AI Builders
Digest

Andrej Karpathy, known for training large deep neural nets and previously at Tesla and OpenAI, has joined Anthropic to focus on R&D at the frontier of LLMs. He remains passionate about education and plans to resume that work. This move highlights the intense competition among leading AI labs for top talent.

Google introduced Gemini Spark, a 24/7 personal AI agent for proactive task management. Project Genie received updates including Street View grounding, a new library, and external sharing. Gemini 3.5 Flash shows significant improvements in complex document tasks and agentic coding.

Box CEO Aaron Levie reports that token costs are a dominant concern for Fortune 500 CIOs. Strategies include workload prioritization, access tiers, spend caps, and long-term commitments. Sam Altman announced OpenAI's discounted tokens for 1-3 year commits to provide capacity certainty.

The Takeaway: Make building automation as simple as performing the manual task itself to drive real enterprise adoption. Serval founder and CEO Jake Stauch is building an AI-native platform for employee support that uses natural language to generate workflows and keep databases updated via a cogen engine. Key insight is separating admin agents (for building tools with permissions) from helpdesk agents to balance autonomy and control. Serval maintains strong unit economics by generating reusable code rather than reselling tokens. Customer obsession through daily Slack engagement with over 100 enterprises remains central.

Anthropic Engineering: Scaling Managed Agents by decoupling the brain from the hands. New self-hosted sandboxes and MCP tunnels keep execution within enterprise perimeters. This architecture virtualizes sessions, harnesses, and sandboxes for flexibility and security as models evolve. Supports providers like Cloudflare, Daytona, Modal, and Vercel.

Designer Ryo Lu uses Composer 2.5 for planning, building, and debugging, especially effective for UI work with Design Mode. Swyx outlines a 4-part AI software development lifecycle including tests, planning hot paths, breaking compatibility for refactors, and steering.

Guillermo Rauch announced Vercel CDN pricing that smooths traffic spikes. Claude Managed Agents integrate with Vercel Sandbox. Also highlighted rerun.io built with Svelte, Three.js, and Vercel.

Nikunj Kothari notes AI moving from assistant to coworker to autonomous workers, with labs collecting long-horizon task data. Jake Stauch highlights the gap between individual desire for autonomy and organizational need for control in enterprise settings.

Serval emphasizes 'fewer, better' hiring. Multiple founders discuss questioning roles, delaying hires, and using AI tools like Serval internally for recruiting and operations. Returns to talent density are higher than ever.

Google Labs updates on Project Genie and Flow. Cognition's Scott Wu featured in Claude series. Various builders experimenting with agents, sandboxes, and new models like Gemini 3.5 Flash.