a product of its own class. I talk about this in a short YouTube video this week.

If you’re mainly using CLI’s for your AI work, this week I also wrote an article how about I use Grok Build, the Grok model harness, to select free models - that’s right, using DeepSeek but with Grok Bot harness goodies.

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading for an exciting update about my latest app, UnslopifyMy.ai.

Read more

SpaceXAI just made Grok Bot templates shareable. I had early access to the feature, meaning I published / shared my first Grok Bot before the feature went live. In the video above I’ll walk you trough the process of creating and sharing your own. But before that, be aware the SpaceXAI confirmed a marketplace for bots will be up and running as early as next week. That means you could actually sell bots and make a profit out of it in a few days.

I already called Grok Bot the iPhone moment for agents. Not because any of its parts were brand new. In the case of the iPhone: the music player, the computer and the phone — they all existed before. But the way they were put together created a product with its own class. The same thing was true for Grok Bot, but now we have even more evidence. What App Store was to iPhone, the upcoming bot marketplace could be for Grok Bot.

Shareable templates it’s what makes this possible.

Anything you build inside the app — give it a job, let it grow a routine — can now be packaged, shared and it gets its own link. Someone else opens that link and drops the bot into their own Grok Bot. It’s that simple.

Go ahead and watch the vide, and if you have any cool bot use cases, let me know in the comments, I’ll check them out.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about how to choose an open weights model, how to transfer files with light only, and, if you’re pre ium, a very exciting update to my app directory project, AppCharts. If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

How To Choose An Open Weights Model

I think open source, local AI is the future. But the pace of development in this sector is insane, we have a new open weights release almost every day. So I decided to make a short, under half an hour video about how to actually choose your free AI inference. I ended up splitting everything in 4 smaller videos and the total watch time is under 30 minutes. Here’s what’s in it:

• Episode 1 - 4 metrics to evaluate a model on

• Episode 2 - run your model via a third party provider, when and what do you need to know

• Episode 3 - run your model locally (my favorite), 2 tech stacks available, Apple Silicon and NVIDIA

• Episode 4 - what is harness and why do you even need one

There’s also a neat playlist on YouTube for your convenience. If you find the videos useful, consider liking and subscribing, and suggesting to me what kind of content do you want me to add in the future.

Transfer Files with Light Only: ShareByLight

We already talked about this tiny utility I wrote in our last newsletter, when I announced the Mac version was up in the Mac AppStore. Meanwhile, the iOS version was approved too, so the circle is complete.

You can use this pair anytime you want to get one file from your phone to your computer, but don’t have access to WiFi, Bluetooth or any other kind of connectivity, including physical cables. All you need is a direct line of sights between the two devices (and it works in all 4 combos, iPhone to iPhone, Mac to iPhone, Mac to Mac, and iPhone to Mac).

Get it for free in the AppStore.

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading for an exciting update about my app directory AppCharts.

Read more

If you ever tried to use AI, with models like Claude or Grok, you ended up with anything betwee $50–100 a month if you use them seriously. Open weights (or open source) models are now good enough that you can get same level of quality for much less — or, if you run them locally, even free.

Choosing an open weights model is not that easy, because there’s no company behind telling you what it does. You have to figure it out by yourself. So I made a short 4-episode mini-series that outlines the entire process.

Episode 1 is out.

• The actual difference between open source and open weights

• The four metrics that you need to know when picking an open weights model

• Parameter count and how different size ranges are used

• Dense model vs Mixture of Experts (MoE)

• Quantization levels and the quality/size trade-off, 2-4-6-8 bits

• Text-only vs multimodal (video, image, text)

Next episodes will cover third-party providers, running models locally, and harnesses.

Long answer, it’s a bit more complicated. To get the full picture, just watch the video, it’s roughly 5 minutes.

