svpino.com

The language of the future

2026-02-25T00:00:00Z

When building agents, we care about two things: intelligence and cost.

We want to maximize intelligence while minimizing the cost of producing it.

Should agents even produce human-readable code?

Human-readable code is an intermediate layer that's only useful if humans must stay in the loop.

But what happens when agents become truly competent? At that point, humans will become a bottleneck.

Future agents may not write Python or JavaScript, or any other modern programming language. They may generate something entirely new, maybe even closer to bytecode.

It will be unreadable to us, but much more efficient for them.

Mathematics of Machine Learning

2025-05-30T00:00:00Z

I met Tivadar during Covid.

We were all stuck at home, unsure what to do with all the extra time, so we started talking about building something together.

I wanted to teach people Machine Learning. I had this idea about building a website that would ask random questions for people to answer. I wanted the site to do a hundred different things, but one thing was non-negotiable: I wanted people to leave feeling they had learned something different.

Tivadar was the answer to that.

Machine Learning is tough, and unfortunately, most educational content you find online suffers from chronic handwaving syndrome: overused buzzwords, skipped intuition, and more confusion than when you started.

At the time, Tivadar was already writing online about math. He wasn't the only one, but he was different. He was taking seemingly mundane topics and telling a story around them that was surprisingly effective.

There wasn't any handwaving or burying people under a mountain of theoretical ideas. The writing was different, sharp, and fresh.

I had never been excited about math before. I read every single one of Tivadar's posts. I wasn't just learning the rules—I was learning how to think. And—shockingly—I was entertained.

I had never seen that combination before.

I asked Tivadar to help me with the site, and he did—for a while until he decided to move on to start writing this book. I remember telling him I understood, but I was secretly sad. Really sad.

Today, I'm thrilled this happened the way it did.

Mathematics of Machine Learning is the inevitable consequence of those short posts that excited me about math for the first time. It's not just the best book I've read on the subject—it's the one I wish existed when I started.

This book does something rare: it teaches you the math behind machine learning without boring you with vague concepts—or making you forget why you showed up in the first place.

The book is laser-focused on what you need and says nothing about what you don't. The explanations are vintage Tivadar: sharp, detailed, and entertaining. You can't just read or memorize them; you'll understand them.

I've been reading this book since it was an idea and a bunch of notes and sketches. I've watched it grow from online posts to something polished and powerful. And I've learned a lot—not just about math, but about how to explain math.

I'll leave you to it. You're in for a treat. Enjoy the journey. I know I did.

The writing's on the wall

2025-04-11T00:00:00Z

If I had a nickel for every person who gave up on Large Language Models because they didn't find them helpful after ten minutes, I wouldn't be sending my kids to public school.

People learn about a tool for the first time, try it once or twice, don't like the results, assume it's useless, and convince themselves they're too smart to benefit from it.

I understand the average Joe is too busy to care, but professional software developers turning up their noses is baffling.

How do I put this delicately?

Technical people who don't embrace Artificial Intelligence won't make it. And I don't mean they'll "fall behind" or have a "hard time" or any of your favorite euphemisms. I mean it as it sounds: those believing they are above what's happening will end up working in some shoelace factory. At best!

The writing is on the wall.

Artificial Intelligence went from something cool to a revolution changing how we design, develop, and maintain software. It's a force multiplier like nothing we've seen. It's a massive advantage to those who know how to use it, and it'll be the end of the road for those who ignore it.

AI won't take your job. A person using AI will.

Once that reality sinks in, we gotta answer a much more interesting question: where should we focus to ensure we don't get steamrolled? There are three areas I'd recommend every developer look into:

Learning how to use AI as a copilot
Learning how to be a copilot for AI
Learning how to build tooling around AI

So far, the first category has attracted the most attention from technical folks. We have tools, papers, videos, courses, and a lot of money focused on augmenting people. Using AI as your sidekick is table stakes. If you ain’t here yet, you're already playing catch-up.

The third category is newer but growing at lightning speed, with things like the Model Context Protocol (MCP) and the Agent2Agent Protocol (A2A). These promise to be the pillars of a future where autonomous agents are the protagonists, not people. It's uncharted territory. A few folks will build billion-dollar businesses on top of it.

