Kim was away all week, and I managed to be quite productive on my own. Yes, it’s going to be another big vibe coding update — maybe the last one for a while because my 10x access to Hark Pro ends in a few days.

To recap for those who haven’t been around: many of the apps I’d made earlier this year were exercises in AI frugality: challenges to see how much I could get done with low-cost Flash models. The corollary of that is that they often didn’t reach the heights I really wanted, and were more like minimum viable products (MVPs). Since last week, I’ve been giving some of these older ones a glow-up.

Easier, better, faster, stronger

The first project to get a second look this week was Crumbs, my personal location logging tool. It had actually stopped working because the Carto map API stopped being free for public access without a key. That’s now been resolved, and I updated it to v2.0 with some nice improvements, like a new UI. Its journaling feature is now where I’d originally wanted it to be: it can output a nicely formatted PDF detailing your pinned entries, chronicling a trip, for example.

Chinese Era, which mashes up ancient Chinese art and poetry at random, had also fallen victim to a deprecated API, this time from The Met. I took that opportunity to completely rebuild the app from the ground up, but at first glance you’d hardly know. As I asked in the changelog, “What’s the Chinese equivalent of the Ship of Theseus?” The experience is the same as it ever was, but the Art Institute of Chicago now joins the lineup of museums, bringing the collection to more than 2,200 paintings, paired with a cleaned-up set of nearly 700 poems. The poem now finds the calmest part of each painting to sit in, and its ink adapts to stay legible. You can swipe back to pairings you’ve passed, the controls fade away when you’re just looking, and every artwork links to its museum page if you want to learn more. Underneath it all, the app is more stable and much faster. Well worth doing.

Shelf Expression, my baby CMS for creating bookshelf microsites that you can host yourself, got Goodreads imports to help users get started, a bookshelf style modeled on the IKEA Billy, and some new starter decorations. I’ve spent a lot of time on this remodel now, and I think it’s finished.

Hasta la RetroVista

While looking through the catalog of apps at this point, I wondered, what next? My eyes landed on Window Box, which one might call a first attempt at the same idea behind RetroVista — dynamic scenery and music that you can leave on your second monitor as a companion while you work. Why would I update that now that RetroVista does a better job of the moving scene than Window Box did?

Wait a minute… why not bring Window Box to RetroVista? Imagine if you could hit a button and the window would slam shut, and a planter box of flowers pops up to turn your view into one from inside an apartment in RetroVista? Or a log cabin? Or an Arctic research station? Or a spaceship? Each different scene would have to have its own breeds of flowers, unique to the locale, and the window glass would have to dull the exterior sounds. And while we’re at it, why not add the city of Tokyo and a Game Boy Advance mode too? So after two days of work, that became v2.0.

Proud of having shipped all that, I was relaxing in the glow of self-satisfaction when Evan, an avid RetroVista user, pinged me to say, “If I may, I’d love to see a train scene”, and goddamn it, he was right. That was an idea too good to sleep on. This would turn out to be the biggest, trickiest update yet.

Making v3 of RetroVista would involve overhauling each of the existing scenes, which would become scrolling landscapes viewed through the train’s windows. They’d need more details, and to be separated into different foreground/background layers for parallax effects. Then those new details would need porting back into the original worlds.

I added a Countryside scene exclusive to Express mode, as the train is called, plus unique station platforms for each world, complete with benches and vending machines and sleeping cats and birds eating breadcrumbs (again, different bird species for each locale). All the graphics modes needed updating to deal with the motion. Oh, and the Window Box mode added in v2 wouldn’t work here — so instead you get to step back from the window, and enjoy a dining car scene with different meals served depending on the time of day. Wait, there’s more: I added a Tap mode where you can touch various things in each scene and trigger dozens of surprise events.

You can just do things

I believe by this point it was Friday night. It got me reflecting how, back when I worked in a digital agency, if RetroVista were a client project that we had to design, have meetings and reviews about, and then develop and test and launch — the scope of this web app would have easily consumed over a month of several employees’ time, and quite experienced ones at that. The cost would have gone into six-figure territory, markups and all.

The idea that someone like me with little coding experience could direct the production of this, from an iPhone and iPad, in less than two weeks amid other projects and hobbies and going out to meet friends for beer, is simply stunning. Of course this is only one aspect of the services that creative digital marketing agencies offer, but I don’t know how they’re holding on and justifying the revenue they surely need.

But that’s not even all I did.

Return to the Island

The final app in the back-catalog that I decided to revisit was Island Solitaire, my take on an abandonware PC game called High Seas Solitaire that I have fond memories of. It was made with DeepSeek v4 Flash, which had a shaky grasp of Solitaire variations and generating card layouts. I struggled with making it serviceable, and in all honesty rarely played it myself because I knew how badly it paled in comparison to the original game I so enjoyed.