Key points:

• Watermarking is implemented because EU regulations

• It survives copy and paste, because it’s based on probabilistic token choices, not on invisible characters

• Heavily editing the text basically removes this type of watermarking

• If you want to avoid it entirely, using local, open source models is the safest way

In this video:

• How Anthropic’s implementation (using Google’s SynthID) actually works

• Why longer text is much easier to detect than short replies

• The surprising side-effect: even human text you paste into Claude for proofreading can get watermarked

• Two practical ways to remove or avoid the watermark

Timestamps:

• 0:00 — Intro: What AI watermarking is

• 0:55 — How the watermark is applied (token-level changes)

• 2:28 — Why short chats are safer than long-form content

• 2:48 — The proofreading trap

• 3:13 — Method 1: Heavy human editing (recommended)

• 3:48 — Method 2: Open-source models (with an important caveat)

• 5:14 — EU regulation context + final thoughts

Key takeaway: Watermarking is not baked into the model weights — it is applied after generation.

Local open-source models on your own hardware currently give the strongest protection.

If you found this kind of content useful, like and subscribe to my channel, it really makes me wanna publish more.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about Anthropic watermarking all of its content, and why China already won the AI race. If you’re premium, happy to report that 2 new apps are up in the App Store: Coloring Kiddo and ShareByight If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

AI Can Follow You Everywhere Now. Unless You Outsmart It

Recently, Anthropic announced that all text produced by its models will be watermarked, making your entire interaction not only recognizable as coming from Claude, but also traceable to your identity. According to their press release, the watermarking will survive any copy and paste, and it will be “invisible” to the user.

Two things worth mentioning here.

First, this is a regulatory measure, it’s a EU law that is enforced starting August 2. FWIW, Anthropic did more than the law asked for, but let’s leave this for another time. For now, just understand that this is coming from the top to the bottom, it’s not one AI player going rogue. Everybody will follow suit (Google already did it, even open sourced their SynthId SDK for this in 2024).

Second, the so called watermark is a steganography technique used to hide a message in plain sight, by encrypting token choices. In other words, Anthropic recognizes the text because it choses the next token according to a proprietary algorithm.

So, the output of the model is now permanently “styled” in such a way that it will always be traced back to the model.

Unless the text changes, that is. If you edit the output in a very meaningful way, the token choices are now broken, and the text is “free floating”.

Read the rest of the article here.

China Already Won the AI Race

For the last week I’ve been on a trip to China. Although I’m quite used to Asia (see below) I’ve never been to mainland China. My only previous contact was a trip to Hong Kong more than 15 years ago. I just returned yesterday, so the experience is still fresh. If you’re in a rush, here’s the TLDR:

• China is way more relevant than Western media portrays it

• It has a big advantage in AI, and the current context is favorable for an even faster evolution

• The “communism” vs “democracy” narrative has limited application here

Head over to the blog for the full article.

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading for an annonucement about my latest apps, Coloring Kiddo and ShareByLight.

Read more

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about the Singularity — and why I still don’t think we’re in it — through my favorite metaphor: a medieval chevalier dropped into a supermarket. If you’re premium, stay for a first look at **Adulthood**, plus the app promo codes (now on the paid side). If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

The Singularity and the Chevalier in the Supermarket

Everybody on X is screaming that we’re in AGI. In the Singularity. The world will never be the same. This changes everything.

But does it, really?

My test is simple. Imagine a medieval chevalier — year 1450, armor on, sword out — short-circuited through time into the same spot on the map. Only now there’s a supermarket there. Lights. Cold meat aisles. Tiny colored packages. TV faces talking in a language that almost sounds familiar and means nothing to him. He can still walk around. The world still feels almost hostile.

That cognitive gap — not a better chatbot — is what the Singularity would feel like to me.

What we have right now is a lot of compute crammed into very advanced prediction machines. You give them a status, they loop until that status is more or less achieved. Excellent coding tools. Useful. Impressive. Not “I don’t understand electricity, refrigeration, or chemistry” territory.

For me, the Singularity shows up when the *whole environment* changes that deep: interplanetary travel that doesn’t destroy your body, ubiquitous intelligence (yes, including the toaster), printing a replacement appliance at home, biological aging that actually stops. Unimaginable right now. Chevalier-in-the-supermarket mode.

We’ll get there eventually. I’m pretty sure it won’t be in the next 6 months.

Full post: The Singularity and the Chevalier in the Supermarket

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading.

