Bit by Bit #5
We just crossed 105 subscribers! Triple digits feels different, just like I thought it would. Each Monday you choose to read this instead of doing something actually productive, and that means a lot.
This week I've been working mostly on my reading habits. Almost finished Atomic Habits and got really inspired about building good habits and dropping bad ones. Reading is becoming my main side activity to keep my mind fresh when I'm not coding, though I always keep one technical book in the rotation.
Current reading list:
To keep three books going at once, I use a simple setup. One paperback always in my bag, one title on audio, and one on my Kindle. The audiobook unlocked reading time I didn’t know I had.
I'm pushing myself to fill every free window with reading. When I'm putting my kids to sleep and just laying there - audiobook. Driving in heavy traffic - audiobook. On the treadmill at gym - audiobook. At this pace I can finish one book per week, which means 4 books per month. That's 48 books per year. Audio books is so amazing, I am so sad I was not using them in past.
My git diff renderer for iOS got its first really active contributor this week. Some guy added macOS support, then found performance bottlenecks, and today opened another PR with fixes. That's why I love open source.
We announced the biggest hackathon in Georgia for October 18-19, Hacktoberfest Tbilisi 2025, and our main sponsor is Lovable (https://lovable.dev/). We're bringing tech talks and workshops during those 48 hours. The whole Devtherapy team is part of it, and I'm genuinely excited about this one.
Raylib is still my favorite library for learning new languages. It has bindings in almost every language and it's super simple. The creator is a humble and great guy. Raylib is peak engineering and simplicity together.
Saw this tweet about how "vibe coding" is the future. And think it is totally wrong.
Coding is creative expression within hard constraints of boolean values. We're bitflippers who build different types of bitflipping on top of others bitflipping. That's our art. AI and beginners can't do our creative expression because they don't know the rules. Experts understand the rules and can bend/break them at will, creating their own rules. There's no such thing as "vibecoding properly." What you’re effectively saying is “I can splash paint on a canvas and create better than Van Gogh” Its just simply impossible.
Handmade hero is one of the biggest gem in the internet.
Doesn't matter if you're a game programmer or not. Watch the first 50 episodes from 700. While your colleagues fight over variable names, you'll be confidently delivering features. When others rely on second-hand programming dogmas, you'll be able to inspect and address performance directly. When language SDK is the lowest level others can operate on, you'll understand OS primitives.
Last week there was Apple Event. Most of it was the same stuff, not worth the hype. But 120Hz for all models is good, people with budgets can finally have more than 60Hz. AirPods 3 with live translation sounds interesting, but I wonder how it'll work in a coffee shop with multiple conversations happening.
My favorite announcement was the iPhone Air with A19 Pro. The A19 Pro has matmul acceleration in its GPU, like Nvidia's Tensor cores. This would make future Macs extremely viable for local LLMs. Currently, Macs have high memory bandwidth and VRAM capacity but slow prompt processing speeds. If M5 generation gets this GPU upgrade, the era of viable local LLM inference is coming.
"I never, ever, ever use ChatGPT or any other LLM in a way where I cannot verify the answer because you just cannot trust it."
And: "Please do not neglect developing fundamental skills the hard way. You need to fail. You need to learn. Don't get lazy and let LLMs outsource your critical thinking." This podcast had a lot of truth and no hype:
Found this great article about building an OS kernel in Zig.
This dude uses OpenSBI for portability and explains context switching really well, it hijacks the normal interrupt routine by swapping the stack pointer, forcing the CPU to restore a different thread's state. It's a clever demonstration of how time-slicing works at the lowest level. If you are interested in low level systems hacking and engineering, don’t miss this.
This piece resonates with me so much.
It describes that feeling when your internal processing speed doesn't match the world's demand for instant responses. The "blur" or "zoned out" feeling is a perfect analogy for that internal process. For a long time, this felt like a liability.
Neovim gets native LLM completions - Neovim now officially supports native inline completions for LLMs, configurable from LSP. It's sad none of my friends use nvim - I just want to drink coffee and chat about it.
See you next Monday.