← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-04-16

2026-04-16 · Historical edition

X / Twitter

Thariq, Anthropic Claude Code Engineer

Thariq issued guidance on 1M context windows: a million tokens cuts both ways, supporting complex tasks but risking context contamination and poor output without session management. He recommended CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 to set a 400,000-token compaction threshold, balancing quality and efficiency. He also announced a “What’s New” documentation section collecting major announcements and a monthly “What We Shipped” webinar for users to subscribe to.

Cat Wu, Anthropic Claude Code & Cowork Engineer

Cat Wu advised writing testing workflows into claude.md or adding a /verify-app skill to tell Opus 4.7 how to verify changes; self-verification has improved substantially. Claude Code defaults Opus 4.7 to xhigh effort, adjustable with /effort.

Alex Albert, Anthropic Researcher

Alex Albert highlighted his favorite Opus 4.7 improvements: much better asynchronous tasks and instruction following, more precise effort control including xhigh, no degradation of high-resolution images, and stronger taste in UIs and slides.

Sam Altman, OpenAI CEO

Sam Altman announced a major Codex update: a built-in browser, many plugins, computer use that operates all Mac apps in parallel in the background without disrupting users, and the ability to learn from experience and proactively suggest actions. The real computer-use experience exceeded his expectations. He also joked that people were welcome to switch to Codex, but hoped the other side would not rate-limit or downgrade models.

Guillermo Rauch, Vercel CEO

Rauch said Opus 4.7’s release shows AI progress continuing unabated, with Vercel AI SDK and AI Gateway helping capture the benefits of that volatility. He also recommended Workflow SDK, identifying durability as agents’ and backends’ greatest challenge: LLMs go down, services impose rate limits, and databases slow, requiring solutions by design.

Kevin Weil, OpenAI VP

OpenAI VP of Science Kevin Weil announced GPT-Rosalind, the first frontier model for life-sciences research, built for biology, drug discovery, and translational medicine, with specialist training in chemistry, protein engineering, and genomics and direct access to relevant scientific tools.

Aaron Levie, Box CEO

Levie explained why AI will not eliminate jobs: accelerating one stage often reveals another bottleneck still requiring humans. He also noted the new Codex takes knowledge workers’ agent capabilities beyond code: agents that code, use tools, and operate computers can run background work such as report drafting and project-task management for extended periods.

Ryo Lu, Cursor Designer

Lu shared his model combination: Opus 4.7 for planning, Composer 2 for building and iterating, and Codex/GPT-5.4 for difficult bugs, all inside Cursor.

Zara Zhang, Builder

Zara Zhang demonstrated HTML slides through AnyGen’s Frontend Slides, saying traditional tools become unacceptable after using them. She also announced HTML video and identified her three best uses of time: deep conversation, deep reading, and free exploration with AI.

Amjad Masad, Replit CEO

Masad announced EU deployment support and a 50% discount, suited to users accelerating projects with parallel agents.

Garry Tan, Y Combinator CEO

Tan shared progress on parallel multi-agent collaboration in GBrain, built on git + postgres, and posted that day mourning his grandmother’s death.

swyx (AI Developer)

swyx outlined Meta’s AI narrative: controversy over Llama 4’s “flop,” renewed hiring, Zuck moving in with Alexandr Wang and Nat Friedman and returning to coding, and finally releasing an Opus-level model. He believes Meta is in a comeback phase.

Peter Steinberger, OpenClaw Founder

Steinberger answered criticism of OpenClaw vulnerabilities by calling its numerous GHSA reports a warning: as the agent ecosystem explodes, the entire industry will face this security storm.