Let's get back to the dog house —I've started implementing the next step of the plan: reinforcing the surface.

I originally intended to do this in the style of papier-mâché and even bought a roll of thin kraft paper for the purpose.

Papier-mâché involves soaking the paper in water and then in a diluted PVA glue solution. This approach was a complete failure—the kraft paper started falling apart just in the water, and when applied to the frame, the whole structure became overly wet and softened.

So, I scrapped that idea and returned to a simpler one: applying glue directly to strips of paper and sticking them onto the frame. The result wasn't perfect either—it takes a massive amount of time, but at least it's producing results.

Next, I’ll add a few more layers until the surface is rigid enough for putty—after that come sanding and composite work.

If I were starting over, I’d seriously consider printing the entire frame with a 3D printer, even if it meant two weeks and 10 kg of plastic.

I decided to install a roller blind so that the light wouldn't disturb my sleep in the mornings.

After looking for ready-made solutions, I only found separate motor units and separate blinds.

I didn't order a complete roller blind—instead, I only ordered blackout fabric in the required dimensions. That's where I shot myself in the foot.

There turned out to be two main problems:

1. The fabric arrives folded many times over, and even after ironing, it’s not perfectly smooth.

2. The fabric doesn’t have a bottom weight, which is almost essential for proper tensioning.

So I bought a 6×6 mm profile in a hardware store for weighting. Then I modeled inserts in Fusion 360 that clamp the fabric and then get tightly fitted into the profile. I had to dust off the resin printer—some parts failed to print, but I had enough successful ones with some to spare.

I was away for a short while due to a vacation trip to another country — Kyrgyzstan.

I try to travel to new places several times a year.

Here are some interesting things I noticed from an engineer’s perspective:

1. We stayed on the 16th floor, and there was a gas stove (in my country, gas is prohibited above the 10th floor), even though the region is seismically active.

2. Throughout the city, there is an "aryk" system used for irrigation (watering parks and green spaces) and cooling. Allegedly, the water flowing through these small channels helps lower the air temperature. I won’t argue, but the efficiency doesn’t seem particularly high.

3. I rented a car for a few days — and there, the main units of measurement were miles, so it took me some time to get used to the imperial system and set the cruise control to 56 mph instead of 90 km/h.

Here are a few photos for illustration, and I’ll get back to my regular schedule of technical posts during the week.

A new device for filming (yes, there are devices, but no actual filming yet) – macro lenses for a phone.

Two pieces, each with x20 magnification, complete with a clip and lighting.

Unfortunately, my phone doesn’t have a native macro lens at all – only an ultra-wide and a telephoto lens (the latter up to x5).
So I had to upgrade the main camera with some extra gear.

Example videos show footage from the original camera, with one lens (x20), and with two lenses (which should theoretically be like x400).

I’m not sure yet where to use two lenses – the macro is just too strong. But one works quite well; perhaps I should also consider intermediate options like x5, x10, or x15 to cover the full range of possibilities.

About GPT-5 and Cursor

A week ago, a new model was released that was offered "for free" in Cursor.

I decided to use this opportunity to build a pet project for tracking income and expenses.

I’ve been tracking all my income and expenses in Google Sheets for the past six months. The main goal is to understand my average monthly expenses.

I even managed to "vibe-code" a Telegram bot to help with these spreadsheets, but now I felt like trying something different. And the models have improved a lot in the last six months.

So — GPT-5 is like night and day (though not drastically better than Claude 4). I deliberately chose a stack I wasn't familiar with — Go, React, and gRPC. And GPT-5 helped me build a working MVP in just a couple of days, which I’m now refining to fit my needs. I’ll announce the launch separately, but for now you can check out the repository.

P.S. The free GPT-5 in Cursor turned out not to be free — it stopped working after $200 worth of API usage (all under the $20 Cursor subscription).

Everyone knows what an RNG is—an algorithm that generates a random number.

However, all algorithms are susceptible to attacks and vulnerabilities, so services with high security requirements tend to avoid relying on them—cryptography, encrypted communication, digital signatures, and so on. Some banks or systems ask users to move their mouse or press keys a certain number of times to generate random events.

Physical RNGs can be hardware-based or rely on physical phenomena, but in any case, they use a real source of entropy.

Some unusual sources of entropy used or studied:

1. Lava lamps – the chaotic movement of wax in dozens of lava lamps.

2. Geiger counter – intervals between events of radioactive decay.

3. Goldfish in an aquarium – random movement of fish in water.

4. A laundromat with a microphone – noise from washing machines.

5. A dance floor with pressure sensors – chaotic movements of people.

So your security might depend not only on software and algorithms, but also on fish in an aquarium.