Can't See Inside AI Models? This Teardown Workflow Changed That

I Got Confused by AI Too

Late last year, Xiaolin (a fellow solopreneur I know in Hangzhou) sat across from me in a café and said: "I'm using seven or eight AI tools now, and I have absolutely no idea which ones are actually reliable. Every time I pick one, it feels like gambling." I nod ded — because I was in exactly the same place. AI assist ants, writing tools, customer service bots — every landing page sounds incredible, but you have no idea what logic is actually running underneath.

This isn't a post telling you to go learn programming. It's me sharing a mindset shift: AI tools aren't total black boxes. Some parts of them can be seen through — even if you're not technical.

What This Workflow Is, and Who's Using It

Sebastian Raschka is an AI researcher who has spent years doing one specific thing: turning complex AI model architectures into diagrams that non-specialists can actually follow . His method isn't mysterious once you hear it.

Step one: find the official documentation (think of it like a product spec sheet). Step two: if the docs are vague or incomplete — and he says most are these days — go directly to the model's "config file." That's essentially the factory settings sheet for an AI tool: it lists the parameters, structure, and how it actually runs. Step three: cross-reference with run nable code to verify. His core belief: "Code that runs doesn't lie."

This workflow is built around "open-source models" — AI systems that make their internal architecture public, like Meta's LLaMA or Mistral. Closed commercial products like ChatGPT or Claude don't expose their internals, so this approach doesn't apply to those.

For those of us who aren't engineers, the value here isn't " I'm going to tear down models myself." It's a judgment framework: an AI tool that's willing to make its structure public is generally more trust worthy than one that's completely opaque.

How Much of This Can You Repl icate Today

Cost: $0. The resources Raschka references — the Hugging Face model hub, published papers — are all free.

Time: If your only goal is "figure out whether this AI tool is open-source and whether its structure is transparent," a search plus some reading takes about 30 minutes.

Technical barrier: No coding required. You do need to be comfortable reading English-language pages and willing to click around on Hugging Face (it's basically an app store for AI models — there's a Chinese interface option too).

First step: Go to huggingface.co and search the name of an AI tool you're currently using. See if it shows up. See if there's a public config file. If yes , that's a point in its favor. If not, don't panic — it's just one data point.

Advice by Stage

If you're just starting out with AI tools and still evaluating options: Honestly, I'd say ignore all of this for now. Getting fluent with one tool you already have beats everything else at this stage. Bookmark this piece and come back in six months.

If you already have one or two paying clients and you're starting to use AI for client- facing work: I'd suggest spending one 30-minute session searching your core AI tools to see if they're open-source or have public documentation. Not for technical reasons — but so that when a client asks "is this AI safe to use?" you have an actual answer you can articulate.

If you're leading a small team and starting to embed AI into your business workflows: This teardown mindset is worth investing time in seriously. You don't need to do the teardowns yourself — but being able to hold a real conversation with technical people and understand basic terminology will give you a lot more confidence when evaluating tools or negotiating partnerships .

This workflow isn't something everyone needs right now. If you're still in exploration mode, it's completely fine to set it aside. I got stuck at that same stage longer than I'd like to admit.