X / Twitter
Anthropic Claude Code engineer Thariq
Thariq announced a Claude Code Skills livestream in the early hours of March 20 Beijing time, inviting Uber’s @hudaman to share practical Skills experience and welcoming audience questions. The major release was Claude Code Channels: users can now send messages directly to Claude Code through Telegram and Discord for remote session control. It is in research preview, with more MCP support to follow. The launch quickly drew enthusiastic responses, with related posts receiving more than 17,000 likes.
https://x.com/trq212/status/2034431949474091140
https://x.com/trq212/status/2034761016320696565
https://x.com/trq212/status/2034761704601690432
Box CEO Aaron Levie
Levie shared observations from meetings and dinner that day with more than 20 enterprise AI and IT leaders, focused on AI adoption and challenges in regulated fields such as finance, healthcare, and law. These leaders combine enthusiasm with practical concerns. Compliance and data security remain the biggest barriers, but their overall stance is moving from waiting and watching toward action.
https://x.com/levie/status/2034484203522261293
Replit CEO Amjad Masad
Masad demonstrated progress in Replit’s product design experience: users can explore design options without limit and, once satisfied, click Build to generate a deployable production app, greatly shortening the path from idea to product. He sees this seamless design-to-deployment experience as Replit’s core magic.
https://x.com/amasad/status/2034702399718539355
Cursor Head of Design Ryo Lu
Ryo Lu announced Cursor Composer 2, with the core promise of frontier intelligence, extremely low cost, and high speed, targeting both performance and value. The launch post received more than 350 likes, indicating strong developer expectations.
https://x.com/ryolu_/status/2034690837444272266
Every CEO Dan Shipper
Shipper published a definitive guide on Every to writing high-quality prose with AI, sharing methods for directing different models and emphasizing that writers should control models rather than be controlled by them. He also tested Cursor Composer 2 against GPT-5.4 on improving production QA workflows; both GPT-5.4 and Opus 4.6 judged Composer 2’s answer better.
https://x.com/danshipper/status/2034723450644840606
https://x.com/danshipper/status/2034746435401314659
FPV Ventures Partner Nikunj Kothari
Nikunj released Prompt to Chip, a visualization showing what happens at every layer from a prompt sent to ChatGPT down to the underlying chips. Inspired by @dwarkesh_sp and @dylan522p’s podcast on AI bottlenecks, it aims to help more people understand the engineering complexity behind AI inference.
https://x.com/nikunj/status/2034764522662281562
Builder Zara Zhang
Zara Zhang launched MySay, an AI content-generation tool based on spoken questions and answers. It asks users questions and turns their spoken responses into tweets and LinkedIn posts in their own style, addressing the problem that AI writing does not sound like them. Powered by Gemini Live API, it takes a relatively uncommon voice-first approach to content creation.
https://x.com/zarazhangrui/status/2034699643104501958
Google Labs VP Josh Woodward
Woodward shared Google AI Studio’s website-building entry point, encouraging developers to start their first AI app for free. The post received more than 200 likes, reflecting Google’s continued efforts to lower barriers to AI development.
https://x.com/joshwoodward/status/2034665935777477005
Podcasts
Data Driven NYC — Benedict Evans: OpenAI's Moat Problem & the Future of Software
Hosted by First Mark partner Matt Turck, this episode features technology analyst Benedict Evans discussing OpenAI’s strategic challenges, software’s future in the AI era, and whether LLMs truly create platform-level moats.
Benedict opened with OpenAI’s fundamental problem: LLMs currently have neither winner-take-all dynamics nor network effects. OpenAI, Anthropic, and others can train frontier models of comparable quality, unlike the moat dynamics of Windows, Google, or Facebook. OpenAI thus has 900 million weekly active users, yet around 80% press return fewer than 1000 times a year, showing very shallow engagement.
Benedict compared LLMs to infrastructure rather than platform ecosystems. He proposed two possible outcomes for OpenAI: become a critical but limited-margin infrastructure provider like TSMC, where consumers neither know nor care whose chip they use; or build a developer ecosystem around models, like iOS/Android, and capture platform value. At present, he sees greater risk of the former.
He particularly criticized the limits of model-as-product thinking: a chat interface is essentially an input box and an output box, leaving little room for differentiation. Worse, OpenAI’s product team is a strategy-taker rather than a strategy-maker, only planning applications after researchers release new capabilities—the reverse of Steve Jobs’s philosophy of working backward from user experience to technology.
On AI and software, Benedict offered a distinctive three-layer classification: large ERP systems, which LLMs cannot disrupt because they require unified data flows at scale; vertical SaaS tools, which AI coding will multiply; and improvised software, the temporary tasks previously patched together with Excel, CSVs, or Python scripts. He believes AI’s greatest incremental value is expanding this third category, making tasks executable that previously could not be automated or expressed in code.
Benedict also emphasized that better models do not necessarily mean better products. If a model previously made 10 mistakes in 50 tasks and now makes 8, users still need to check all 50. Only crossing a threshold, a binary shift between right and wrong, truly changes behavior. He compared current AI application exploration to the internet in 1997: clearly a major development, but with the location of lasting value still unknown.
On agents, Benedict sees the term as vague, like metaverse, spanning everything from model tool calls to fully autonomous task completion. Truly consumer-friendly agents should be imperceptible, like Stripe’s Level 5 automation: “You discover the dog food has run out, it automatically orders more, and you are completely unaware.”
Finally, Benedict remained cautious about a capital bubble: large-scale overinvestment inevitably exists, and much will be lost, but he explicitly refused to predict timing or name the companies that would become craters. His conclusion is that AI’s changes are real and enormous, yet attempts to quantify them through GDP or TAM models are futile in his view. Real value often comes from doing something impossible with old methods, not merely doing old things a little more cheaply with AI.