Read more

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about the death of software engineering, and, if you’re premium, about my upcoming Coloring Kiddo app.

The 1000th Death of Software Engineering

I’m fortunate enough to have witnessed a few key moments in our recent tech history. One of them is web 3, the dot com revolution. Back then, having a website was considered peak flex (in modern slang). I see something similar happening with AI agents, now, it seems equally fashionable. And, just as back then people predicted the end of many verticals, now other pundits are claiming that software engineering is dead. Some business models will die, that’s true. The most prominent is the monolithic App Store business. But new ones will emerge. The world will go on.

These thoughts – and handful of ramblings – are packed in a 5 minutes video I just published on my YouTube channel.

If it’s the first time you hear about it, I recommend you subscribe, and learn why it is called “Bio Content“.

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading.

Read more

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about the poor guy guide to Fable-level vibe-coding and, if you’re premium, about how I’m automating my App Store and ASO processes (yes, there’s a free GitHub repo inside). If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe

Promo Codes

Before going forward, a little bit of housekeeping: as you know, all subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. This time, though, I chose to skip them.

Going forward, I’ll have to give it a serious thought, and decide if they will be still kept, if I will change the offer design or if I’ll skip them entirely. Don’t get me wrong, promo codes are great for product discovery, but not so much for retention. And this newsletter already covers a lot of ground for free subscribers, so maybe promo codes need to be somehow gated for the premium , these people are constantly supporting my work and I’m really grateful for that.

The Poor Guy Guide to Fable-level Vibe Coding for under $80/month

At the beginning of June I decided that the whole vibe-coding landscape entered a new era, one in which cheap but powerful model are replacing the SOTA offers from Anthropic or OpenAI. I ran a few tests, which resulted in me cancelling the Claude Code subscription ($100/month). I am now on a $80/month spending layer, but with much better options.

The article also contains a few mental adjustments needed when you switch from a monolithic model to a collection of specialized ones.

To all my subscribers, thank you for being around, and the premium ones, hold on, I’m going to talk about my App Store and ASO automation in a bit.

Read more

Subscribe now

Promo Codes

Before going forward, a little bit of housekeeping: as you know, all subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier. Make your pick below:

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of the world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager - this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

Vibe Coding for People Over Fifty

That would be my generation, fwiw. I was lucky enough to grow up before any digital malformation of our lives. I played in the dirt, took care of cows at my grandparents house in the country side, knew how to feed the pigs, the chickens and the turkey and slept in a chariot full of freshly cut grass (I will never forget that fragrance).

For decades, these made me a little bit obsolete. Then AI come. Now I can build whatever I want in plain English, and the main differentiator is not coding skill: it’s what questions I ask. And, just like that, the apparent obsoletion became my bigger advantage.

Because I grew up in a time where true interaction was a the core of our lives, unmediated by screens and the commoditization of attention, I simply know what matters, just like that.

By the way, when I say “I know”, I literally mean, “we know”, if you’re from the same generation. Anyway, here’s the link to the full article.

Introducing Mood 4 Love

I have 2 apps in my pipeline now, both in late testing stage, meaning those who signed up for TestFlight builds already have their hands on them. In the paid section, below, I will detail what one of them, Mood 4 Love, is all about.

To all my readers, thank you for your support, see you next week, and for the paid ones, keep reading.

Read more

I began June with a fixed plan.

I had already cancelled the $100 Claude Code Max subscription and committed to running a full month on MiniMax 3 at $20/month. The explicit goal was to test whether coding quality remained usable across real, large codebases at roughly one-fifth the previous cost.

About two weeks into that test, a different offer surfaced.

Grok was available at $35 for the first three months, then $35/month ongoing. In addition to the coding model, it included image and video generation that neither Claude nor MiniMax delivered at a comparable level.

At that point the original experiment was no longer the strongest available option.

I changed direction.

What I was not willing to abandon was the interface. I had built months of familiarity with Claude Code — the commands, the way it navigates projects, the overall flow. That accumulated comfort had practical value. At the same time, the economics plus the new creative capabilities on Grok were clearly superior to continuing the MiniMax test for the sake of completing it.

