sethserver.com Subscribe
The Third AI Future: Boring, Practical, and Quietly Useful | Small Models That Actually Work

The Third AI Future: Boring, Practical, and Quietly Useful | Small Models That Actually Work

By Seth Black • Updated: March 18, 2026

AI · 4 min read

Like this kind of writing? Get one email a week with notes on startups, AI, and the occasional strong opinion about Python: subscribe to the newsletter.

I like sci‑fi as much as the next nerd. I grew up thinking AI would either give us flying cars or turn us into human batteries. But the future I keep running into is a third option: boring, practical, and quietly useful.

It's also the only one that works in real life.

The "third AI future" is the one nobody makes movies about

The hype cycle loves extremes. Either:

Real Machine Learning work is usually closer to: "Can we reduce errors by 3% without breaking the existing workflow?" Nobody buys a ticket for that. It's still a real win.

And yes, I'm guilty. I've built ML systems that were... not heroic.

Some of my ML projects were not helping humanity

I've worked on ad tech systems that picked ads and popups for websites. You know that moment where you're trying to read an article and a newsletter modal lands on the page like a brick? Yeah. That genre. I helped make those happen. Not proud.

I also built an NLP system that was supposed to help with job seeking. The pitch was: match resumes to job posts using text. Clean idea. Messy reality. Job matching depends on a pile of stuff that isn't in the text. Timing. Location. Internal referrals. Whether the company already promised the role to someone's cousin. The model can score well and still waste the user's time.

Then there was machine vision in cold, dark, damp parking garages. Computer vision demos love bright lighting and clean backgrounds. Parking garages are the opposite. They're dark, wet, and terrible for cameras.

What I learned wasn't "ML is bad." It was simpler:

ML is not a silver bullet. It's a specialized tool for a specific, well-defined job.

Small models, real constraints, and why edge wins

I worked at a physical product company that had a successful crowdfunding campaign. The genius dev behind it used the fame to land a FAANG job, which left a software crater behind. I walked into that crater.

That job hammered a few things into me:

It also taught me something ML people love to forget: smaller, specialized models can beat bigger general-purpose ones if the task is narrow and the data is real.

That's why I'm so interested in tiny language models that run on a phone. Not everything needs a cloud call, an API bill, and a privacy policy nobody reads.

And yes, I want this stuff in manufacturing. Give me a small robot, a small chip, and a clear goal. A Jetson GPU attached to a few small servers could handle real factory work.

The boring ML applications that actually matter

Here's the ML future I actually want:

Daily, unglamorous measurement

None of these need a chatbot personality. They need consistency.

Control systems that quietly improve the world

My favorite example is vehicles adjusting fuel/air mixtures based on environmental data and driving habits. That's not sci‑fi. That's immediate, measurable impact:

It's the opposite of flashy. It works anyway.

A future of tiny chips doing tiny jobs

The "third AI future" I want is thousands of non‑internet connected microchips doing one job each:

No grand consciousness. No robot overlords. Just orchestrating specialized systems working in harmony, lots of little optimizers quietly reducing waste.

This is also why I still respect RF tech. People keep predicting it'll vanish like fax machines. It won't. Radios are stubborn. They work in places Wi‑Fi doesn't. RF tech keeps proving people wrong.

My personal rule: RTFM, then build the smallest thing that works

I read documentation. That's basically a superpower in this field. That habit saved me with LLMs too. When I started working on crowdwave, ChatGPT 3.5 was barely out. Tiny context windows. A few thousand tokens. Hallucinations everywhere.

The first time it invented a fake fact, I remember thinking: What the hell? It just invented a completely fake fact.

The actual starting point for practical AI is accepting the limitations, designing around them, and stopping the model from playing wizard.

If you want the third AI future, you need clear problems, smaller models, and honest feedback.

And maybe fewer popups. We can dream.

-Sethers

Share this post
Newsletter

One email, once a week.

Notes on databases, systems, and the occasional strong opinion about Python. No spam, unsubscribe anytime.

Seth Black
Written by

Seth Black

Engineer and founder based in Texas. Writes about databases, AI, and running things in production. Embeds in small teams as lead engineer.

More from AI

View all →
AI

How to be a Good Farrier: What the Car Teaches Us About AI and Your Job

Nov 10, 2026
AI

Principle of Least Privilege in the AI Age

Apr 14, 2026
AI

How to Actually Secure MCP (Model Context Protocol)

Apr 14, 2026