← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-04-02

2026-04-02 · Historical edition

X / Twitter

Replit CEO Amjad Masad

Masad declared that the “one-person billion-dollar company” had arrived, congratulating @galligator. He considers this one of the fastest periods of wealth creation in human history. Replit also launched zero-configuration enterprise-grade authentication for all users, including free users, and opened a sales office in Salt Lake City that is hiring. He also shared a practical feature: auditing your website's SEO directly in Replit.

Vercel CEO Guillermo Rauch

Rauch revealed that Vercel signups grew 52% month over month, following 23% and 17% in the preceding two months—the growth rate itself is accelerating.

Y Combinator CEO Garry Tan

Tan posted frequently this week. He called local models “a very, very good thing” and expressed enthusiasm for OpenClaw, saying its “Apple II moment is still ahead.” He responded with a screenshot to online “haters” criticizing his comments on lines of code (LOC), declaring he “won't stop building things.” He also backed California's BASED Act (SB 1074) as a strong response to monopolistic behavior by major platforms including Amazon, Apple, Google, and Meta.

OpenClaw Founder Peter Steinberger

Steinberger shared two widely discussed posts. First, he never uses Plan Mode, calling it a feature for people who are “Claude-pilled but can't change their habits”; simply talk to the agent. Second, citing the Linux kernel security mailing list, he said reports of vulnerabilities in AI-generated code rose from two or three per week two years ago to more than ten weekly, predicting disruption for some open-source projects.

Box CEO Aaron Levie

Levie launched Box Agent, an AI agent that works across the entire Box file system, fully inherits enterprise security and access controls, and is deeply optimized for enterprise content.

Andrej Karpathy

Karpathy shared his most valuable recent personal practice: using LLMs to build a “personal knowledge base,” or LLM Wiki, for research topics of interest. He feeds the model material he reads, asking it to integrate everything into an evolving wiki. Much of his recent token consumption has gone toward this. He envisions each future question to a frontier LLM automatically summoning a team of LLMs to iteratively build a temporary wiki, gather evidence, and produce a complete answer, scaling the entire retrieval–understanding–synthesis process. The two posts received more than 19,000 likes combined.

Cursor Designer Ryo Lu

Lu officially announced Cursor 3, emphasizing a combination of power and simplicity across all local and cloud projects. Its interface uses progressive disclosure: start simple, reveal additional tools when needed, and preserve focus and control.

Every CEO Dan Shipper

Shipper shared the team's full review after a week testing Cursor 3.0. He also questioned the “one-person billion-dollar company” claim: although @galligator is impressive and uses AI extensively, he does not consider it a true “one-person AI company,” because a growth marketing team works behind the scenes.

Sam Altman (@OpenAI)

Altman called TBPN, the technology and business podcast, his favorite tech show. He announced OpenAI's intention to support its continued operation, without expecting the show to “go easy” on OpenAI as a result.

Anthropic's Official Claude Account

Claude announced that Computer Use in Claude Cowork and Claude Code Desktop now supports Windows. The post received more than 7,700 likes.

Podcasts

Unsupervised Learning — Ep 83: Owning the System of Record, AI-Native Org Charts, & Why ITSM is The Most Vulnerable Legacy Category

This episode interviews Serval founder and CEO Jake Stout. Serval is an AI platform for employee support across IT, HR, legal, and other areas, automating workplace support requests end to end. It recently raised funding led by Sequoia at a billion-dollar valuation and ranked first among mid-stage companies in the latest Enterprise 30 selection.

Jake's central argument is that AI-native enterprise software must also build its own system of record instead of depending on legacy platforms. He believes reliance on platforms such as ServiceNow fundamentally limits AI products, whereas owning the system of record gives companies more control over features, sales, and customer relationships.

Serval's organizational chart is highly AI-native: collaboration between engineers and AI directly affects how flat the team can be. AI has taken over repetitive tasks traditionally assigned to junior engineers, enabling a smaller, stronger engineering team in which everyone makes higher-level decisions.

Jake explains why ITSM, or IT service management, is the most vulnerable legacy software category: its core value is helping employees solve problems, precisely where agents can be faster, more accurate, and cheaper than people. Traditional ITSM vendors burdened by legacy systems, extensive custom code, and poor user experiences face the most direct replacement threat from AI.

The MAD Podcast with Matt Turck — AI is Already Building AI | Google DeepMind's Mostafa Dehghani

Google DeepMind researcher Mostafa Dehghani is a key contributor to the Vision Transformer, Universal Transformer, and Gemini's multimodal architecture. He and Matt Turck explored the current state and future of AI self-improvement.

Dehghani's central claim is that recursive self-improvement is already happening, although most people do not realize it. Nearly every frontier lab relies heavily on its previous model generation to build the next. Long-term planning and full automation remain missing, but the direction is clear. Once the loop is fully automated, humans will cease to be the bottleneck, which he expects to produce a major leap in capability.

Dehghani distinguishes two levels of “looping.” Inference-time looping—chain of thought, test-time compute, negative sparsity, and similar techniques—devotes more compute to a particular problem. Development-level looping, where models improve the next generation, is the highest-level extension of the same direction. He cited Coparty's automated research project as a strong early example of AI making substantive research contributions.

Dehghani also discussed the evolution of Google DeepMind's Imagen 3 image-generation architecture and how continual learning could fundamentally disrupt RAG systems and enterprise data pipelines. If models learn from the world and update their weights in real time, RAG's role as an external retrieval mechanism would shrink substantially.