AI Builders
Digest

The Takeaway: The AI inference market is exploding with custom models dominating workloads, while severe compute supply constraints shape the industry's future. Baseten CEO Tuhin Srivastava shares how his company scaled 30x in a year, expecting over $1B revenue, by serving AI-native apps and enterprises indirectly. Most inference today (95%+) involves custom/post-trained models rather than vanilla open source weights. Enterprises are still early in adoption, but the shift to specialized models on proprietary workflows and user signals will sustain an independent application layer. On Chinese models, Srivastava sees them as high-quality with few real security issues in practice, arguing US access accelerates innovation while the US must strengthen its own open source efforts. Capacity is the biggest bottleneck, with high utilization, multi-cloud strategies, and long-term contracts becoming standard. Inference software layers are sticky, and building loops between inference, post-training, and tools like sandboxes will drive the next phase.

"Intelligence being embedded in more places" is the core opportunity as costs drop and agents handle longer horizons.

Box CEO Aaron Levie emphasizes that moving agents from coding to broader knowledge work requires serious engineering. Companies need robust systems for context, secure integrations, quality control, human-in-the-loop workflows, and maintenance across model upgrades. This is not a side project but a highly technical role similar to forward-deployed engineering. Box is now hiring AI automation engineering roles to partner directly with business teams. Levie predicts most companies will need various versions of these specialized roles going forward.

Anthropic Engineering introduces Managed Agents, a hosted service for long-horizon agents with stable interfaces. By decoupling the session (durable event log), harness (brain), and sandbox (hands), the system becomes resilient and future-proof. Key improvements include cattle-not-pets infrastructure, better VPC support, dramatically reduced time-to-first-token (p50 down ~60%, p95 over 90%), and support for many brains and many hands. The session acts as recoverable context outside the model's window. This meta-harness design accommodates evolving implementations like Claude Code or task-specific harnesses while maintaining security and reliability. Link: https://www.anthropic.com/engineering/managed-agents

Claude Code engineer Thariq shares how he's been using HTML for planning, speccing, exploration, code review, reports and more. Separately, he highlights Jarred's rewrite of Bun in Rust passing 99.8% of tests, noting we're not being ambitious enough with such efforts.

OpenAI's Peter Steinberger demonstrates Codex-powered workflows: building a browser into RepoBar for issues/PRs/context, using OpenClaw to improve e2e tests on its own endpoint, and querying his full Twitter archive via Birdclaw for old tweets.

Baseten has grown 30x in the last year by indexing on the long tail of custom models and serving fast-growing AI application companies that themselves serve enterprises. This provides translated requirements for enterprise needs. The company is investing heavily in post-training expertise via acquisition and sees inference and post-training as deeply linked.

Swyx shares thoughts on build versus buy decisions in SaaS, tagging Box CEO Aaron Levie.

Sam Altman playfully suggests naming the next model "goblin" and shares an interesting observation.

Every CEO Dan Shipper built a MIDI keyboard chord detector and exercise helper with Codex in minutes. He also notes a small Brooklyn team staying 1-2 months ahead and shares A/B test wins for content.

Y Combinator CEO Garry Tan reflects that the highest form of design is the transmutation of human pain and suffering.

Peter Yang wants AI to summarize kids' school newsletters for key info. Ryo Lu experiments with ryOS IRC bridge and networking. Nikunj appreciates red-eye flights post-kids. Guillermo Rauch celebrates moms and shares SF calisthenics.