AI Builders
Digest

The Takeaway: AI is still in its early technical phase; the real consumer-scale network-effect product has not yet arrived, and the next breakout will be a multiplayer experience with UGC communities and status incentives.

Benchmark partner Sarah (early at Pinterest) argues that consumer tech cycles always slide from deeply technical founders toward product geniuses. Google and ChatGPT were both “text box + backend magic” technical wins; Character.AI followed the same pattern. As infrastructure matures (DeepSeek and others lowering cost and capability barriers), product intuition and social DNA become decisive. She calls Custom GPTs “criminal” — the team is extraordinarily capable yet completely non-social: no trust-building visibility, no real following of creators, no status-seeking loop. The real opportunity is making it trivial for ordinary people to copy power-user prompts and personalized setups, creating a Reddit-style prompt community that is native to AI. She is also bullish on stablecoins for removing real friction in high-inflation countries like Argentina, and she is drawn to founders for whom the work is an affliction rather than a cool job.

Box CEO Aaron Levie argues that 99% of the world’s tokens will be consumed in enterprise contexts: writing code, combing through life-sciences research, automating manufacturing, securing enterprises, detecting fraud, and thousands of other high-economic-value tasks. Even agents that most impact consumers will often be wrapped as end-to-end services where the user never thinks about an AI system on the other side. The key point is that diffusion will take years because workflows must be re-engineered to incorporate AI — a complete change to how most work happens today. Anyone expecting this overnight should update their timelines.

Meta Senior Director of AI Madhu Guru (previously led Gemini and Veo at Google) says AI diffusion has been slow because we force users to understand our nerdy lab speak: a blank window, write a prompt, pick a model, decide whether they need an agent, and somehow know what context windows, reasoning, MCP, memory and skills are. Most people do not care. They need a thing done. Give them something that does that and they are happy. He expects that breakthrough product within the next 12 months.

Practical AI tutorial creator Peter Yang released a new /human-review skill. After installing, type “/human-review (doc name)” and a visual editor opens so you can edit HTML/Markdown directly, resize images, and leave feedback like Google Doc comments. Click “Send to agent” and the agent applies every change. Built for the final 10% polish on PRDs, landing pages and similar work.

AI engineer and podcaster Swyx notes that a very primitive form of the near-term multi-agent AGI future is setting up one thread to ping back once finished, creating an implicit kanban/waterfall graph of dependent threads while each agent preserves its own work. He cannot wait for a proper UI, but the pattern can already be hacked together in most coding agents.

Vercel CEO Guillermo Rauch announced infinite agent compute: 10,000 concurrent + 5,000 CPU cores per minute, with quotas that can still be raised. He also quipped that writing a banger tweet is AGI-complete.

OpenClaw-related developer Peter Steinberger gave Codex a video-enabled remote KVM so it can automate end-to-end testing of the iMessage integration on OpenClaw (iMessage is unreliable in VMs and certain features require SIP disabled).

OpenAI Codex & ChatGPT team member Thibault Sottiaux revealed he receives on average one DM or email every six minutes asking for a reset. He occasionally obliges when the request comes with solid feedback or good banter. He also recommended founders explore the powerful /goal loop inside Codex.

Google Labs announced Dreambeans is expanding: in addition to US AI Ultra users, AI Pro subscribers in the US can now receive a fresh daily collection of personalized stories that surface the deep dives and hidden gems they actually want to read.

Y Combinator President & CEO Garry Tan said he cannot wait for AI to get so good that detecting whether something is AI-generated no longer matters. Silverware used to be handmade, but nobody complains that their dinner fork is stamped by a machine. The quality of ideas matters; the important thing is people can eat.

Every CEO Dan Shipper noted that to stay competitive today Google needs to catch up on frontier coding, yet Demis believes different fundamental research directions (such as world models) matter more to his long-term goal even if they are less important competitively right now.

Linear Head of Product Nan Yu asked the blunt question: how is ChatGPT *not* an agent? The post sparked discussion about where the boundary of “agent” should be drawn.

FPV Ventures partner Nikunj Kothari listed the words AI tech people will use a lot more in the next 6-9 months: out of distribution, control plane, unverifiable fields, rails, intelligence per watt, cope, angst. Some are already in rotation; frequency will rise.

Replit CEO Amjad Masad remains struck by the amount of founder talent packed into such a small space every time he visits SPC.

Swyx observed that Paul Erdős prompted fellow mathematicians with bribes in a manner remarkably similar to how people prompted early LLMs.