Classifiers used to be homework
We went from hand-tuning SVMs to asking frontier LLMs to pick a label. System One models like Jev keep zero-shot learning and bring back classifier speed and cost.
Agent design starts with the objective: lessons from a 16,000-word protocol prompt
Anthropic's protein-design system runs on a 16,000-word prompt, and only a third of it is about proteins. OpenAI's proposed Navier-Stokes proof now shows a similar process, alongside harder questions about validation and provenance.
Nobody voted on this: how to build a team that can still change its own mind
A team can get faster, flatter, and happier while quietly losing the ability to decide how it works. Notes on how to tell whether your team still authors its own arrangement, through Guattari's idea of transversality, and what to do about it.
Skin in the game: how to build vertical AI that survives
Foundation models are eating software from both ends. Seven observations on where the value actually lives in vertical AI, and whether AI-native services companies really are the future: why owning the outcome beats owning the workflow, why mandatory gates are a real moat and which five traditional software moats are already gone.
Work
- Company
- Stealth
- Role
- Let's chat!
- Date
- Company
- Ogment.ai
- Role
- Interim CTO
- Date
- Company
- Claryo
- Role
- Co-Founder, Chief Architect & VP of Engineering
- Date
- Company
- Robust.AI
- Role
- Founding Engineer
- Date
- Company
- Stanford
- Role
- Research Assistant
- Date
- Company
- CERN
- Role
- Research Assistant
- Date
- Company
- EPFL
- Role
- Student
- Date