The Long Backstory

I’ve been making scones lately, which means I’ve been searching the internet for recipes.

That also means I’ve been reminded why most websites are so painful to use.

You land on a recipe page and immediately get attacked by ads, pop-ups, newsletter boxes, autoplay videos, and cookie banners. If you can’t find the “Jump to Recipe” button, you’re forced to scroll through the creator’s life story before you get to the ingredients.

You know the one.

Their great-grandmother’s heirloom chocolate scone recipe. Made with rainwater, stoneground Einkorn flour, and ancient baking wisdom passed down since 2,000 B.C.

It’s annoying, but I try not to get too frustrated. I’m getting the recipe for free and they’re trying to make a living.

The problem is how they make that living.

For years, creators depended on Google to send traffic to their websites. A food blog with 500,000 monthly visitors could support a business with fairly normal sidebar ads.

Then Google changed their search engine model.

Between search updates and AI-generated answers, some sites that once received 500,000 monthly visitors now get a fraction of that traffic. If your revenue depends on ad impressions, an eighty-percent decline in visitors creates a brutal math problem.

You either need more visitors or more ads per visitor.

That is how we ended up with recipe pages that feel like digital obstacle courses.

This is a classic example of platform dependency.

Since the beginning of the internet, businesses have built on platforms they do not control. Search engines. Social networks. App stores. Marketplaces. Payment networks. Each one offers distribution, convenience, and leverage.

Each one can also change the rules without a moment’s notice, disrupting or even killing businesses overnight.

The latest version of this risk is AI.

Fable 5

Anthropic released Fable 5 last week, its newest high-capability model. It was built for longer, harder, more autonomous work such as coding, research, analysis and agentic tasks.

In other words, the kind of model people start building substantial business workflows around.

The initial excitement around Fable 5’s capabilities quickly dampened.

Users discovered in some cases, Anthropic could route sensitive requests to a different, lower level model. In others, safeguards could quietly modify prompts, steer outputs, or reduce the model’s effectiveness without clearly telling the user.

Anthropic claimed that these measures were for safety reasons. Whether or not they were is irrelevant. These kinds of guardrails shouldn’t be hidden or buried in vague language in the terms of service.

If the model silently changes the request or degrades the answer, you may not know what you are actually using and you should.

That is the platform dependency problem.

When the intelligence layer lives on someone else’s servers, using model architecture and routing that is invisible, your access is conditional. The model, the rules, the output and your access can change without you knowing. Even as your workflow may depend on all of it staying the same.

Even worse, these issues and others related to the supposed safety vulnerability resulted in the U.S. government directing Anthropic to remove access to Fable 5.

Intelligence Independence

The AI space is evolving rapidly with new models, tools and platforms arriving daily. No matter how you’re using these tools, you should be evaluating how dependent you are on any model or platform.

It’s not just about access. Platforms can change price, speed, output quality, etc. Not thinking about mitigating these risks as you build AI into your business is a mistake.

The following are the three ways I’m thinking about mitigating the risk across my businesses.

BYOM

Whatever you build, don’t do it in a way that you’re locked in to a particular model. Bring Your Own Model (BYOM). Fine if you want to use a closed source model but be able to switch it out (and understand the implications of doing so).

For example, Anthropic has built incredible tools around their models such as Claude Cowork and Claude Code. Many businesses have built workflows with these tools. I would instead use agentic platforms or tools that are model agnostic and provide the ability to assign the model of my choice.

For example, I use Paperclip for my most important agentic workflows. I can swap out models in a few clicks.

The downside of this approach is that you increasingly cannot use base subscriptions from Anthropic or other providers and must use API (Application Program Interface) access which, depending on usage, can incur higher costs.

It’s well worth the additional spend in my opinion.

Open Source

Open source models are AI models where the code, training data, weights and everything associated with the model are publicly available. The obvious benefit is there are no surprises as to models’ functionality.

However, there are significant trade offs to consider when using open source models. Open source models are often not as capable as the latest frontier models for certain tasks.

For most business use cases, open source models perform just as well as the latest frontier models. Even if you can’t remove all of your agents from frontier models, there are certainly parts of your workflow that can utilize open source models.

Also, the source of the model matters. Many of the best open source models are Chinese which present security issues, privacy and data exfiltration concerns.

There are many good non-Chinese open source models, however, including Google’s Gemma 4.

Local

The third step toward intelligence independence is local AI.

Running a model locally sounds more intimidating than it is.

At the simplest level, it means the model runs on your own machine or a rented cloud instance instead of a model provider’s data center. For many people, this can now be done with consumer tools that let you download a model and chat with it on your laptop.

Local models are useful for a few specific reasons.

The first is privacy.

If you are working with sensitive documents, client data, financials, customer records, legal materials, or internal strategy, local models reduce the amount of information you send to third-party providers.

The second is continuity.

If a cloud model is unavailable, throttled, restricted, or too expensive, a local model can keep certain workflows running.

Third, cost.

For repeated, narrow, high-volume tasks, local inference can be cheaper once the system is set up.

Finally, experimentation.

Local models let you test, customize, and iterate without worrying about usage limits, policy shifts, or surprise bills.

But hosting models locally has its limits.

They require hardware, setup and may be impractical for non-technical users beyond basic experimentation.

Use local models when privacy, resilience, cost, or independence matter.

The Next Step

We’ve “seen this movie” before as platform providers such as major search engines, social media companies and payment processors disrupted businesses by suddenly changing access rules.

AI models will not exist in isolated parts of the business. It’s bad enough to lose your audience, but if AI is implemented throughout your operation, every part of your business is at risk.

As you integrate AI in your business, consider what would happen if you lost access to a model or a platform unexpectedly? Seriously evaluate data privacy and cost.

Where possible, build model optionality into your systems. Experiment with open source and locally hosted models. The time for evaluation is now, not after your business have been compromised by your AI model provider.

My goal with The Leap is to provide you each Saturday with the knowledge, tools and lessons learned to help you get started and keep going toward building your future. 

Whether you are making the leap to startups, solo-entrepreneurship, freelancing, side hustles or other creative ventures, the tools and strategies to succeed in each are similar.