I kept the existing client and changed only what sits behind it.

The full technical steps — how I routed Grok’s model into the current Claude Code installation using a local proxy — are documented here.

This is the third distinct pricing and capability setup I have used in five months: $100 on Claude alone, $120 when I layered in MiniMax as backup, and now $35 with Grok handling both coding work and media generation.

The relevant point is not that every experiment must be driven to completion. It is that when a materially better combination appears midway, the rational response is to adjust the plan rather than continue the original test out of attachment to the initial framing.

The real constraint is speed of recognition: how quickly you can see that the current arrangement has been overtaken by a stronger one.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about how you can manage AI cognitive overload, and, for my paying subscribers, I share my latest discovery: how can you get paid for vibe-coding. If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

Promo Codes

Before going forward, a little bit of housekeeping: as you know, all subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier. Make your pick below:

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of the world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager - this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

Managing AI Cognitive Overload

There’s literally nothing you can throw at a decent LLM model these days that would NOT be created. Making something with code used to be the biggest bottleneck, now this is gone. And the biggest bottleneck shifted to… you.

I call this AI cognitive overload and it’s something that we, as a species, never experienced before. I first hit when I shipped 9 apps in 4 months. The first 2-3 were ok, but as the number grew, the complexity if just managing my workspace also grew… but exponentially.

So, I came up with a little strategy of how to solve this, and it involves 3 steps: increment in very small steps, save aggressively and take frequent breaks. I put all the details in this blog post.

Getting Paid to… Vibe Code

The last area of unclaimed attention was the feedback given by your agent while it’s building. You know, the “Discombobulating…” thing. Well, not anymore, there is a way to actually turn that into revenue. As usual, I share all the details to my paid subscribers.

To everyone else, thank you for being with me on this journey, and see you next week!

Read more

I shipped 9 iOS apps in 4 months - assisted by AI. This also almost broke me — but not in the way you’d expect.

The AI was mostly ok the whole time. It didn’t drift, never got confused, never needed a coffee. I did - especially the coffee part.

Here’s what happens when you try to move at AI speed: you end up context switching constantly. Nine different apps, nine different codebases, nine ongoing conversations with the AI, each one picking up mid-thought from wherever you left off. The AI handles that with no effort. You handle it by slowly losing the big picture, your passion, and maybe, fi you don’t address this upfront, your mind.

I calling this “AI cognitive overload”. It’s not exactly burnout — it’s more like running too many tabs in your browser and feeling all of them slow down.

When I realized this is actually serious, I started to engage it. I found out 3 things that helped.

The first was shrinking my steps down further than felt reasonable. The temptation with AI is to go big — hand it something chunky and let it run. That almost always goes sideways. It drifts. It solves a slightly different problem than the one you had. By the time you notice, you’ve lost a lot of ground. Small steps — one function, one screen, one bug — keep it honest and keep you in control. It feels slower, until you realize you’re making constant progress.

The second was saving obsessively. Not because I’m afraid of losing work, I am a big fan of version control, but because a commit is a note to future-me about where things stand. When I switch to a different app and come back three hours later, that checkpoint is the difference between a clean re-entry and 20 minutes of reconstructing what I was even doing. At the pace I was moving, that reconstruction cost added up fast.

The third was breaks. Actual ones, not “I’ll just fix this one thing first” breaks. The AI doesn’t need to rest. I kept acting like I didn’t either. Well, back to reality, here comes gravity: I did. The mental cache only clears when you step away. There’s no shortcut for it.

None of this is groundbreaking advice. Small steps, save your work, take breaks — you’ve heard it. What changed for me was understanding why it matters more now than before. When you’re working with a tool that never slows down, never loses track, and never needs to stop — you become the bottleneck. All the friction that used to pace you naturally is gone, and you have to pace yourself deliberately or the pace does it for you, badly.

The AI will keep up with whatever you throw at it.

