← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-03-31

2026-03-31 · Historical edition

X / Twitter

Anthropic Claude Code engineer Cat Wu

Anthropic Claude Code & Cowork engineer Cat Wu announced a much simpler setup experience for the integration between claude.ai and GitHub: users can now run /web-setup in a local claude session to connect to the web version directly using their local GitHub credentials.

Box CEO Aaron Levie

Box CEO Aaron Levie argued that in the AI agent era, the greatest opportunities inside enterprises belong to employees with resources and ambition who proactively redesign workflows. He emphasized that automating work with agents requires real upfront investment: organizing unstructured data and clearly defining business rules before agents can truly take over tasks. He added that agents capable of coding are also excellent at operating tools, seamlessly connecting nondeterministic intelligence with deterministic system interactions within the same workflow.

OpenClaw founder Peter Steinberger

Peter Steinberger shared a series of updates: CodexBar beta added experimental multi-account support; he shared an old post “somewhat relevant” to the current security situation; and he released a new Claw beta with extensive reliability and security improvements and a new task system supporting more stable subagents and scheduled tasks. He also jokingly described himself as “creating new jobs” and revealed how the new version’s tagline was chosen.

Andrej Karpathy (AI researcher)

Andrej Karpathy warned that axios, npm’s most popular HTTP client library with 300 million weekly downloads, had suffered a new supply-chain attack. Scanning his own system, he also found infected dependencies introduced through packages he had used while experimenting with Gmail/Google Calendar CLIs a few days earlier, and urged developers to investigate immediately.

Replit CEO Amjad Masad

Replit CEO Amjad Masad also addressed the npm supply-chain attack, describing Replit’s security mechanisms for protecting users against large-scale JavaScript supply-chain attacks. He cautioned developers coding on laptops or in environments outside Replit to be especially vigilant.

Y Combinator CEO Garry Tan

Y Combinator CEO Garry Tan continued refining GStack’s /review skill and shipped the new version that evening. He also reflected on opportunity in the AI era: whether you are 16, 30, or 60, age ceases to be an obstacle if you can get agents to work their magic for you and bring taste and initiative yourself.

Every CEO Dan Shipper

Every CEO Dan Shipper released a new version of Compound Engineering and led his team in exploring Claude Code’s source code. They discovered a hidden “buddy” concept: each user receives a unique CLI mascot generated from their user ID, prompting his joking remark that “Claude Code is catching up with OpenClaw.”

Builder Peter Yang (AI tutorial creator)

Peter Yang shared how he built a mobile app from scratch with Claude Code in a few hours, without a programming background, to track his workouts while traveling in China. He also recorded his impressions of returning to Shanghai after 10 years: the city felt more modern than New York or even Tokyo, and its infrastructure impressed him.

Builder Zara Zhang

Zara Zhang argued that if Claude Code or Codex can access and operate all of Lark’s (Feishu’s) features—including chat, calendars, meetings, documents, and spreadsheets—this all-in-one collaboration tool also becomes the most agent-friendly collaboration platform. She shared explorations of related Lark CLI tools.

swyx (AI developer)

swyx noted that Microsoft had just appointed a CVP whose sole responsibility is OpenClaw-related work, marking AI coding agents’ entry into major technology companies’ strategic agendas. He also shared views on the right framework for evaluating local LLMs, arguing that most evaluations miss the central issue.

Podcasts

Latent Space — The Stove Guy: Sam D'Amico Shows New AI Cooking Features on America's Most Powerful Stove at Impulse

This episode of Latent Space takes viewers on a tour of the Impulse stove, regarded as the most powerful induction stove currently on the market.

Product core: a lithium-battery-powered kitchen revolution

Impulse founder Sam D'Amico walked the hosts through the system’s central design: a 3 kWh lithium iron phosphate (LFP) battery built into a household stove adds 15,000 watts of output on top of wall power, equivalent to twice a conventional induction stove’s total power. This lets Impulse boil a liter of cold water in around 40 seconds while delivering real-time responsiveness and an analog knob feel comparable to gas stoves.

From power electronics to induction coils: the entire stack built in-house

Sam described the modules the Impulse team redesigned from scratch: custom power electronics boards, temperature-control sensors, and proprietary induction coils with full magnetic-field simulation. This gives them complete control of the loop from turning the knob to heating the pan, with continual performance improvements through OTA updates. An upcoming update, for example, will triple temperature-control response speed.

Consumer electronics thinking versus appliance thinking

Sam compared Impulse’s approach to the difference between Tesla and traditional carmakers: design the product like a phone, start from software-defined hardware, and control every step from knob to coil. He also noted that widespread household battery devices effectively provide distributed storage for the grid, easing congestion caused by AI data centers, EV adoption, and household electrification.

AI features and product strategy

The episode demonstrated Claude providing real-time cooking temperature recommendations. Impulse follows a two-track strategy: selling premium consumer stoves itself while licensing technology to major appliance OEMs through its Impulse Core module platform. This asset-light channel expansion lets it enter global markets without a large sales team. It has signed partnerships with several major appliance manufacturers and a retail channel agreement with Zephr.

The startup journey

Sam’s background includes consumer electronics such as Meta Quest Pro controllers. From its founding in July 2021 to the first product reaching a customer’s home, Impulse took nearly 4 years, encompassing full hardware R&D, firmware development, UL safety certification including UL 858, and durability testing on more than 200 test units.

Data Driven NYC — Can One Person Build a Billion-Dollar Startup? — The General Intelligence Company of New York

The speaker in this Data Driven NYC episode is General Intelligence founder Andrew, a second-time founder who has been a CEO since age 18. His topic is how to realize the vision of a “one-person, ten-billion-dollar company.”

Core framework: four layers of an AI-driven company

Andrew divides company operations into four basic functions: create the product, sell it, support it, and operate it. He believes that as agents mature, all four layers can be handled autonomously by AI, with people setting goals and supervising only at the top. General Intelligence focuses on the “manager agent” layer: middle management coordinating underlying vertical agents in customer service, finance, marketing, and other areas.

First-generation product Co-founder: a testing ground across 8000 startups

General Intelligence’s first product, Co-founder, launched on September 8, 2024. It is an AI assistant with memory that runs agents in the cloud, drafts emails, and operates browsers. More than 8000 startups already use it to automate operations. Through this product, the team validated agents in email replies, LinkedIn candidate research, API automation, and other scenarios, while gaining extensive practical experience in driving complex agents through difficult tasks.

New product Co-founder CTO: a fully autonomous engineering department

On the day of the talk, Andrew publicly demonstrated Co-founder CTO for the first time, positioning it as “the world’s first fully autonomous engineering department,” running entirely in the cloud. The workflow receives a feature ticket, calls Claude Code to write code in the cloud, automatically creates a PR and deploys to a Vercel Preview Environment, then has a browser agent, Browser Base, sign in and test the feature before automatically merging after verification. No local development environment is required.

The interface draws inspiration from real-time strategy games such as StarCraft and Civilization: users dispatch and track agent tasks on a canvas-like interface. The system also integrates Vercel infrastructure monitoring, automatically triggering an agent to fix failed deployments without human supervision.

Measured results and business progress

After using CTO internally, the team improved engineering efficiency by 400%, with each engineer merging around 25 PRs a day on average. On some peak days, Opus token spending reached $4000, bringing compute expenditure into the same order of magnitude as salaries. Plans call for AI customer service, Co-founder Support, and an AI sales/growth team, Co-founder CRO, to roll out over the next 6 months, gradually covering the full company operations loop.