When GPT-3.5 Turbo dropped, we all tried asking it how to make meth and got an answer we didn’t want, unless we jailbroke it.

Well, this isn’t a problem nowadays, since local LLMs can be uncensored. All you need is some memory. You can run pretty bad models right on your phone or pretty good ones on your PC.

Memory is your biggest constraint, since it determines how big/smart your model can be. Model size is measured in parameters, so you’ll see something like 27B right after the model name. Note that proprietary models such as GPT or Claude don’t disclose their size, but you can always search for estimates.

Recently, I finished Metal Gear Solid: The Twin Snakes (a remake of the first MGS; the original is superior btw), and it reminded me of my favorite speech in all of gaming. It comes from the ending of MGS2.

BTW, the speech doesn’t contain spoilers, so you can watch it even if you haven’t started playing the series. But I still recommend waiting until you finish the game to hear it in all its greatness.

As we all know, AI tools are constantly getting better.

Every ~4 months, there’s a new SOTA model, and the last-generation models (~4 months old) are basically no longer worth using.

So, should we just stop coding and rely only on AI agents to do all of the architecture, design, and implementation?

NO.

I use AI daily for my job and for other unrelated, complex work. AI agents are very good if you know what you are doing, and they will massively increase your productivity. They will save you a lot of time during those six-hour debugging sessions where there’s one small edge case, but you just can’t find it. Or when you need to do some annoying array indexing, get it exactly right, and are losing your mind trying to keep track of all those indices (I personally suck at algorithms). Basically, AI agents don’t care how boring a task is. They can work nonstop for 10 hours on the most boring debugging sessions.

In my life, I’ve written a lot of libraries. A lot of them are unfinished; some are finished. Over that time, I’ve kind of learned what makes a good library.

Have you ever asked yourself: do I even need to write documentation or a getting-started guide for my library? My library must be perfect. I WROTE IT. I UNDERSTAND IT. Well, that doesn’t mean some random guy on the internet using your library will understand it.

In the last article (The Joy of Learning) I described that it’s actually really fun to learn new things. But how do you even start learning in this day and age, with countless distractions like TikTok and Reels?

If you’ve been wondering why you’ve been so unproductive while other people still get shit done, it’s usually because they are not distracted. If you open Instagram the second you feel even a tiny itch of boredom, you’re basically frying your dopamine and distracting yourself instantly. So you never get that feeling of wanting to learn or build something.

Often when I ask people what they wanna do, like 80% of the time they answer “I wanna do nothing.”, “I wanna be rich.”, or some generic crap.

Most people are uncreative. They just wanna be rich without doing any work.

When I do “nothing” (or something very boring) for a few days, I get bored and start needing to do something creative. That creative thing can be a programming project or something completely different.

I remember in 2022 when OpenAI released GPT-3.5. It was a really interesting model at the time, and we didn’t know what consequences of AI we’d be dealing with today. I used it mostly for annoying school stuff like writing protocols. But for programming it wasn’t great, because all it could really do was write a simple function or a small algorithm.

Nowadays big companies say they’ll replace devs every 3 months, but we already know where that leads. AI is still not good enough for building big applications.

Hello World!