The question now is whether you can keep up with yourself.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about why you should try migrate towards local, private AI models, and, if you’re premium, about my surprising findings after deciding to move from Claude to MiniMax 3. If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

Promo Codes

Before going forward, a little bit of housekeeping: as you know, all subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier. Make your pick below:

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of the world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager - this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

Local, Private AI Is Not Optional Anymore

I already wrote about this on my blog and here on Substack, but I just want to add something on top. Yes, the pace of AI is so quick that even a few days can make a big difference.

Since I wrote the piece on local AI, a new provider announced an upgrade to their model. It’s about GLM 5.2. It’s still a Chinese provider, and still an open-source model.

The difference: everybody who tested it says it’s on par with Claude Opus 4.8. That’s a huge statement, but that’s where the consensus seems to be.

Now, to run GLM 5.2 you will need AT LEAST $10,000 worth of hardware, either 4 Nvidia Sparks, or a big Mac Studio with 512 GB. That’s just the entry level, you’d probably need slightly more to run it comfortable. But that will put you literally on the same level with SOTA models, for which you paid, just until a couple of weeks ago, hundreds, if not thousands, monthly.

And with that, we’re segueing to the paid part of this newsletter, where I give an update on my quest on the Claude replacements I’m testing. To all my readers, thank you for being here, and see you next week.

Read more

There is a pattern we keep missing and then we tend to kick ourselves when we miss it.

I happened with social media, for instance. Let’s say you built an audience on a platform. Then suddenly, and usually without notice, the platform changed the algorithm, or got bought, or got banned in your country, or decided your content violated a policy you never agreed to. Your audience — the one you spent years building — was rented, not owned. You were a tenant, not a landlord. Most of us figured this out too late.

Well, this is happening again with AI.

Last week, Anthropic released Fable 5. Three days later, the US government ordered them to shut it down for non-US users. Anthropic complied. Not because they wanted to — they said they simply couldn’t filter by nationality fast enough, so they pulled the plug for everyone. Three days. Then gone.

I’m not writing this to criticize Anthropic. They were in an impossible position (and they kinda brought this unto themselves, but that’s another story). I’m writing it because the moment was clarifying in a way that’s important.

Here’s the thing: you probably don’t think of your AI setup as a dependency. It just works. You open the app, you type, it responds. That’s what dependencies feel like — invisible, until they’re not. The Fable shutdown made the dependency visible. And visible dependencies are the ones you can do something about.

The good news is that the alternative already exists, and it’s actually good now.

Running local, private AI models used to mean accepting a significant quality gap. You ran something that was maybe 60% as capable as the frontier, and you made peace with that for the sake of privacy or cost. That trade-off is mostly gone. The open-source models released in the last few months — Kimi K2, GLM-5.2, MiniMax M3 — are legitimately competitive with closed frontier models on the tasks most people actually care about: coding, reasoning, long-context document work.

And the hardware is there too. A MacBook Pro M4 Max runs 70B models at 70 tokens per second. A $3,000 AMD mini PC gives you 128GB of unified memory. An Nvidia DGX Spark is the size of a Mac Mini. You don’t need a data center anymore - a strong, recent machine will do it.

This isn’t about being paranoid or anti-establishment. It’s about a much simpler thing: owning your tools.

A carpenter who rents their tools from someone isn’t really a carpenter. They’re a temporary operator of someone else’s equipment.

We are on the verge of a moment — not here yet, but approaching — where the users of commercial AI become the product in the same way social media users became the product. Your prompts, your workflows, your patterns of thought will be valuable training data. The incentive to harvest that is enormous and the infrastructure to do it already exists. I’m not saying every provider will do this. I’m saying the incentive will be very hard to resist, and by the time it’s obvious, it’ll be too late to matter.

The window where you can set up your own, private and local AI, for a reasonable cost, with models that are actually good, is open right now. But beware, these windows have a way of closing without giving notice.

Full breakdown of models and hardware here: https://dragosroua.com/running-local-private-ai-models-how-and-why/

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about the difference between prompting your agent and writing loops for your agent, about attention harvesting with Tetris, and, for premium subscribers, my take on why Anthropic Fable’s model was banned by the US government. If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

Promo Codes

Before going forward, a little bit of housekeeping: as you know, all subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier. Make your pick below:

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of the world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager - this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

From Prompts to Loops

For about two years, working with a coding agent meant one thing: you wrote a good prompt, shared enough context, read what came back, typed the next thing. The agent was a tool and you were holding it the entire time, one turn after another. Like a sculptor uses the chisel.

