X / Twitter
Claude, Anthropic’s Official AI Account
Claude Cowork added interactive charts and diagrams, available to beta users on all paid plans. This extends collaboration beyond text into visual data presentation, letting users generate interactive rather than merely static charts directly in Claude conversations.
https://x.com/claudeai/status/2047047633416397076
Aaron Levie, Box CEO
Levie offered two substantive enterprise AI insights. First, real enterprises face legacy stacks, data scattered across tools, undigitized knowledge, and the need to maintain daily operations—barriers very different from startups’. This creates major opportunities for domain- and workflow-specific agent deployers, whether software or service companies. The FDE (Field Deployment Engineer) model will persist because enterprises need vendors to drive implementation and change management. Second, ChatGPT’s new custom sales-assistant agent can access Box content through MCP and CLI, working in foreground or background to assist knowledge work. He sees this as a milestone in software becoming headless, supporting his view of headless-platform opportunities.
https://x.com/levie/status/2046805326784319663
https://x.com/levie/status/2047028112626749645
Ryo Lu, Cursor Designer
Ryo Lu reflected on “overcooking.” AI sharply reduces the cost of adding features, encouraging continual additions whose individual decisions are reasonable but collectively create chaos. “Most new ideas aren’t really new, just repackaged existing concepts,” he observed. Accumulation feels like progress, but thinking has not deepened; people merely replicate themselves amid confusion. A product has a core, with interconnected, balanced elements, and continual additions weaken that pull. The AI era needs the ability to see through confusion, return to essentials, and cut everything not serving them, rather than more tools for producing rubbish.
https://x.com/ryolu_/status/2046957675079237668
Amjad Masad, Replit CEO
Replit formally testified before Congress supporting the BASED Act, intended to prevent large technology companies manipulating software markets. Masad opposes platform monopolies squeezing small developers and independent tools, consistent with Replit’s mission that anyone can create software. This was Replit’s first public legislative intervention, marking increased policy engagement.
https://x.com/amasad/status/2046762468199071765
Garry Tan, Y Combinator CEO
Garry Tan announced a major GBrain Minions upgrade. This job server he built for his agents schedules and executes background work. Getting GBrain to guide OpenClaw/Hermes agents correctly required extensive debugging, he said, but the configured system works impressively.
https://x.com/garrytan/status/2046804979160334433
Zara Zhang, Builder
Zara Zhang shared an impressive learning technique: asking Claude Code to generate an HTML file visualizing the current context window’s state, a “wild” way to understand how it works. She also recommended three older books relevant to the AI era: The Mythical Man-Month (1975, Fred Brooks), Diffusion of Innovations (1962, Everett Rogers), and Player Piano (1952, Kurt Vonnegut).
https://x.com/zarazhangrui/status/2046758723998425433
https://x.com/zarazhangrui/status/2046853431554719753
Peter Steinberger, OpenClaw Founder
Steinberger released three updates. discrawl 0.3.0 adds Git-backed archive synchronization: Discord archives can be published to private repositories and queried locally without each user needing bot credentials, alongside automatic refresh, activity reports, field notes, and faster imports. CI parallelization cut builds from 8 minutes to 2, with resources sponsored by Blacksmith. OpenClaw 2026.4.21 fixes a critical npm-update bug breaking plugin runtime dependencies, adds Docker end-to-end tests to prevent Telegram, Discord, and Slack plugins breaking after upgrades, and adds OpenAI Image 2 support. Update with `npm i -g openclaw@latest`.
https://x.com/steipete/status/2046748122928263345
https://x.com/steipete/status/2046787353906167992
https://x.com/steipete/status/2046803162590335240
Josh Woodward, Google Labs VP
Woodward recommended DESIGN.md in Google Stitch, calling it one of his favorite current designs in Stitch; details were undisclosed, but it prompted community discussion. He also announced Pomelli, a newer Google Labs experiment, entering Europe and attracting substantial SMB adoption. Both point to continued investment in AI productivity tools.
https://x.com/joshwoodward/status/2046754179499356594
https://x.com/joshwoodward/status/2046763674199794061
Nikunj Kothari, FPV Ventures Partner
Kothari described his most effective networking approach: give 10-15 trusted people complete permission to make blind introductions without asking first. He considers it the fastest way to meet interesting people, creating more serendipity than traditional double opt-in. He also reposted a real-time pixel-generation demo, saying every pixel will eventually be generated in real time; it is only a matter of time.
https://x.com/nikunj/status/2046821454143250636
https://x.com/nikunj/status/2047024714116419665
Podcasts
AI & I by Every — The AI Sandwich: Where Humans Excel in an AI World
Key takeaway: In the agent era, human value concentrates at the beginning—defining problems—and the end—polishing output—while agents have largely solved the execution in between.
Kieran Klaassen is Quora’s GM and creator of the Compound Engineering framework and plugin. Every’s engineering team uses it extensively, and it has spread widely in the agent-programming community.
Compound Engineering has four steps: Plan, Work, Review, and Compound. The last is the most important and distinctive: useful lessons discovered during execution are saved in the repository so future planning or work can retrieve them and avoid repeating mistakes. Klaassen considers this the plugin’s strongest capability and the real reason one engineer can deliver like a five-person team.
His claim that execution is solved is bold: with a sufficiently good plan, LLM execution can be fully trusted without a human in the loop. Models are mature at deep, sequential work lasting hours or days. Review and planning are catching up quickly. This raises a question: if the middle is automated, where does human work remain?
Compound Engineering’s evolution offers an answer: humans occupy the sandwich’s ends. At the beginning, Ideate and Brainstorm, AI should support deep human thinking, ask questions, and avoid rushing into autonomous planning. Humans must really think; LLMs support rather than replace them. At the end, Polish, after automated tests pass, humans click through, feel the result, and fine-tune it. Klaassen compared music: “Practicing a piece 100 times is the boring middle, but performing on stage is special, as is composing from nothing. Both ends remain where humans are most valuable in software.”
On when to remain in the loop or relinquish control, Klaassen rejects the assumption that humans must always be present. Knowing when to intervene and when to let go concentrates energy where thinking matters. He acknowledges ideation will eventually be automated too, but believes anyone releasing something must at some point say “this is mine.” That agency cannot be fully outsourced. Like art, if it is to belong to you, it must come from you or connect genuinely with you.
Dan Shipper offered a useful AGI criterion: it is true AGI when running an agent continuously for 24 hours makes economic sense. OpenClaw and similar tools still need plans and heartbeats, rather than truly continuous autonomy. He believes reaching this requires fundamental changes to language-model architecture.
Klaassen advises practitioners to find what brings joy and use LLMs to do more of it. Engineers may pursue elegant code and beautiful architecture; product managers may polish user-experience details. Engineers will become more like product managers and managers, with the overall direction toward the interesting parts rather than repetitive execution.