X / Twitter
Claude, Anthropic’s Official AI Account
Claude Security has entered public beta, built into the web version of Claude Code and available to Claude Enterprise customers. Users can point it directly at their codebases for security analysis. Since the research preview in February, hundreds of organizations have used it to find real vulnerabilities in production code. Its purpose is to apply Opus 4.7’s security analysis directly to actual code without additional security tools or a separate process.
https://x.com/claudeai/status/2049898741772021991
https://x.com/claudeai/status/2049898745051886013
Sam Altman, OpenAI CEO
Altman announced that GPT-5.5-Cyber is beginning to roll out to critical-infrastructure partners. OpenAI built this model specifically for frontier cybersecurity capabilities. He also announced a major Codex upgrade that day and encouraged using Codex for noncoding computer work, suggesting its positioning is expanding beyond a pure coding tool into broader computer automation.
https://x.com/sama/status/2049712078836170843
https://x.com/sama/status/2049946120441520624
Andrej Karpathy, AI Researcher
Karpathy shared key ideas from his fireside conversation at Sequoia Ascent 2026. He highlighted a line he has frequently quoted recently: “You can outsource your thinking, but you cannot outsource your understanding.” He believes this captures the core of human value in the AI era: understanding must genuinely enter your mind before you can frame problems correctly, catch agents’ errors of judgment, and make tasteful choices. He sees understanding as the true human bottleneck in current workflows and the part least amenable to outsourcing.
https://x.com/karpathy/status/2049903821095354523
https://x.com/karpathy/status/2049907410303865030
Cat Wu, Claude Code Team Member
Claude Security is now in public beta within Claude Code’s web version. Security teams can point Opus 4.7 directly at codebases for scanning, combining GitHub event triggers and scheduled execution to make security review part of continuously automated workflows.
https://x.com/_catwu/status/2049964403177689130
Amjad Masad, Replit CEO
Masad announced a new Replit feature: generating an LLC directly from a prompt. Users can incorporate a company through a prompt, the latest extension of Replit’s AI capabilities into business infrastructure. He also shared its “customer zero” philosophy: Replit inc. itself is the first real customer, with internal tools and processes built entirely using Replit products. This goes beyond simple dogfooding to having the whole team live inside the product, maintaining deep awareness of the user experience.
https://x.com/amasad/status/2049921597499445677
https://x.com/amasad/status/2049934937688854993
Aaron Levie, Box CEO
Levie disclosed that Box has begun hiring and retraining for “agent engineering” roles supporting internal functions, confirming that the position he described in detail last week is becoming a real hiring initiative. He believes agent deployers combining business-process understanding with technical ability will be central talent in enterprise AI transformation.
https://x.com/levie/status/2049714403050918067
Garry Tan, Y Combinator CEO
Garry Tan demonstrated GBrain’s new ability to infer and build knowledge graphs automatically from frontmatter properties. It can infer increasingly many relationships without manual annotation. He also completed and shared an end-to-end installation harness, standardizing GBrain setup in OpenClaw/Hermes environments and lowering the barrier to installation from scratch.
https://x.com/garrytan/status/2049737217208397925
https://x.com/garrytan/status/2049742376307147110
Guillermo Rauch, Vercel CEO
Rauch tested an interesting v0 prompt: “What if Vercel launched GitHub?” In only 2 prompts, v0 generated a complete design. The demo showed its product-design capabilities and invited speculation about Vercel’s future product portfolio.
https://x.com/rauchg/status/2049959307941179678
Ryo Lu, Cursor Designer
Ryo Lu offered a product-design principle: having no position is itself a position. “Opinionated” and “general purpose” are not opposites; truly good products can be both. He also shared integrating Cursor SDK into his operating-system project ryOS, which can now be modified directly through conversation. Integration was extremely simple: he pasted the Cursor SDK documentation link into Cursor, which completed it automatically.
https://x.com/ryolu_/status/2049866003287576978
https://x.com/ryolu_/status/2049872551955013713
Nikunj Kothari, FPV Ventures Partner
Kothari identified a useful trend: the proliferation of MCP and CLI tools supports the conclusion that big models will become the primary interface between people and the world. He sees rapid expansion of this agent-first infrastructure as a clear signal that the era of models as general-purpose interfaces is approaching faster.
https://x.com/nikunj/status/2049871924105531672
Peter Steinberger, OpenClaw Founder
Steinberger shared his experience exploring the AI security ecosystem over recent months, expressing admiration for its complexity and professional depth. He also announced major improvements to OpenClaw group chats: anyone disappointed by previous multi-person conversations should try again, as reliability and feature completeness have improved substantially.
https://x.com/steipete/status/2049976855617314991
https://x.com/steipete/status/2049988836160074022
Dan Shipper, Every CEO
Dan Shipper is experimenting with Codex and Chronicle to build a focus tracker. This is his latest exploration of agents for personal productivity and part of Every’s continuing search for new forms of AI-assisted knowledge work.
https://x.com/danshipper/status/2049913064561258986
Aditya Agarwal, General Partner at South Park Commons
Agarwal offered a judgment on cybersecurity in the AI era: attackers have AI, and defenders do too. Cybersecurity is at a critical inflection point, he believes, with AI fundamentally changing the balance of capabilities. Only AI-based defense against AI threats can maintain equilibrium.
