My Honest Take on AI - 2026

I’m not an AI expert. But I know the math behind these models well enough to have an opinion.

The way I’d describe a language model to someone — it’s like a very advanced lorem ipsum generator. Except instead of filling space with random placeholder text, it fills intent with something contextually meaningful. It doesn’t actually think. But it operates at a high enough level that it often looks like it does. That gap between appearance and reality matters, and I think a lot of people miss it.

I’m genuinely on the fence about whether AI is a good thing overall.

The good part is hard to argue with. Compressing hours of reading and searching into a single prompt — that’s real. For people learning things, building things, solving problems — it genuinely speeds things up. I’ve felt that myself. And at a larger scale, that speed matters more than people realize. The same acceleration that helps someone debug code faster could help a researcher find a new drug, spot a pattern in climate data, or make a discovery that would have taken a decade of human hours. That part genuinely excites me.

The bad part is about control.

Whoever owns the most powerful models gets to decide what they say, what they refuse, and how they’re restricted. That’s not a small thing. I think about it like the Internet. Early Internet was open. Anyone could build on it, anyone could publish. But now Google, Meta, and a few others hold most of the traffic, most of the reach, most of the power over what information people actually see. They also made the Internet easier and more accessible — I’ll give them that. But the consolidation still happened.

I think AI follows the same path if we’re not careful.

The reason the server world didn’t get completely locked down is Linux. Because of open source, any company can run their own infrastructure today on just the cost of hardware — no license fees, no permission needed from Microsoft. That openness created enough of a counterweight that no single company could own the whole stack.

AI needs the same thing. Without open models, big tech will own not just the infrastructure but the intelligence layer on top of it. That’s a different kind of control than owning a search engine.

I’m realistic though. Large companies will always have the biggest models. The compute alone makes that almost certain.

But I think the more interesting work is in the other direction — small models, fine-tuned for specific tasks, that are actually good at what they do. This is basically the Unix philosophy applied to AI — do one thing, do it well, and compose with other simple things to complete something larger. The same way CLI tools are built to be minimal and chainable, models should be too. A well-tuned small model is often more useful than throwing a massive general model at every problem. I’d like to see more energy go there.

One thing I can’t get past — the environmental cost of all this is real and it bothers me. Training these models takes enormous energy. Running them at scale takes more. I don’t think progress justifies ignoring that. Future generations will deal with what we build now. That has to be part of the conversation.

The other thing that quietly bothers me is dependency.

When a tool is fast enough and good enough, you stop practicing the thing it replaces. That’s fine for some things. I don’t sharpen my own knives. But thinking feels different. Working through a hard problem — really sitting with it, getting stuck, finding your way out — that process builds something. I’m not sure what to call it. Judgment, maybe. Pattern recognition that’s actually yours.

There’s actually a useful comparison here from how India teaches math. Calculators are not allowed in school exams — through primary, secondary, all the way to board exams. The stated reason is to emphasize mental calculation. And it shows. Most people who went through that system can do basic arithmetic quickly in their head in a way that feels almost automatic. That’s not a small thing. That’s a cognitive habit built over years of not having a shortcut available.

The contrast with how students are using AI today is striking. It’s not being used the way a calculator is — as a tool for a specific, well-understood operation. It’s being used as a replacement for the thinking itself. Essays, code, analysis, decisions — handed off wholesale before the student has even attempted the problem. A calculator doesn’t make you worse at math if you understand the math. But if you never worked through the math, the calculator is just a black box you trust. And at least that black box is deterministic — the same input always gives the same output. You can verify it. A calculator doesn’t guess. AI does. It’s a probabilistic machine — it generates the most likely-sounding answer, not necessarily the correct one, and it does so with the same confident tone either way. Handing off your thinking to a tool that can be fluently, silently wrong is a different risk category than handing it to one that always adds 2+2 correctly. That gap matters more when the person using it hasn’t built enough understanding to catch the mistake.

The harder thing — and I think the more important thing — is to be deliberate about when you reach for it and when you don’t. Build the cognitive muscle first. Use the tool to go further, not to avoid going at all.

If you use AI tools — try to use open source ones where you can. DeepSeek, and other models being built in the open — they’re worth your time. Not because proprietary models are evil, but because a world with only proprietary models hands too much leverage to too few people.

The Internet we have today exists partly because enough people insisted on building things in the open. AI needs that same push, right now.

This is just how I see it. I could be wrong.