It's the end of May, and I'm coming back from a podcasting conference in the UK. I'm feeling so alive, but now I have a problem on my hands. And it's a very embarrassing problem.

Now I can't use AI, and I'm the AI Guy.

I was already tormented well before the conference. I was setting up admin systems for a friend and her $2M not-for-profit. I couldn't use AI. Yes, there was red tape around public perception, but the real blocker in her office is privacy, specifically the perceived and real liability around it. It's the top reason 80% of Canadian businesses haven't touched this stuff in production. Cost is the second.

Before diving into the experiments and the results, it pays to understand the landscape.

I needed to find alternatives to Claude and ChatGPT. Being a child of the 90s, I still remember when we used to own software. I wondered if we could do the same with LLMs.

That search led me to open-weight models. Kimi 2.6 was the first thread of hope (Kimi 3 has come since then, and it's even better!). I was looking up benchmarks at the airport, hoping to find out that the open-weight models have caught up with the frontier ones. The answer: not quite, but they are close enough for most business use.

I explored three paths from there.

What I quickly realized is that the hardware required to run Kimi 2.6, the 1.1 trillion parameter model, requires a datacenter (that's a minimum of 8 x A100 chips for the lowest precision setup). Sure, it's a step in the right direction because it opens up the playing field for smaller inference providers, thus decreasing the reliance on a single company. But ultimately, the point is to use existing hardware instead of stamping out more GPUs (for many reasons, including that it still takes child labour in mines to extract the raw materials). And at the very least, businesses should be able to set up their sovereign, on-prem infrastructure for an affordable cost for what will be (and already are) their critical systems.

The second ray of sunshine was distributed inference, and not the kind like Akash.network. Creating a distributed marketplace of mid- to high-powered GPUs is an effort that I fully support, which is what Akash is doing. Projects like exo and prima.cpp allow you to do something even cooler: run large models distributed over low-powered devices connected by wifi (or cables). It's really magic, but it can hardly support real-time use, and the libraries are more experimental than battle-tested.

So that left me with the last possible solution that could satisfy both the ethical constraints and the ironclad privacy requirement: mobile-first SLMs.

The challenge: they're "dumb" (and yet 12 years ago I would've thought of their capabilities as pure magic).

The beauty is that they're egalitarian: anyone with a phone can run them privately and offline.

Models in the 9B - 12B parameter range (Qwen, Gemma) are actually quite capable. But they won’t run on old hardware, like a 4GB VRAM AMD card I had in an aging desktop (while VRAM is similar to regular RAM, the distinction is critical). What I wanted to see is how far I could push a truly tiny model, so I chose LFM 2.5 1.2B for my experimentation.

I wanted to make it challenging. The point was to find out if it had the capacity to navigate and understand an unpredictable environment. The World Wide Web is one such place.

And the approach would rely on the single advantage of running a tiny model on under-utilized hardware: tokens are essentially free and you can forget about the token burn.

Can a 1.2B model detect a business address in a block of text? Of course. But given a random website, can it navigate around intelligently until it finds an address? I set out to find out with a browser automation MCP connected to a mini-harness.

Given a random website, the harness explored it using the following flow, which was a mix of prompts and MCP calls chained in a sequence:

Step 1: Read the page text. The model is asked if there's a business address, and to respond with it, and if not, go to Step 2.
Step 2: MCP generates a list of interactive elements visible on the page (buttons, links, etc).
Step 3: For each element from Step 2, the model decides if interacting with this element might bring up a page or an element with the business address.
Step 4: The model is asked to make a tool call to interact with the element.
Step 5: Go to Step 1.

The problems showed up immediately:

Ultimately, this tiny 1.2B parameter model is able to do this simple navigational task reliably. It just needs a lot of hand holding. Instead of giving it a list to pick from in Step 3, I run the same prompt over each page element in a programmatic loop one at a time, preventing the model from getting confused. Tool calls had to be very explicit. When depending on a specific output format, adding a follow-up prompt that verifies and corrects to the expected format is essential. And of course, it needs grace: when it fails, going a step back and trying again.

Can it navigate the internet with a more open-ended task, like finding product listings and verifying details or doing price comparison with other sources? How reliable can it be? That requires a lot more skill, like finding and using pagination, dealing with iframes and modals, and understanding that actions change the page state.

Adding memory and the ability to create reusable code snippets (akin to Voyager by NVIDIA) is showing promise.

In some ways it's silly to try to get a 1.2B model to do something complex. But this approach of handholding, as if you're teaching your grandma to use the internet, works (Gen Z doesn't have that problem anymore 😀). And it can be exponentially powerful when combined with 4B - 30B class models running on your own, affordable hardware.

I created a library so that it would be simple for you to try this approach yourself:

The Podcast Show in the UK was a reminder of just how awesome LLMs are. They're no longer strictly secondary to human creators (check out the Epstein Files podcast going viral). But it was also a reminder of just how much hatred there is towards AI. That hatred has a basis: from stolen data to hardware that still requires children working in mines to extract raw materials.

The juxtaposition jolted me to realize my own hypocrisy. You make your own judgments, but for my own integrity I couldn't continue paying Anthropic and OpenAI while hosting the "Apps for Matriarchy" meetups. That was the embarrassing realization, and I had to find alternatives.

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.