A loop hands that holding to a small system that finds the work, hands it out, checks it, writes down what’s done, and decides the next thing — and lets that poke the agent instead of you. It’s like your job is to create the sculptor now – or, to be more precise, a small army of sculptors.

I did my first experiment with writing loops last week, and I documented everything on my blog. This was a real life situation, solving real life problems (app delivery consistency) and the loop itself proved to be rather useful. Have a look at the whole process, the bugs it found and the lessons I learned.

Harvesting Attention with Tetris

One of the most impressive benefits of AI is how drastically it shortens the path from idea to product. Here’s an example: the other day, when I was waiting for Claude to finish compacting my session, I had an idea. What if I can intercept that, and then fire up a small Tetris game that runs inside the Terminal (no need to leave the Claude window) just to have some fun. It took me less than an hour from idea to the github push.

What was more important, though, was the mental exercise in finding out how valuable attention became, and how the last spots available are popping out where you least expect them. Have a look at the whole brainstorming here.

To all my readers, thank you, and see you next week. And if you’re on the premium layer, keep reading for my thoughts on the Fable debacle.

Read more

A few days ago I wrote up a small script — a loop that audits my iOS apps for two dull bits of release hygiene. If you want the code and the war stories, that’s over there. This article is about something else. It’s the part I didn’t put in the technical post, because it isn’t technical.

It’s what the whole loop paradigm made me think about for the rest of the week.

Do we need another fancy name: “loop engineering”?

Boris Cherny runs Claude Code at Anthropic. He’s said, more than once now, that he doesn’t really prompt the AI anymore — he writes loops that prompt it, and his job has become writing the loops. He mentioned, almost in passing, that he uninstalled his IDE in November because he’d stopped opening it.

The internet did what the internet does. Within seventy-two hours there was a name — “loop engineering” — a wave of breathless posts, and a counter-wave insisting it’s the end of programming as a craft, that we’re all going to forget how to code while autonomous loops quietly fill production with garbage nobody can fix.

I want to sit in the uncomfortable middle of that, because I think both ends miss what’s actually interesting.

The sculptor and the chisel

After sitting for a couple of days with it, something finally clicked.

For roughly two years, using a coding agent meant holding a chisel. You type, you read, you adjust, you type again. The agent is the tool and you are the sculptor, working the marble one strike at a time. Good work, but it’s your hands on it the whole time, and there’s only one of you.

A loop changes your job. You stop being the sculptor and start building the sculptor — an almost autonomous system that finds the marble, makes the cuts, checks its own work, and decides what to carve next. And once you’ve built one, the obvious move is to build a few more and point them at different blocks. You go from one pair of hands to an army of them.

That’s the leverage everyone’s excited about. It’s real. But notice what just quietly happened: you’re no longer the one who knows where every cut went. That’s the whole game, and it’s where the excitement and the danger live, sharing the same narrow space.

What my boring loop revealed

The loop I built is genuinely unremarkable. A for loop with an API call inside it. An average developer writes it in half an hour. No architecture to admire.

And yet running it changed how I see my own portfolio.

Not because the AI was clever. Because the loop forced me to write down what “done” means. To ship my apps consistently, I had to state — in plain English, precisely enough that a machine could check — what a finished app looks like. In this case, I had to define 2 very clear outcomes:

• A cross-promo section linking to the others.

• A version string in its exact shape, so that when I’m juggling six or seven apps in a day I can tell at a glance which build is which.

The moment those rules left my head and became text, they stopped being vibes and became a standard.

That’s the part the hype misses. The value wasn’t the automation. It was being made to articulate the thing I’d carried implicitly for months. The loop is just the forcing function. You could get most of the benefit from writing the checklist and never running the script — but you’d never write the checklist, because nothing forced you to. The loop is the reason you finally do.

