Host: If you follow AI at all right now, one question keeps coming up: why aren't open models running your daily life in myOrbit yet? They're free to download, and honestly, some of them are extraordinary.

Listener: Right, that's what I don't get. You see benchmarks every week showing these open models practically matching the big proprietary giants. If they're that good and free, why wouldn't you switch tomorrow?

Host: It's a completely fair challenge. Stanford's AI Index found the top closed model leads the best open model by only three point three percent. And Epoch AI calculates open models lag behind frontier closed ones by just four months on average.

Listener: Four months is nothing in tech! So what's the hold-up? Are you worried about where they come from?

Host: Not at all. Excellent models come from everywhere today—Meta in California, Mistral in France, Alibaba's Qwen and DeepSeek in China. It has nothing to do with geography. The issue is trust. When a model reads your private messages, drafts your emails, or plans your week, you have to know what it will actually do.

Listener: Wait, but isn't 'open' supposed to mean transparent? Like open-source software where anyone can inspect the code?

Host: That's the big misconception. Most so-called open models are actually open-weight, not open source. You get the final file of billions of numbers that the model learned. But you don't get the training code or the dataset. Projects like the Allen Institute's OLMo share the full recipe, but they're rare exceptions.

Listener: Okay, so you have a giant file of numbers. Why can't you just scan those numbers to see if it's safe?

Host: Because weights aren't human-readable instructions. And research shows you can plant deliberate backdoors inside them that survive normal safety training. In one recent test, a researcher spent under a hundred dollars with just ten poisoned examples to implant a trigger that made a model reliably output insecure code.

Listener: Under a hundred dollars? That's wild. But what if you just run the model on Amazon, Google, or Microsoft? Don't their clouds keep everything isolated so your data can't leak out?

Host: The cloud providers do great work on isolation. When an open model runs in Amazon Bedrock or Google's Vertex, your prompts aren't phoned home to the model maker. But hosting it in a secure building doesn't tell you who you let through the door. As Microsoft plainly notes in their documentation, they don't scan all catalog models for vulnerabilities. Cloud hosting protects your data from leaking out; it doesn't verify the model's behavior inside.

Listener: Fair point. But let's be real—you can't inspect the inside of Anthropic's Claude or OpenAI's GPT models either. Isn't that a double standard?

Host: It's the right push. We can't see their weights either. The difference isn't inspection; it's direct contractual accountability and exhaustive evaluation. We have direct enterprise agreements with them, clear legal terms, and thousands of hours evaluating their specific behavior on our workflows. A downloaded file off the internet doesn't come with that accountability.

Listener: So what would it take for an open model to actually make it into myOrbit?

Host: Three things. First, genuine transparency—like NVIDIA releasing training data alongside Nemotron. Second, better technical tools for detecting sleeper backdoors. And third, our own rigorous evaluation, probably starting with narrow, specific tasks before anything touching personal conversations.

Listener: And you'd never train a model on our private conversations to customize it?

Host: Never. Research shows fine-tuning on sensitive user data dramatically spikes leakage risks—sometimes up to seventy-five percent. When your personal twin needs information, it looks it up in real time; it's never baked into model weights.

Listener: Got it. So it's not 'never open models'—it's 'not until we can personally verify what's running.'

Host: Exactly. Open models are vital, and when the verification catches up to the capability, we'll gladly run them. If you want to see the full citations and security breakdown, check out the complete note on our site.