https://x.com/adityaag/status/2049874181509034151
Swyx, AI Podcast Host
Swyx relayed a real ML project story: someone attempted to train a “vintage” language model using public-domain tokens from before 1931. The engineering was meticulous, with specially trained OCR models, an anachronism classifier, and carefully selected hundreds of billions of pre-1931 tokens. Training finished and validation passed. The result: “the most confidently racist model humanity has ever released.” The humorous story reveals a real mechanism of training-data bias: the purer the historical texts, the more concentrated the period’s prejudices, with no way to filter them out.
https://x.com/swyx/status/2049652947408372187
Podcasts
Training Data — Andrej Karpathy: From Vibe Coding to Agentic Engineering
Key takeaway: Vibe coding raises everyone’s floor, while agentic engineering makes engineers tens of times faster without sacrificing professional quality standards. They serve different audiences and solve different problems.
Andrej Karpathy is an OpenAI co-founder, a practitioner who brought Tesla Autopilot into real use, and the person who coined “vibe coding” last year. He described a turning point in December 2025: code generated by the latest models almost no longer needed corrections—he “couldn’t remember the last time I corrected it.” He then explored agent workflows deeply, seeing a qualitative leap rather than simply greater speed.
On the Software 1.0/2.0/3.0 framework: 1.0 means writing code; 2.0 means curating datasets and training neural networks; 3.0 means prompting as programming, with the context window providing control over the LLM “interpreter.” He cited OpenClaw installation: instead of a bash script, users paste text into an agent, which intelligently installs it according to the environment—a typical 3.0 paradigm. More radically, his MenuGen app for menu translation could, in Software 3.0, be replaced by giving Gemini a menu photo and saying “overlay it with Nanobanana,” rendering dish images directly on the original menu. “My entire app shouldn’t exist. It solves the problem using the old paradigm.”
On “jagged intelligence”: Opus 4.7 can refactor a 100,000-line codebase and find zero-days, yet suggest “walking to the car wash 50 meters away,” even though washing a car requires driving it there. This unevenness comes from training RL environments: reinforced and unreinforced domains create jagged capabilities. His advice is that building an RL environment in a verifiable domain can make your own fine-tuning significantly effective even if major labs have not focused on it.
On vibe coding versus agentic engineering: vibe coding raises the floor for everyone, letting anyone build quickly. Agentic engineering accelerates professionals without lowering security, code-quality, or other standards. He estimates that those truly proficient at it gain far more than 10-fold speed. For hiring, most companies still use the old programming-puzzle paradigm. A new approach would give candidates a real large project, then have 10 Codex instances attack their deployed system to see whether it holds up.
On what remains human: “You can outsource thinking, but not understanding.” API parameter details can be forgotten because agents remember them, but underlying logic—such as efficiency differences between memory views and copies—must be understood to catch alarming agent mistakes, such as linking accounts by email instead of user ID.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8
Training Data — Demis Hassabis on Building DeepMind, AlphaFold, and the Final Stretch to AGI
Key takeaway: Demis Hassabis expects AGI before 2030, matching his 2010 prediction of a “20-year mission.” He considers machine learning the perfect language for biology, as mathematics is for physics.
Demis Hassabis is DeepMind’s co-founder, Google DeepMind CEO, and also CEO of AI drug-discovery company Isomorphic Labs. At a Sequoia event, he reviewed his journey from planning for AGI as a teenager to today’s “final stretch.”
On early intellectual foundations: at 15 or 16 he decided AGI mattered most, and subsequent game development—including Bullfrog’s Theme Park—and a neuroscience PhD served that goal. Theme Park sold over 10 million copies, with an economic AI model driving thousands of virtual people. Seeing players’ delight interacting with AI strengthened his conviction. At Elixir Studios he attempted to simulate an entire country in Republic, modeling a million people on a Pentium CPU; being too early hurt the project. The lesson: “Be five years ahead, not fifty.”
On founding DeepMind in 2009: they felt “like they were keeping a secret.” Geoff Hinton and others had just invented deep learning, reinforcement and deep learning had not yet been combined, GPU compute was beginning, and computational neuroscience supplied ideas. Academics even openly mocked AGI researchers, but Hassabis thought: “At least we’ll fail differently from the nineties—that alone makes it worth doing.”
On AI for Science: the day after AlphaGo won its Go match, they formally launched the AI for Science group. AlphaFold solved a 58-year challenge, he believes, but drug discovery is multistep. Isomorphic Labs aims to automate compound design: finding molecules that bind precisely to a target protein’s specific site without binding other proteins and causing toxic side effects. He predicts this will compress average discovery timelines from 10 years to months or even weeks, enabling personalized medicine.
On machine learning and biology: machine learning is the perfect language for biology, just as mathematics is for physics. Biology contains weak signals, enormous data, and highly emergent behavior that traditional mathematics cannot fully describe—precisely where machine learning excels. He is researching virtual cells, hoping to build precise biological models analogous to physics simulations.
On the period after AGI: he recommends first building AGI as an “extremely intelligent tool,” then addressing deeper questions of consciousness and autonomy, using the tool itself to help answer them. His favorite book is David Deutsch’s The Fabric of Reality, and his post-AGI mission is to use AGI to answer its questions.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8