Supportiv just published
an article about their mental health language model.
I know it's incredible technology.
Because I built it from scratch.
The systems they're highlighting:
- Crisis prediction (self-harm, danger, suicidal ideation)
- Real-time matching that forms "emotionally resonant micro-communities"
- Teen detection and specialized routing
- Three-layer safety filter (regex + neural classifier + human review)
- Bias audits across gender, age, religion, socioeconomic factors
- MinHash near-duplicate detection
- Topic oversampling to fix underrepresented issues
235 million data points. 10+ real AI applications in production.
Head and shoulders ahead of competition. No one could come close.
Here's what's interesting:
A year and a half after I left as President, this is still their highlight.
Not new tech. Not new applications.
The same systems I built.
Meanwhile, AI moved faster in the last
18 months than it did in the previous 5 years combined.
GPT-4 to o1 to Gemini 2 Flash to Claude Sonnet 4. Multimodal. Reasoning models. 2M+ context windows.
The entire landscape shifted.
And they're still running what I built in 2024.
Here's what I learned:
Building something so good they can't move past it 18 months later? That's one kind of validation.
But in AI, standing still = falling behind.
Infrastructure that lasts is valuable.
Infrastructure that becomes a ceiling is dangerous.
The best systems I ever built weren't the ones that lasted longest.
They were the ones that made it easy to build what came next.
