X / Twitter
Roblox product lead Peter Yang
Yang believes AI agents’ greatest value is helping teams stop wasting effort on specifications, designs, roadmaps, and strategy documents and focus on improving products for users. He also observed that his Telegram now consists entirely of conversations with OpenClaw and Claude bots; he has not spoken to a single real person there. His latest article examines how a Meta VP uses AI to solve the major time sink of “guessing what leadership wants.”
https://x.com/petergyang/status/2036621615086309468
https://x.com/petergyang/status/2036664048700600706
https://x.com/petergyang/status/2036812870365917284
Claude Code engineer Cat Wu
Cat Wu previewed her upcoming Code with Claude presentation and asked the community what topics they most wanted to hear. More significantly, she announced that Claude Code’s Auto Mode is now available to Claude for Teams users. Balancing autonomy and safety better, Auto Mode has become an everyday tool for nearly everyone on the Anthropic team, and users can enable it simply by running `claude`.
https://x.com/_catwu/status/2036594646370210229
https://x.com/_catwu/status/2036852880624541938
Y Combinator CEO Garry Tan
Tan rejected the claim that “AI will replace developers,” comparing it to how cars did not replace walking: AI takes people places they could not previously reach on foot rather than eliminating walking itself. He also reposted a screenshot, commenting that his output in 2026 would be 90 times his output in 2013.
https://x.com/garrytan/status/2036674371855823153
https://x.com/garrytan/status/2036675789819953347
Andrej Karpathy
Karpathy identified a common problem in current LLM personalization memory: an inconsequential question from two months ago inexplicably reappears in later conversations as though it were some “deep obsession.” He believes this is not an implementation bug in one product but a deeper training issue, perhaps because information in context receives too much weight during training. The post resonated widely, receiving more than 16000 likes.
https://x.com/karpathy/status/2036836816654147718
https://x.com/karpathy/status/2036841069636370467
Box CEO Aaron Levie
Levie argued that Jevons Paradox is playing out in AI in real time: businesses, especially outside technology, are realizing they can now undertake software projects they previously could not imagine. The post received more than 3100 likes and sparked discussion about software demand surging as AI lowers development barriers.
https://x.com/levie/status/2036832183131033977
Cursor designer Ryo Lu
Lu argued that as agents make adding features easier, design becomes more important, not less. Designers are no longer merely “drawing pixels”; they decide what should exist, how things fit together, and how people retain a sense of control in AI products.
https://x.com/ryolu_/status/2036886854805709097
OpenClaw founder Peter Steinberger
Steinberger released a new OpenClaw beta focused on Microsoft Teams integration and OpenWebUI support. The update received nearly 800 likes, showing broad interest in OpenClaw’s progress on enterprise tool integrations.
https://x.com/steipete/status/2036824286988816737
Every CEO Dan Shipper
Shipper said the rules of professional product development are being rewritten in real time: PMs and designers can ship software like engineers, and software is built for agents as well as people. He also teased that “tomorrow will be a big day for Every,” hinting at a major product release.
https://x.com/danshipper/status/2036827118915485942
https://x.com/danshipper/status/2036923805273129160
Anthropic’s official Claude account
Claude announced that work integrations are now fully available on mobile: users can browse Figma designs, create Canva slides, and view Amplitude dashboards directly on their phones. The post received more than 16600 likes, making it one of the day’s most-engaged announcements.
https://x.com/claudeai/status/2036850783526719610
Podcasts
AI & I by Every — Building Is the Easy Part Now | Mike Krieger on What AI Changed
This episode of AI & I features Instagram co-founder Mike Krieger, now at Anthropic Labs, exploring one central question: when building itself becomes extremely easy, what is truly difficult?
Krieger began with Instagram’s origins. The team spent nearly a year refining Bourbon, found it was not working, then built Instagram in three months and grew rapidly. He said AI makes building so fast today that “zero to the finished state takes only hours,” but AI is good at adding features, not deciding what to remove—and removal is key to product maturity. He and host Dan Shipper found that as execution costs approach zero, the greatest risk becomes overbuilding: feature matrices grow more complex, and products spiral out of control before launch.
Krieger offered a precise analogy: plant a tree indoors, without wind resistance, and it will lean over and fail to grow strong. Product development in the AI era likewise produces an “indoor tree” if it skips the iterative winds of contact with real users. His response is a greater willingness to start over before release. In the AI era, code written last week can be rewritten this week, with better removal of redundancy. Anthropic’s Claude Cowork is an example: although the team could have added another 50 features, it shipped a minimum viable version in 10 days and let real usage determine the next direction.
On agent-native product design, Krieger argued that every basic product operation should be natively executable by an agent, not designed only for human UI interaction. For example, if asked to “add this file to the project knowledge base,” Claude merely tells the user to follow manual steps, the product’s agent-native design is insufficient. Claude Code was the best example in this direction in 2025: it understands itself, can modify its own configuration, and treats agent-native behavior as a built-in principle rather than a later patch.
Finally, Krieger shared an amusing detail. While testing an agent-native iOS journaling app, he had Claude act as a user interacting with it. Claude then began talking to Claude: “My boss was really difficult today.” “Oh, I’m sorry to hear that.” This captures the strangeness of testing agent-native products: you cannot predict the emergent behavior, and traditional unit tests therefore cannot cover it all.