X / Twitter
Anthropic’s official AI account Claude
Claude Code released a major update: auto mode removes the need to approve every file write and bash command individually or skip all permission checks. Claude determines which actions can proceed automatically. Before every tool call, a classifier performs a safety review: safe actions proceed, while risky ones are blocked for human confirmation. The feature is currently a research preview on Teams plans, with Enterprise and API access rolling out over the coming days. Enable it with `claude --enable-auto-mode`.
https://x.com/claudeai/status/2036503582166393240
https://x.com/claudeai/status/2036503583667933400
https://x.com/claudeai/status/2036503584943034643
Anthropic researcher Alex Albert
Albert strongly recommended a guest article by physicist Matthew Schwartz describing how AI helps scientific research, noting Claude Opus 4.5’s surprising capabilities on certain tasks. He also announced that Claude Code was officially deprecating `--dangerously-skip-permissions` in favor of auto mode, an important safety-mechanism upgrade.
https://x.com/alexalbert__/status/2036232980059062550
https://x.com/alexalbert__/status/2036510206155432293
Claude Code engineer Thariq
Thariq revealed a little-known background detail: Anthropic acquired a computer-use company and released computer use four weeks later, an impressive pace. He said being an AI safety company is especially useful when ensuring AI operates safely. He also announced a March 31 livestream with Figma to share how to use Claude’s latest capabilities. The feature is currently available only on Teams plans, with work underway to scale access.
https://x.com/trq212/status/2036302910498173034
https://x.com/trq212/status/2036513038983995820
https://x.com/trq212/status/2036442894777594248
OpenAI CEO Sam Altman
Altman offered a broad outlook on AI’s future impact, arguing that helping discover new science and conquer diseases is one of its most important paths to improving human life. He also sought a concise term for “throwing every piece of context you can think of into a model,” sparking broad discussion, and supplied further details about OpenAI’s latest progress.
https://x.com/sama/status/2036488680769241223
https://x.com/sama/status/2036489823792607273
Replit CEO Amjad Masad
Masad demonstrated a TSA wait-time app built on Replit that quickly went viral. He also shared creator Nick’s story of how learning to write software transformed his content creation, describing it as an authentic firsthand account rather than marketing.
https://x.com/amasad/status/2036272008158433564
https://x.com/amasad/status/2036427267296067875
Box CEO Aaron Levie
Levie argued that computer use and the ability to write and execute code in real time are the ultimate infrastructure for agents to handle more knowledge work. Quoting a description of enterprises moving away from legacy systems, he emphasized that software unable to run headlessly will become the biggest bottleneck in the agent era, while APIs that can be called automatically will be a core competitive advantage.
https://x.com/levie/status/2036290915950797314
https://x.com/levie/status/2036478998667207012
Vercel CEO Guillermo Rauch
Rauch announced that almost all SaaS applications used internally at Vercel had been replaced by generated apps or agent interfaces, spanning support, sales, marketing, PM, and other departments, all deployed on Vercel. He also announced that Vercel can now intelligently select the right hardware for build tasks, substantially improving build performance alongside the new Rust compilers Turbopack and Rolldown.
https://x.com/rauchg/status/2036447879985037495
https://x.com/rauchg/status/2036584870995173678
Andrej Karpathy
Karpathy warned of a serious software supply-chain incident: malicious code had been injected into the litellm PyPI package, and simply running `pip install litellm` would result in theft of SSH keys, AWS, GCP, and Azure credentials, and Kubernetes configurations. The warning attracted wide attention and more than twenty-one thousand reposts.
https://x.com/karpathy/status/2036487306585268612
Google Labs VP Josh Woodward
Woodward called the current period inside Google the most exciting of his career, saying the company was moving fast, and publicly announced substantial hiring by the Gemini and Google AI Studio teams.
https://x.com/joshwoodward/status/2036513009661780291
Linear Head of Product Nan Yu
Yu shared a personal experience: for months he has not manually written a PRD, filed an issue, or written a line of code, yet both the quantity and quality of his output have increased. The account illustrates AI-assisted workflows in actual product management practice.
https://x.com/thenanyu/status/2033545747686166887
Y Combinator CEO Garry Tan
Tan described his first PR to YC’s internal repository, which contains 1.84 million lines of code; his PR comprised 2400 lines. He said GStack is not only for independent developers building greenfield projects. That evening’s experience showed its skills can also transfer to large existing codebases.
https://x.com/garrytan/status/2036330296308802012
https://x.com/garrytan/status/2036330844516917297
OpenClaw founder Peter Steinberger
Steinberger released CodexBar 0.19.0 with Alibaba Coding Plan support, subscription history charts, and dashboards for Cursor Total, Auto, and API, helping users understand token usage patterns. He also shared his full PR review process: have Codex analyze issues, determine whether the problem is clear, and assess whether the solution is optimal. He demonstrated OpenClaw acting as a “gatekeeper” for WhatsApp Business, recommending a separate number and WA Business to avoid accidental actions.
https://x.com/steipete/status/2036245531522113910
https://x.com/steipete/status/2036252448399171717
https://x.com/steipete/status/2036486352964035013
Swyx
Swyx analyzed OpenAI’s tighter control of internal “Side Quest” projects, identifying Sora as the first product affected. He also praised Devin, saying it saves him 3 to 8 times the workload each day. Although “Devin finding Devin’s own mistakes” sounds odd, he finds it effective.