Addy Osmani, a Google engineer who’s written sharply about all this, makes a related point I haven’t stopped chewing on: writing your project knowledge down where the agent reads it every run is really just intent paid for once instead of over and over. An agent starts every session cold and fills any gap in your intent with a confident guess. I’d push it further — writing it down is how you discover what your intent even was. Most of mine was undocumented because it had never been challenged. The loop challenged it.

Where the skill-decline critics are half right

The fear is that if loops do the work, the human atrophies. True — for one kind of work, and false for another.

If you let a loop write code you don’t read, you’ll lose the ability to write that code. That’s just practice, or the lack of it. The critics are right about that.

But the loop I built didn’t replace a skill I had. It replaced a chore I was avoiding. I was never going to manually diff every Settings screen before a release. That work simply wasn’t happening. The loop didn’t take a craft away — it did a thing I’d quietly decided wasn’t worth my hours. And in doing it, it surfaced a real bug that had been shipping for months.

So the honest question isn’t “will loops make me worse?” It’s “what am I actually automating — a skill I should keep sharp, or an avoidance I’d dressed up as a decision?” Completely different things, and the answer changes per task. I’ll write loops for the second category all day. I’m in no rush to hand over the first.

The line that reframed the whole debate for me

Osmani ends his piece on something that’s stuck harder than anything else I read this week. Two people can build the exact same loop and get completely opposite results. One uses it to move faster on work they understand deeply. The other uses it to avoid understanding the work at all. The loop doesn’t know the difference. You do.

That’s the thing. The loop is neutral. It’s an unbound amplifier. Point it at work you understand and it compounds your judgment. Point it at work you’re avoiding because you don’t want to understand it, and it compounds the avoidance — faster, and with a clean console output that makes the avoidance feel like progress.

Your army of sculptors will carve whatever you point them at, beautifully, including a statue of your own misunderstanding. They won’t tell you which one it is. Only you can.

The completion

If you take one thing from this, take the reframe, not the actual script.

The loop is boring, just fifty lines of vanilla Python. You won’t be impressed and you shouldn’t be. But building it makes you do two things you’ve been putting off: state precisely what “finished” means for your work, and find out where you’ve been trusting a process you never actually checked.

That’s available today, with the tools you already have, for the cost of one boring afternoon. You don’t need to be Boris Cherny. You don’t need an army running while you sleep. You need one repeating chore you’ve been avoiding and the willingness to write down, exactly, what doing it right looks like.

Build one sculptor first. Make it carve one good statue.

I write about building a portfolio of small software as one person, in public, without pretending it’s tidier than it is. If that’s your thing, you’re in the right place.

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about the post 30-day challenge blues, we look at how Twitter resurrected (at least for me), and, for premium subscribers, give the early results of my Claude Code to MiniMax switch If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Subscribe now

Post 30-day Challenge Blues

Revisiting posts from 10-15 years ago, one at a time, in a 30-day challenge, proved to be a more intense challenge than I anticipated. The writing volume was not that bad. But the psychological consequences of revisiting so many beliefs and approaches was deeper and unexpected.

Most of the things I’ve written about seemed set in stone 10-15 years ago. Back then, they formed the foundation of my lifestyle, my business, my relationships. Revisiting that big slice of my life activities, with the “now” persona, put a very uncomfortable light on many of them. Some were shallow. Some were just borrowed motivational mumbo-jumbo. Some stuck, but not too many.

So, I’m in a little of a blues period, asking questions, waiting for the answers and just letting the dust settle over this. The 30-day challenge, as useful as it proved to be for my readers, accidentally brought a little bit of psychological chaos in my personal life. But chaos, in small dosage, is good for humans.

Twitter Resurrection

I wrote last week that Twitter seem to be living a second life, at least for me. I joined that platform in 2009, 17 years ago, and I’ve seen it go through many transformations. For years in a row I had a blast there, then there were years of silence.

The current algorithm, or just the quality of people there, seems to warrant the beginning of a new intense period. Recently, I’ve been participating in many interesting debates, seen many new founders and builders showing up, putting together businesses and interacting genuinely with their followers.

This is so contrasting with the previous Twitter management team, choked by the woke propaganda, that it seems like a completely new entity being activated there. Literally seems like Twitter is rising up from the dead. Let’s hope it sticks, you can follow me here, if you’re curious.

