AI Builders
Digest

The Takeaway: The real bottleneck is not model capability but the protocols, memory, and ecosystem that let agents take real actions in the world—the same role HTTP played for the internet.

Microsoft CTO Kevin Scott (long-time programmer and woodworker) notes that last year’s scaling-law doubts have given way to the agentic web. Reasoning has pulled ahead of product delivery, creating a capability overhang. Current agents are transactional and memory-constrained, limiting complex delegation. True usefulness requires tool use, system changes, and rich information sources—an internet-like open ecosystem. MCP is the HTTP of this world; NL Web the HTML. Simple, composable open protocols are emerging. Inside Microsoft he pushes every internal system to speak a standard agent protocol so the company does not ship its org chart. On security he favors open foundations plus agent identities, entitlement queries, and explicit user permission rather than closed stacks. On the craft of coding he compares the old hand-tools-vs-power-tools (and now CNC) debates in woodworking: passionate makers will always have strong opinions, but stay curious—try new tools and keep what works. The future is many specialized product-focused agents, not one ruler. A year from now, people still waiting for “readiness” will be behind; asynchronous agents that run long autonomous loops before reporting back will be normal.

“Be curious. Try. And if it works for you, use it. And if it doesn’t, don’t.”

Builder Zara Zhang notes that talented individuals can achieve 10x their potential with AI when working on their own thing, yet the same person inside a large organization at best gains 20% (sometimes less). This is why more high-caliber people are leaving big companies—except top AI labs like OpenAI and Anthropic. She also observes that everyone who is ahead in using AI still feels behind.

Box CEO Aaron Levie argues that AI diffusion is far more rate-limited by good evals than most realize. Public model-release benchmarks show the shape of general progress, but the much larger opportunity is evals for the actual workflows enterprises run—down to the specifics of an individual company. You cannot automate what you cannot measure; enterprises will not proceed on vibes alone.

OpenAI Codex & ChatGPT’s Thibault Sottiaux reports progress on rate limits: inefficiencies with images in long multi-compaction sessions, high p95+ usage for Computer History, and a conversation-title feature that was over-consuming. A tiger team ships fixes tomorrow and will fully reset usage for all paid subscriptions. Separately they found a novel efficiency approach unrelated to these issues and will work on it next week.

Meta AI Sr Director Madhu Guru (ex-Google Gemini/Veo) continues his series: hill-climbing on evals simply means picking a dimension that matters and optimizing hard—better quality on high-value journeys, adjacent use cases, lower cost or latency. The work is better harnesses and model selection via prompt/context eng, memory, post-training, or old-school deterministic code. Your failure-mode taxonomy is the compass. Example: if tool-calling is the top failure, you may be stuffing 20 tools when each task needs only 3-5—context-engineer the right tools at the right stage and iterate. For cost, launch with the best model first to nail quality users love, then hill-climb to a smaller/faster model. The essential is evals that actually tell you whether you are moving the right way.

Practical AI educator Peter Yang shares a privacy workflow: go to myaccount.google.com/permissions, run a prompt in Codex or Claude Code telling it there is an open browser tab, then have the AI select and disconnect the apps that should no longer hold your Google data. He just cut half of them. He also credits the Instinct team for quickly shipping the ability to delete external data (including Gmail records) under Data Privacy. Separately he is building a new AI skill called /fuck-cancer to help patients and families navigate the process and stay informed—seeking input on what to include.

Replit CEO Amjad Masad doubles down on shipping cadence: “A week has 7 days. That means 7 ships.” Another post notes that what was once called “pretty soon” turned out to be three months.

Every CEO Dan Shipper posts that the company is hiring and quote-tweets support for “agent native ftw,” reinforcing the bet on products and workflows built natively around agents.

Vercel CEO Guillermo Rauch reflects that AI can do wonderful things and grant many wishes, yet we have only one Earth, one California, one Patagonia, one Mendoza. The US and Argentina are among the freest countries that also possess the most glorious geography—he is very long 🇺🇸🇦🇷. He also calls San Francisco a paradise on Earth.

FPV Ventures partner Nikunj Kothari calls out young founders who ragebait investors and then send SAFE docs expecting a wire: that is not how investing works. He stresses how critical it is to seek advice from the right people and is angered watching talented young people with real substance play these games. He also notes that LinkedIn has sunk to unimaginable levels of slop.