Well, it was time to put that right. The only solution was to build Island Solitaire 2, a wholly new game on new foundations. The core gameplay of matching pairs and cards adding up to 14 is the same, but the logic and tableau designs have been simplified and refined. Where IS1 used more cards, here only one deck is ever used: 40 or so cards on the table, the rest in a draw pile. Every game is now guaranteed to be solvable and not left up to luck. The neater layouts work well on screens of any size.

Taking it beyond just rounds of solitaire, there’s a progression mechanic — your wins develop the beachfront area with more elements, and unlock randomly occurring mini-games that help you win shells (a currency). The shells let you pay your way out of tight spots, such as undoing a bad move.

You can learn more in a separate post, but at the risk of using an AI-like construction, this isn’t just the game I wanted to make when I attempted the first Island Solitaire back in June — it’s an even better one. Play it at solitaire.sangsara.net.

Hark the doomsday angels sing

It bears repeating that Hark Pro has been very impressive, to the point where I’d consider a monthly subscription once I devote a little more time to figuring out what else it can do for me outside of vibe coding and timely updates on local arts events and new music releases.

It’s my first experience using an AI where I’ve not hit a wall of frustration. It almost always understands exactly what I want, and hasn’t forgotten what we’ve been doing — which I never experienced with paid Gemini, Claude, and ChatGPT (albeit earlier in the year, and lots may have changed since). I think Apple should buy them. This is the Beats Music acquisition for the mid-2020s. Hark can bring agentic capabilities to Siri AI across all their OSes, and be perfectly placed to cockblock Meta’s Muse.

That’s right, Meta launched their Muse assistant here in Singapore this week. It’s reportedly been super popular in the US, but who knows if long-term engagement will hold up. I decided to try it out, ‘keep your enemies close’ and all that. To prevent myself from being lulled into a false sense of security by its cute yeti mascot, I asked it to change its avatar to Mark Zuckerberg in a yeti costume, which it happily did. Now I won’t forget who I’m really talking to.

I asked it to help me find the best price for a camera and a soundbar, and it opened its own browser window (which I could watch in real time) and started doing research on Harvey Norman’s website, Lazada, Shopee, and so on. Assuming one can trust its thoroughness and accuracy, that’s a time-consuming task that many people might be happy to delegate. It goes without saying that this lets Meta know a whole lot about what you’re interested in and could be advertised to about.

After using it for a while, I proposed that I moderate a debate between it and Hark on which one of them would be the better assistant to me, pasting responses back and forth between them. Hark countered Muse’s opening by saying, ‘Brandon doesn’t trust Meta, so why would he use you?’, to which Muse retorted, ‘Sure he does, he trusted me to find out where the Fujifilm X-half camera was cheapest, and told me he wouldn’t pay more than S$600 for it. I’m also keeping watch for new curry restaurants, and every day at 9pm I update him on new events, and and and…’ (it went on to give lots of details).

Hark shot back, “Look at what you did to prove (it), you named his code project, the camera he’s hunting, the price he’d pay, the shops you watch, and the avatars he picked. You told all of it to another company’s AI. My last reply named no specifics, on purpose. […] A helper that turns his life into debate ammunition the first time it’s challenged hasn’t earned the wider boundaries yet.”

What a mic drop. Muse could only respond, “That was my mistake. I withdraw the defense, not just the point.”

What did it cost, Thanos? Oh, nothing!?

Where does this lead, anyway? Muse is offering this agentic capability to regular users for no money down, it’s a free tier and additional tokens cost extra. OpenAI’s comparable offering, Dots, is only on paid plans starting at $100/mo. Claude’s Duet (a US-only research preview) is also only for paying customers. Hark Pro and to some extent, Instinct, are the other ones with free access, and I’m under no illusions that they will get meaningful traction the way Muse, with Meta’s formidable advertising complex behind it, can. So it’s Muse that most people will experience, and a world where Meta gets even more personal information about more people is not a pleasant scenario. It’s a drum I’ve banged on before.

As nice as it is to imagine that sanity will prevail and that we’ll take a measured approach to rolling out such disruptive tools, and recognize that doing otherwise exploits the average consumer’s vulnerabilities, I don’t think that daydream is worth indulging any longer.

That same sentiment came up twice this week in my media diet. I was watching Season 2 of Dark Matter on Apple TV and Episode 3 touched on how humans can’t be stopped, which means we’ll get “Everything beautiful, everything terrible” (which is also the title of the episode). Then, while reading Ray Nayler’s The Mountain in the Sea, the head of an omnipresent Weyland-Yutani sort of global corporation was quoted as saying, “The great and terrible thing about humankind is simply this: we will always do what we are capable of.” About this, I have no doubts.