Promo Codes

All subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier.

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager - this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

To all my readers, thank you, and see you next week. And if you’re on the premium layer, keep reading for my Claude Code to MiniMax switchover.

Read more

Hello and welcome to another installment of my low-volume, no-nonsense, high-value weekly newsletter, where I recap what happened on my blog and, for the paid subscribers, give previews of interesting projects. Today we’re talking about a sudden surge in my Twitter activities, look back at the 30-day challenge, which just ended yesterday, and, for premium subscribers, do deep dive in switching from Claude Code to MiniMax (and potentially to a local first coding environment). If you’re not premium yet, you can subscribe below, it takes less than 5 seconds.

Twitter Surge

Something happened to my Twitter lately: I got a surge of new followers and I got days with 100k impressions overall, for my tweets and replies. This wasn’t in my plan, and I think it’s a combination of 2 things:

• some kind of threshold that triggered the algorithm and made me more visible. I do post targeted content, about AI, and vibe coding and X / Twitter is a good place for this.

• consistency: as part of my marketing activities I became more engaged and focused on more platforms. Coding has became the easy part, now the bottleneck is distribution, a.k.a. attention

I you’re on Twitter, let’s connect.

“Then and Now “ 30-day Challenge Finished - What Gives?

This is probably the most intense 30 day challenge that I ever attempted. It doesn’t show in the blog posts voice, and that’s the idea, maintain balance and keep producing relatable content. But it was tough. The challenge was started while I was in Europe, spanned Romania and Greece and was finished in Vietnam. Just making sure I won’t skip a day was a logistics feat.

It was also mentally taxing to go over to all the viral content I wrote 10-15 years ago and understand how it relates (or, to be more accurate, how it doesn’t relate anymore) to my current reality. All 30 articles are now side by side with their initial post in the main challenge hub post, if you’re curious.

Promo Codes

All subscribers to my newsletter get a handful of perks and one of them is the collection of promo codes. They are applied to all my existing apps (as of now) and they offer substantial benefits, up to 50% discounts on the heaviest tier.

• MosquiGo. A mosquito deterrent help, using a specific approach: 3 sound layers, combined with anti-habituation patterns. Promo code BUZZFREE, 50% off of Pro.

• 999 Wise Quotes. A curated and verified collection of world’s greatest minds, spoken and supported with AI reflection. Promo code 999NOW, 50% off of Wisdom Vault (full access)

• Shape Walk. Walk shapes to stay in shape, my very own fitness apps, which lets you design shapes over your neighborhood and walk them. Promo code WALKNOW for 50% off of the Pro annual plan.

• AI Kiddo. A simple app, made for kids, teaching your toddler their first 300 words in 8 languages. Promo code SMARTKID, 50% off of Premium.

• Just 5 words. An experiment in AI storytelling and image generation, all happening on device, no external API calls. Promo code FIVEFOREVER, 50% off of Pro.

• Zen Tales - Buddhist Stories. The app I wanted to build 10 years ago, when I was first exposed to real Buddhist teachings. Promo code: BREATHE, 50% off of Monk Wisdom (full access)

• Light Njoy - Turn your iPhone into a mesmerizing lamp, AI-powered. Promo code: NJOYLIGHT, 50% off for Pro

• addTaskManager- this is the flagship app in my portfolio, the Assess-Decide-Do implementation started more than 6 years ago. As of now, there’s a new tier: lifetime purchase - that’s right, no more subscriptions, if you want you can buy it and it’s yours forever. Promo code: ZENFLOW for 50% off of lifetime purchase.

• Flight Lens - Traffic Pulse - Flight intel and ticket prices. If you’re a digital nomad, you just need this. Promo code: LAUNCH, 50% off of the first year for the Pro layer.

• Hacking Habits - Flowing financial resilience - Automate finances with a privacy-first, habit creation app. Promo code: GRAVITY for 50% off of lifetime purchase.

To all my readers, thank you, and see you next week. And if you’re on the premium layer, keep reading for my Claude Code to MiniMax switchover.

Read more