AI Isn’t Magic. It’s a Data Decision.
Join the data collective.
Andrej Karpathy recently said something that I think really helps explain AI.
GPT-3 was terrible at chess. GPT-4 was great at it. The difference wasn’t some magical breakthrough. Someone at OpenAI decided to include more chess examples in the training data. That’s it. One person (maybe a team). One data decision.
That means every capability you see in an AI model, and every gap, comes down to what data someone chose to include. And right now, those choices are being made by a handful of people at a handful of companies, and they’re shaping how AI works for the entire world.
What This Looks Like in the Real World
I tested 8 of the top speech-to-text models against a subset of voice data. Word error rates ranged from 2.9% to 33.4% (lower is better). Most of them failed badly on women from New Orleans and let alone my southern voice.
That’s not a technology failure. That’s a data decision. Someone decided which voices to include in the training data, and Southern voices, especially women’s, weren’t a priority. The model works exactly as well as whoever chose the data decided it would work.
Now think about all the people building businesses on top of these models. Healthcare companies using voice AI for ambient scribes. Education platforms using speech recognition for language learning. Customer service tools. Accessibility tools. All of them dependent on data decisions made by someone they’ve never met, at a company that may not know their users exist.
If your use case was included in the training data, great, the model works for you and you can build on it. If it wasn’t, you either miss out or you fine-tune a model for your use case. Most people don’t have the resources to do that.
The Opportunity Nobody’s Talking About
But if AI is just data decisions, then whoever controls the data controls who AI works for.
Right now, that’s a handful of companies. It doesn’t have to stay that way.
Creators can pool their data together; with consent and build or fine-tune AI that actually works for their communities. Not hoping some engineer in San Francisco decided to include your accent, your genre, your art style, your voice. Actually building it yourself.
We asked 20,000 people to contribute their voice data to improve AI through Destined AI. Explicitly. No buried terms. No vague clauses. A clear ask. They said yes. Because when you’re honest about what you’re building and why, people want to be part of it.
The lesson from Karpathy’s chess example is simple: AI isn’t magic. It’s data. And your data has more power than you think. The question is whether you let a handful of companies decide what to do with it, or whether you decide for yourself.
“AI isn’t magic. It’s data. And your data has more power than you think.”