The second category is different. How would anyone in their right mind let an imperfect tool take the driver's seat?

I've spent decades cementing a career based on my ability to build good, working software. The simple idea of copiloting is terrifying—and, sometimes, I like to think, irresponsible.

I've criticized many people who've shared software they didn't understand. I've even used my smug face while letting them know how irresponsible they've been and how I'll never let my kids do the same.

But it might be time to swallow my pride and really grapple with what's happening.

People are building software without understanding what's happening behind the scenes, just like I drive my car without knowing the first thing about combustion engines. I'm not a mechanic, and they aren't software developers, but we’re still getting where we need to go. And you know what? It doesn't matter.

This phenomenon is not new. In fact, it repeats itself right before the new replaces the status quo. Every major shift comes with holdouts, convinced their way is the one way. And every single time, people like me who dig our heels in become a footnote in history books.

I've always told myself I'm much more modern than my parents were. "Look—I embrace technology! I know about the latest and greatest!" Except that's bullshit. As soon as something threatens my comfort, I start screaming and pointing to my titles and résumé.

Copiloting AI will be a big part of the future. People will keep building software without formal training. Although imperfect today, their software will improve. We all know where this is going.

I recommend you learn how to help Artificial Intelligence build software. I recommend you become comfortable controlling and guiding, not only driving.

This is happening, and it's not asking for our permission.

Coding is over. Now what?

2025-04-03T00:00:00Z

I never thought things were going to change this much.

From the beginning, it was obvious that Large Language Models were not a joke, but I thought they had fundamental limitations preventing them from getting too far.

I was wrong.

Around 2012 or 2013, my former employer went on a hiring spree for software developers. The criteria to qualify for a role were owning a computer and convincingly stating you wanted to learn programming. The bar was low. We had the budget and patience to train people in basic front-end coding, and we needed warm bodies, so we hired everyone, some unqualified.

It was a different time.

Large Language Models have blown past every limitation I had imagined to a point where today, you'd have to be crazy to hire untrained people to write code. I'd go as far as to say you'd be crazy to hire anyone without "significant" experience in the field. We might disagree on the meaning of "significant" in this context, but it's clear the bar is much higher because of AI.

Change in this field is nothing new. I've been building software for long enough to appreciate how much we have to reinvent ourselves every year. But it has never felt this fast.

Every week, we get a new version that leaps and bounds around everything we've seen. Every new model is faster and better. They are now solving problems that nobody—and I mean nobody—thought possible a few years ago.

Yes, of course, a part of me is worried! I never planned for this future, and I still don't know how much things will change in the coming years, but I also have some encouraging thoughts.

The fundamental fear in everyone's mind is whether we'll wake up tomorrow and find a model that can do everything we do today but cheaper, faster, and better. I don't think that'll happen any time soon.

Large Language Models have changed how we write code, but code is just the expression of a much more complex process. Coding is how we communicate our ideas to a computer, but that has never been the hardest part.

The most competent people you know didn't earn that title because they remembered what code to write, memorized documentation, or typed faster than anyone else. Instead, they consistently do three things better than everyone:

They know how to identify the right problem to solve.
They know how to frame this problem in ways where finding the correct solution becomes inevitable.
They know how to shape that solution into an elegant, maintainable, and scalable design.

When you get to this point, writing code to solve the problem is easy, and that's where Large Language Models shine.

Coding will become fully commoditized in the coming years, but your brain will not. Artificial Intelligence is a hell of an assistant, but it can't replace people yet.

In fact, I believe the opposite to be true.

Every single invention in the history of computing has enabled more people to write software, not fewer. The easier something gets, the broader its inner circle becomes. Artificial Intelligence will help the artist write code. The accountant, the writer, the banker, and the history professor can now join a club that was previously exclusive to a few.

We'll build more software. Faster and—hopefully—better software. The amount we need is unbounded, so we'll keep building.

Assuming you've been writing software like me, I'd recommend you learn as much as possible about AI and how to use it to write better code. With the excitement and the amount of information out there, this shouldn't be hard.

People asking others to ditch programming because it doesn't have a future are doing you a huge disservice. History will not look kindly on them.

Building software is here to stay.

It will look very different—it already does—but there's never been a better time to be a developer.