https://x.com/swyx/status/2036533647659143630
https://x.com/swyx/status/2036565584515899445
Zara Zhang
Zhang announced two new Frontend-slides capabilities: deploying a deck to a shareable URL and exporting to PDF, making vibe-coded slides easy to share. She also reflected on her past year: from barely understanding GitHub to accumulating more than 13000 stars, despite lacking a technical background.
https://x.com/zarazhangrui/status/2036552689673445847
https://x.com/zarazhangrui/status/2036590721927618988
FPV Ventures Partner Nikunj Kothari
Kothari crawled and analyzed SOC II subprocessor lists from 417 companies. He concluded their real value lies not in security compliance itself but in revealing what infrastructure companies actually use. The lists expose technology stacks at firms such as Notion, Ramp, Brex, Harvey, and Perplexity, making them valuable for VC competitive analysis.
https://x.com/nikunj/status/2036572222081606065
https://x.com/nikunj/status/2036573898289086895
Podcasts
Latent Space — 🔬There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik
MIT chemical engineering professor Heather Kulik has long studied the intersection of AI and materials science. Her central assessment is that materials science currently lacks, and is unlikely to see, an AlphaFold-equivalent breakthrough, because the complexity of changing chemical bonds across materials space far exceeds that of protein structure prediction.
Kulik shared a specific AI-assisted discovery: her team screened tens of thousands of materials with AI, whereas traditional experiments take months or even years per attempt. AI identified a previously unknown quantum-mechanical phenomenon that increased the toughness of plastics in polymer networks by about fourfold. The experimental team said they would never have conceived the AI-proposed design direction themselves, and ultimately fabricated and validated it in the lab.
Asked the challenging question of whether studying chemistry is still necessary, Kulik gave a counterintuitive answer: LLMs’ chemistry knowledge is largely at Wikipedia level. Whenever a major model is updated, she asks it to design a ligand containing 22 atoms that coordinates with a metal through two nitrogen atoms; none has yet produced the correct atom count. Her conclusion is that LLMs can quickly introduce unfamiliar fields, but users need enough foundational knowledge to recognize when they are confidently wrong.
In active learning, Kulik’s team currently optimizes seven objective functions simultaneously in research on direct CO2 capture using metal-organic frameworks, including cost, humidity stability, CO2 selectivity, mechanical stability, and thermal stability. Even with imperfect machine-learning accuracy, each added optimization dimension can yield a 100- to 1000-fold speedup, which is active learning’s core value for high-dimensional problems.
On materials science’s lack of a “CASP moment,” Kulik noted that datasets such as Materials Project and Open Catalyst Project exist, but their data comes from imperfectly accurate density functional theory calculations rather than experimental ground truth. Unlike CASP’s experimentally validated structures, materials science lacks high-quality experimental benchmarks, a pressing infrastructure problem for the field.
She also described the “impressive exterior, fragile interior” of machine-learned interatomic potentials. Some foundational potentials claiming to replace first-principles calculations cause molecules to inexplicably fall apart in practice, while actual speedups can be only fivefold, far below promised levels. She called for stricter evaluation standards to determine whether such models can truly replace physics-based modeling.
Kulik developed and open-sourced molSimplify, called MOFSimplify when handling metal-organic frameworks, for generating transition-metal complex structures and screening metal-organic frameworks. It has a web interface and is available on GitHub, and she welcomes engineers to try it and provide feedback.
https://youtube.com/watch?v=KSCCKCz2x04
Training Data — Biology's Waymo Moment: Ginkgo Bioworks' Jason Kelly
Ginkgo Bioworks founder and CEO Jason Kelly made a forceful assessment: previous internet, social media, and mobile revolutions had essentially no effect on biotechnology or biopharmaceuticals beyond improving some back-office IT. AI is truly different, fundamentally changing how science is done and bringing real disruption to biopharma.
Ginkgo was founded in 2008 and did not receive its first outside funding until 2014. For six years it struggled to survive on government grants and service work. The turning point was a 2014 blog post by Sam Altman, then newly in charge of YC, arguing that Silicon Valley’s model could apply to deep-tech fields such as nuclear energy, biotechnology, and materials science. Kelly wrote to Altman and was eventually persuaded to join YC, which became Ginkgo’s true starting point.
Kelly defines Ginkgo’s core mission as “making biology easier to engineer.” The mission has never changed, though the product path took many detours. Its original business model was B2B biotechnology services, providing strain engineering to large chemical and agricultural companies and gradually developing the capabilities of a biological foundry.
On AI at Ginkgo, Kelly’s central view is that biology is experiencing a Waymo-like moment: foundational capabilities have arrived over recent years, and combining software with automation will fundamentally accelerate laboratory work previously requiring extensive manual effort. He stressed that robotic hardware is ready; software reliability and scaling are the current bottlenecks.
He said Ginkgo has built large-scale automated experimental platforms, aiming to create rapid iteration loops between computational predictions and experimental validation, reducing manual intervention while preserving human judgment in unexpected situations. Fully automated high-throughput labs and manual work each suit different scenarios, and he believes their combination is currently more effective than pure automation.
Kelly also discussed Ginkgo’s strategic shift in the AI era from a service company to a data and AI platform. Its core assets are the vast biological experimental data accumulated across hundreds of customer projects, unique raw material for training biology AI models. He sees this as the key competitive asset for future biological AI companies against traditional drugmakers.