It seems that the era of HLMs (Huge Language Models) is definitely beginning.

“Now that the barrier to creating software has dropped so low, everyone is building custom applications for specific, personal use cases.” You’ve probably seen posts like this lately (or several of them). But it turns out there’s one group I doubt is applying that principle. I’m talking, precisely, about those who develop software.

If that’s your case, you probably already have a subscription to a coding agent. And if you do, why create an application for a personal need? Most of you will be creating CLI utilities and skills that let the agent work directly with the data, and then asking it for what you need from different interfaces. An application has to be modified if you ever want it to do something differently or add a new feature. An agent just needs you to ask. And if you do it with a technically specific and bounded language, it’s hard for it to get it wrong these days.

Last summer I took part in a Gemma 3n hackathon. In the write-up, I discussed the techniques I’d had to use to control the model’s output locally. I called that set of techniques ’taming’. I wish I’d used the word “harness” 😅

All the closed-source code you run on your machines will soon be closed no more. All the software powering your infrastructure will soon be open source. The end of an era.

The path toward fully autonomous software is moving like a steamroller. The landscape is so disruptive that it’s hard to predict not just what will happen next year, but how people will be coding a month from now.

With the goal of keeping up with it all, I’ve decided to hand over the maintenance of one of my side projects to an agent. Small tasks that used to require daily attention and are no longer on my plate.

This weekend I participated in Google DeepMind’s Nano Banana Hackathon.

The result is BananaMD, a web application for illustrating Markdown files (slides, documents…) with generated images. In recent years I’ve created all my teaching materials in this format and I like them to be accompanied by images, even if most are purely illustrative.

I also usually prefer to change them from one course to another. But thinking up the prompt, generating them, linking… It takes quite a bit of time.

The best advice that can be given to anyone who creates a product that makes use of LLMs: Use only open-weight models.

One of the most interesting use cases I’ve found for @PerplexityComet is to use it as a UX tester for the apps I develop. I’m sure someone has already thought of it, but I haven’t seen it before. If it can’t complete a user story, maybe it’s because your flow isn’t clear enough.

Also, if I had a website in production I would worry about systems like Comet being able to complete most of the functionality without a problem.

Imagine studying with AI for months, but every conversation starts from scratch. The model helps you well… but doesn’t remember you’ve mastered algebra while struggling with trigonometry. We have “memory” in LLMs, but not for learning.

When you search for AI in education, you find the same everywhere: pre-made prompts, task automation, automatic grading… @emollick has written extensively about this challenge of moving beyond surface-level AI integration in classrooms.

There are promising initiatives like @karpathy´s @EurekaLabsAI that I’m excited about. But there’s one specific aspect I find fascinating: applying LLM memory concepts specifically to learning progression tracking.

Software engineering work has always changed continuously, but in the last two years that pace has skyrocketed. There’s constant speculation about its complete automation, but I think we’re not focusing enough on the immediate consequence of that.

If that work becomes completely automated, if truly any software product can be created on the fly, without any kind of supervision… Then, not only would software engineering jobs disappear.