X / Twitter
Peter Steinberger, OpenClaw Founder
Peter Steinberger released clawpatch 0.1.0, a tool that maps a codebase into semantic feature slices, reviews each for bugs and quality, and records explicit fix attempts and verification results. It can be installed with `npm install -g clawpatch`; the post received nearly 1,000 likes. He also revealed OpenClaw's development philosophy: approximately 100 Codex instances run continuously in the cloud to automatically review every PR and issue, exploring the central question, "How would we build software in the future if tokens were no longer a constraint?" That post received more than 3,700 likes. He also shared his experience using Svelte as an alternative to React and a case where Codex's review skill found an unfinished feature in discrawl.
https://x.com/steipete/status/2055364630709448970
https://x.com/steipete/status/2055405041843052792
https://x.com/steipete/status/2055402519841411165
https://x.com/steipete/status/2055203470941061600
Guillermo Rauch, Vercel CEO
Guillermo Rauch offered his central view of engineers' growth in the AI era: if you excel at managing agents and deeply understand fundamentals, you will be unbeatable. He believes agents' amplification of output means neither ability can be neglected; the post received nearly 1,900 likes. He also posted a city photograph, remarking, "This city is so beautiful. Unbeatable."
https://x.com/rauchg/status/2055278852931530784
Aaron Levie, Box CEO
Aaron Levie published three substantial posts. He humorously described today's AI development pace: spend a year building scaffolding for an agent harness, then a new model makes it all obsolete. He noted that AI is disrupting functional boundaries by making it easier for everyone to work in adjacent functions, requiring teams to redefine roles for the agent era. He also proposed "headless software" as the future, receiving 250 likes.
https://x.com/levie/status/2055136119969529993
https://x.com/levie/status/2055143231625818554
https://x.com/levie/status/2055357619888595271
Amjad Masad, Replit CEO
Amjad Masad announced two important developments. Replit has resolved its dispute with Apple and returned to the App Store after four months, with many new features awaiting release. He also announced that users can directly import websites vibe-coded outside Replit and receive a mobile app version for free; this feature received more than 430 likes.
https://x.com/amasad/status/2055185058282226146
https://x.com/amasad/status/2055097107758076254
Sam Altman, OpenAI CEO
Sam Altman shared his attitude toward user feedback: he appreciates the team taking every report seriously, even when the eventual conclusion is "I've grown accustomed to the current level of magic and now want more." The post about managing expectations received more than 3,200 likes.
https://x.com/sama/status/2055356452286640630
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari shared a startling `/goal` experience: an agent automatically worked through more than 2,000 database records, fixing product images, frontend bugs, and product descriptions while using the browser and web search for real-time information, all without human intervention. He exclaimed, "/goal with the right tools is AGI," receiving more than 300 likes.
https://x.com/nikunj/status/2055426430654439485
Swyx, AI Engineer Founder
Swyx argued that blogs begin to die when they become "by some team" rather than bearing an individual's byline: "Ownership creates accountability." He also shared an amusing moment at AI Engineer Miami: a speaker called SAP Concur "soulless software," with SAP employees in the audience, and was subsequently invited to advise on AI transformation for 6,800 SAP employees. "He mocked SAP, and SAP concurred."
https://x.com/swyx/status/2055079344632926603
https://x.com/swyx/status/2055113979857703350
Dan Shipper, Every CEO
Dan Shipper shared a detailed retrospective on building an agent-as-a-service platform on OpenClaw. OpenClaw is powerful, but building on it as a platform provider is extremely challenging: iterations are too fast, rollbacks frequent, and it is difficult to serve users reliably through it as an intermediate layer. The candid product retrospective received 130 likes and offers an important reference for other teams hoping to build services on OpenClaw.
https://x.com/danshipper/status/2055347527457886336
Podcasts
Unsupervised Learning — Ep 86: Yann LeCun on Leaving Meta, Breaking The LLM Paradigm, & Why Hinton is Wrong
In this episode of Unsupervised Learning, Redpoint investor Jacob Efron speaks with Yann LeCun, one of AI's foundational figures. LeCun shared the Turing Award with Geoffrey Hinton and Yoshua Bengio and led Meta FAIR, its fundamental AI research lab, for many years. The discussion covers his fundamental objections to the LLM paradigm, his departure from Meta, the direction of his new company Ami, and the core of his disagreement with Hinton.
On LLM limitations, LeCun maintains that LLMs can never achieve genuine general intelligence because they lack explicit modeling of the physical world. He identifies their fatal weakness as failing to predict action consequences accurately enough, causing failure in practical tasks requiring continuous precise control, such as robotics and industrial process control. He sees world models as the paradigm needed to overcome the LLM bottleneck: AI systems must predict consequences before they can truly execute complex tasks.
On his disagreement with Hinton, LeCun traces the divergence to GPT-4's release in 2023. After encountering GPT-4, Hinton concluded that AI was approaching human intelligence and might have subjective experience, leading to strong warnings about safety risks. LeCun rejects that inference, regarding Hinton's computational analogy—inferring model intelligence from the number of cortical neurons—as fundamentally flawed, and rejecting even more strongly the judgment that LLMs are close to AGI.
On leaving Meta, LeCun said misunderstandings of his role were one reason for his departure: there were fundamental differences in understanding about his relationship with Meta AI head Alex Lebrun and how AI operations should work. He emphasized that when establishing FAIR, his principle was "hire the best people, then get out of the way." He considers this the only route to breakthrough research, and believes it did not persist in his later working environment.
On Ami and world models, LeCun predicts they will become central to future AI, with initial breakthroughs visible within one to two years. He recommended the paper "Le World Model," the latest Ami research on world-model architecture. The name came from team member Randall, and LeCun disclaimed responsibility for it. He expects world models' value to emerge first in robotics, industrial control, and healthcare, becoming a path toward general intelligence as they scale.