vitalnodo

Before, I would write an emulator of an old console. On weekends I would debug why the game shows garbage instead of a picture, compare opcodes with the documentation and learn from my own mistakes. And the emulator itself would be valuable as an artifact, because it holds the experience I accumulated. Now the same thing can be done much faster. Learning from mistakes disappears, only high-level templates remain which looks similar to the academic approach.

On the other hand, you can set up an LLM so that everything goes gradually and you learn like you used to. But that's more like what I would call pet archaeology or something... Learning the tools doesn't give as much benefit anymore, it's more like a reconstruction of past events.

In the essay Hackers and Painters it was mentioned that there is a project approach to learning and an academic one. The idea of pet projects is killed by language models, because what you did on weekends, a language model now does in half an hour. So you can try the other approach: maybe you should have more general knowledge right away, best practices, so more of an academic approach. In the academic approach you learn from the best reference examples, develop taste, and only then recognize what is good or bad and build your own. But here too you can see that LLMs iterate much faster, while you have to collect patterns for years, and you will also always be in the role of the one who is behind.

Nice presentation, I've been asking myself about exactly this lately. Speaking for myself, I like to build a small copy of some big project, and then I put big hopes on it and start feeling guilty that it's unfinished. It turns into a kind of ban on picking up anything else, while the big one keeps getting put off too. And here there are so many different styles and approaches laid out.

There is an Android Translation Layer now, and there were different attempts earlier.

As I found out recently, there's a lighter option: model checkers like Spin. You describe your synchronization logic in a small modeling language (Promela), and Spin tries every possible interleaving of that model.

Mnemotechnics is an art of memorizing by linking visual images. A book I have gives these criteria for a proper image:

you can roughly draw the object from memory

you can name its main parts

you can name its main colors

you can describe its spatial orientation

The image shouldn't be as clear as a hallucination, and it should be a memory of a specific object, not a generic one. Meeting these criteria is already enough to link another image to it, for example by piercing it with another object, though there's also the question of imagining it at a large enough size.

Aphants, do these hold for you? Take Doge, the popular meme dog: not a generic Shiba Inu, but that exact picture. Could you sketch it, name its parts and colors, and say which way it's facing? And could you imagine, say, a pencil piercing it?