TL;DR
Altman said AGI moved from labs to companies to people. Steinberger called autocomplete the real killer app. Rauch pegged AI at 42% of the web, while Dan Shipper said MCP became the new API layer—and Every warned it needed ruthless curation.
He says the big three wins are AGI speeding up research, companies, and eventually every person’s goals. He also points to a major math breakthrough as proof these models are already pushing into real scientific discovery, while OpenAI keeps leaning into startup adoption with $2M in credits for every YC company.
He’s all-in on @cotypist, calling it a must-have because it brings autocomplete everywhere. The take is simple: the best AI tools don’t feel like chatbots — they quietly speed up every place you type.
He says FDEs are here to stay because agents don’t just get deployed — they reshape workflows, and every model update can invalidate yesterday’s setup. That makes implementation a moving target, which is why vendors and partners who’ve done it hundreds of times will keep winning.
Project Genie is now fully available to Google AI Ultra subscribers globally, and the pitch is simple: pick characters, set the scene, and let the model build the game. Google Labs is pushing this as a real shift from playing games to designing them fast.
He says this move will put AI on 42% of the web, with support for every major model, provider, and modality — text, image, video, and audio. That’s a big distribution claim from Vercel’s CEO: the web stack is becoming the AI stack.
He says Exa is the only search stack he trusts for agents — used at YC and inside his own OpenClaw and Hermes agents because it’s fast, reliable, and complete. The bigger point: if your agent needs the web, search quality isn’t a nice-to-have, it’s the product.
She argues AI-native teams will flip old roles: ICs need to think like managers, delegating to agents and verifying output, while managers need to get more hands-on and build again. She also says the “T-shape” is becoming universal — go deeper in your domain, wider across adjacent skills, and layer AI fluency on top.
He says Sam Altman’s old idea of a business that improves with model quality is basically what he meant by “Agent Labs” — and he’s seeing a direct revenue lift as models improve, with a likely Q4 2025 step-change. He also notes a quick Exa bake-off where the team unanimously picked Exa in 1.5 hours, calling it a company on a generational tear.
Anthropic’s Stainless acquisition makes the bet obvious: the boring plumbing around APIs, SDKs, and MCP servers is now strategic. The real edge, Dan Shipper says from his AI & I convo with Stainless founder Alex Rattray, is keeping tools lean, using dynamic mode for messy APIs, and eventually letting models just write code against SDKs instead of juggling hundreds of tools.
MCP won’t scale by exposing everything; it needs ruthless curation.
The Takeaway: The winning MCP strategy isn’t “more tools”; it’s fewer, sharper tools designed for how models actually behave.
Alex Rattray, founder and CEO of Stainless, has spent years building the plumbing that lets computers talk to computers — first through APIs and SDKs for companies like Stripe, OpenAI, and Anthropic, and now through MCP servers. His view is blunt: the dream of agentic AI is real, but the current implementation is clumsy. If you hand an LLM every endpoint in a giant product like Stripe, “you’ve burned through your entire context budget” before the model even starts thinking.
His answer is not to expose more, but to expose better. That means fewer tools, precise names, tight schemas, and responses that return only what the model needs. It also means accepting that MCP is still a research problem, not a solved interface. Humans can learn Python; models can’t learn to “think like an LLM” from the outside, so the interface has to do the heavy lifting.
The most interesting part is how Alex uses AI himself: not as a chatbot, but as a research assistant that writes into a curated Git repo of notes, quotes, and citations. That way, future questions don’t require another expensive search through live systems. It’s a very Stainless answer to the AI era: don’t just automate the task — design the interface so the machine can actually use it.
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.