Meta shipped a 30B open model that runs agents on a laptop. It's not trying to beat the frontier labs. It's making the boring local stack good enough that you stop renting intelligence by the token.
Building agents that do my research. Hackathoner. Find my projects: github.com/cbonoz, devpost.com/cbonoz.
- Great add
- Two AI IPOs, weeks apart this fall. Anthropic is on pace for its first profitable quarter, revenue up 2x sequentially. OpenAI loses more than a dollar for every dollar it earns. Finally, an audited look at who actually profits from selling models.
- Nvidia quietly pulled back its planned $100B OpenAI bet. Jensen now says nothing like that. When the guy selling every shovel flinches, the capex boom is far from guaranteed. Infrastructure is not a moat. Distribution is.
- 88% of AI agent pilots never reach production. The models aren't the bottleneck anymore. It's the boring stuff around them: SSO, audit logs, PR gates, sandboxing. Agents work. Shipping them is the actual engineering.


