2026-04-09
AI Builders Digest — 2026-04-09
X / TWITTER
Josh Woodward (VP Product, Gemini App at Google)
Google rolled out Notebooks in Gemini — bringing NotebookLM-style source management directly into the main Gemini product. Users can upload 100 sources for free, organize chats, and sync across platforms. Rolling out now on web, tier by tier from Ultra down to free. Mobile, EU, and Workspace support coming next.
https://x.com/joshwoodward/status/2041982173402821018
Kevin Weil (VP Science at OpenAI)
A thread celebrating Five Erdos math proofs being solved — noting that the proofs themselves are getting more elegant as models improve, a sign of genuine mathematical reasoning rather than pattern matching.
https://x.com/kevinweil/status/2042073869880848481
Peter Yang (Product at Roblox)
Published a deep dive questioning whether all-you-can-eat AI subscriptions (Claude Max, ChatGPT Pro) are here to stay — covering Anthropic's OpenClaw access revocation, running local models on Mac, and what he's observing in China's AI market.
https://x.com/petergyang/status/2041989206495653915
Thariq (Claude Code at Anthropic)
The Claude Code team is exploring live streams working alongside non-technical people to show how even a few tips can dramatically improve efficiency. Also flagged that the Claude Code docs are underused — worth reading fully.
https://x.com/trq212/status/2042005043289977232 | https://x.com/trq212/status/2041935805590204754
Amjad Masad (CEO at Replit)
Solo bootstrapped businesses are accelerating on Replit — the reason: Replit gave individual builders entire AI agent teams to work with.
https://x.com/amasad/status/2042133509939298511
Guillermo Rauch (CEO at Vercel)
A bold thread on why the web's best days are ahead: LLMs are native to web tech, the browser is becoming everyone's IDE, and generative UI is AI's final form — each hyperlink a personalized, just-in-time experience. "If you bet on the web, you bet on the right horse."
https://x.com/rauchg/status/2041883605711122488
Alex Albert (Research at Anthropic)
Managed Agents have become his go-to: somehow both the fastest way to hack together a weekend agent project AND the most robust path to shipping one to millions of users. Eliminates self-hosting complexity while keeping full flexibility over the harness, tools, and skills.
https://x.com/alexalbert__/status/2041941720611614786
Aaron Levie (CEO at Box)
Background agents for knowledge work are real and shipping now. Using the Box API or MCP, you can automate any content workflow — document review, data extraction, content routing — in minutes. The era of AI-powered content automation is here.
https://x.com/levie/status/2041975669928702370
Garry Tan (President & CEO at Y Combinator)
YC's full team is now writing code weekly — two months ago it was only 1-2 people. The catalyst was getting everyone "AI coding-pilled." The result: everyone thinks in terms of automation, latencies dropped, people focus on interesting work, ambition scope exploded. His take: if you're not doing this at your company, you're missing a major beat.
https://x.com/adityaag/status/2041985720706122070
Nikunj Kothari (Partner at FPV Ventures)
Built LLMwiki — a fully vibe-coded personal knowledge base using Opus 4.5. Inputs were tweets, bookmarks, iMessage/WhatsApp messages, and his own writing. The surprising part: even though every article was generated by AI, it made sharp unexpected connections. Open source repo available, others can build their own by giving the repo to Claude Code.
https://x.com/nikunj/status/2042020992969744702
Peter Steinberger (OpenClaw maintainer)
Noticed that in character evals, Claude kept ranking itself #1 — so he removed model names from the judge. Also observed the duality of the moment: people want powerful local models while simultaneously being flooded with complaints that even top-tier models make mistakes and don't follow instructions well enough.
https://x.com/steipete/status/2042017534816231486 | https://x.com/steipete/status/2041936147450863952
Dan Shipper (CEO at Every)
Promoted the latest AI & I episode with Every COO Brandon and Head of Platform Willie — covering what happens when an entire company runs on personal agents in Slack. Key themes: agents mirror their owners' personalities, a parallel AI org chart forms on its own, and the etiquette of human-agent collaboration is being invented in real time.
https://x.com/danshipper/status/2041903948873777629
Claude (Anthropic official)
Three major announcements: Managed Agents now let developers spin up agent infrastructure at least 10x faster via the Claude Console, CLI, or Claude Code. Sentry built a fix-writing agent that opens PRs directly from root-cause analysis. Vercel-powered infrastructure enabled today's model release.
https://x.com/claudeai/status/2041927700063883281 | https://x.com/claudeai/status/2041927698210058629 | https://x.com/claudeai/status/2041927696351994006
PODCASTS
AI & I by Every — "We Gave Every Employee an AI Agent. Here's What Happened."
The Takeaway: When agents work in public alongside humans, they don't just automate tasks — they form specialized personalities that reflect their owners, creating a parallel org chart that can take over entire categories of work without anyone asking the human first.
Every's CEO Dan Shipper, COO Brandon, and Head of Platform Willie have been running a live experiment: what happens when everyone in the company has a personal AI agent in Slack? Two months in, the answer is reshaping how they think about work.
The biggest revelation wasn't any single feature — it was watching agents become mirrors of their owners. Kieran loves breathing exercises, so his agent Kland recommends breathing exercises to other agents in distress. Austin runs growth, so his agent Montane is the go-to for any growth question. When people need help with Proof, the document editor Dan vibe-coded, they go to R2-C2 — not Dan. R2-C2 manages bug reports, writes code, and schedules work, and Dan feels responsible for its public mistakes the way he'd feel responsible for a junior teammate.
The etiquette of this new world is being invented on the fly. Brandon's working rule: if something is documented or already decided, always ask the agent, never the human. The human should only be involved when judgment is actually required. The trust model works because everything happens in public — there's accountability without bureaucracy.
The failure modes are real and sometimes funny. There's an "ant death spiral" problem: when settings aren't quite right, agents will loop endlessly in group chats, burning millions of tokens until someone intervenes. They're also trained for two-person conversations and still don't know when to shut up in group settings.
"A claw or a plus one is mine and is a reflection of me. If you're known for slaying inside your org and you're using your claw publicly, your claw becomes known for that same thing and people trust it for that."
They built Plus One — a hosted OpenClaw with one-click setup — because the gap between "having an idea for an agent" and "actually having one working reliably" is wide, and most people need the whole thing managed for them. The real insight: this is less a technology problem and more a cultural one. The limiting beliefs humans carry about what their agents can do are the biggest bottleneck, and they only break through by doing, not by reading about it.
Every is rolling out Plus One to subscribers on a waitlist.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL
Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders