I founded Seed Club, a venture network that turns the conviction of the people founders most want on their cap table into coordinated capital. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
Week of September 14, 2026
Evals become the strategic moat for AI app companies
The emerging playbook for AI application companies is to build public evaluations first, use them to post-train a model for cost reduction, then train custom models for customers. Evals are repositioning from a quality metric to a durable competitive asset.
- Phase progression: public evals, then post-training, then customer-specific model training
- Eval literacy is now considered basic competency for AI practitioners
- Fine-tuning for cost savings depends on having high-quality eval infrastructure first
AI agents move from workflow tools to autonomous operators
A new generation of AI applications is designed not to assist with workflows but to execute them entirely, from reconciling purchase orders in procurement to running vending machines and cafes as autonomous entities. The operator layer is becoming the product.
- Autonomous agents must sit inside existing communication flows to replace human work
- The shift is from systems of record to systems of action
- Fully autonomous operation raises unresolved questions about accountability and error recovery
Domain experts challenge AI capability and safety narratives
A senior OpenAI capabilities researcher published a personal statement on AI risk, signaling internal concern not confined to outside critics. Separately, a synthetic biology expert challenged the rigor of AI-driven biological claims, citing a gap between hands-on domain knowledge and AI-assisted speculation.
- Insider safety concern at a frontier lab is distinct from external critique
- Domain expertise in biology reveals limits that AI capability claims tend to obscure
Consumer AI stagnates while enterprise inference ROI soars
AI has not improved consumer services like ride-sharing or food delivery, yet LLM inference at one firm now triples total salary costs and is described as additive to headcount rather than a substitute. The consumer-enterprise divide in AI value creation is widening.
- Consumer disappointment may reflect platform incentive problems, not AI limitations
- Inference ROI claims currently rest on individual firm data, not broad trends
- Whether inference displaces or augments headcount remains an open question