Andrej Karpathy (Former Director of AI @ Tesla, Founding Team @ OpenAI) warns about the litellm PyPI supply chain attack. A simple pip install litellm was enough to exfiltrate SSH keys, AWS/GCP/Azure credentials, Kubernetes configs, git credentials, environment variables (including all API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, and database passwords. The poisoned version was up for less than an hour and was only discovered because it had a bug that caused a crash. Karpathy argues this is why he’s “growingly averse to dependencies” and prefers using LLMs to “yoink” functionality when simple enough. https://x.com/karpathy/status/2036487306585268612
Vercel CEO Guillermo Rauch announced elastic build machines that intelligently pick the right hardware for builds. With new Rust-based compilers like Turbopack and Rolldown, build performance now scales with O(cpus). The key insight: “Too many CPUs and you waste money. Too few and agents waste time.” He also shared that “almost every SaaS app inside Vercel has now been replaced with a generated app or agent interface” — from support to sales, marketing, PM, HR, dataviz, even design and video workflows. The SaaSpocalypse is “understated because the software we are generating is more beautiful, personalized, and crucially, fits our business problems better.” https://x.com/rauchg/status/2036584870995173678 https://x.com/rauchg/status/2036447879985037495
Box CEO Aaron Levie responded to Rauch’s post with a critical insight for platform builders: “If you’re building software that can’t work fully headlessly in a way that agents want to use, you’re not prepared for what the future of software is going to look like. Agents will use software 100X more than people, and people will more and more interact with their data and workflows via agents across many different platforms. Software doesn’t go away, but it becomes the guardrails and business logic for what agents are able to operate on. But if you can’t connect to wherever the agents want to do that work, you’re DOA.” https://x.com/levie/status/2036478998667207012
Anthropic’s Claude launched auto mode in Claude Code. Instead of approving every file write and bash command, or skipping permissions entirely, auto mode lets Claude make permission decisions on its own behalf. “Before each tool call, a classifier reviews it for potentially destructive actions. Safe actions proceed automatically. Risky ones get blocked, and Claude takes a different approach.” This replaces the deprecated --dangerously-skip-permissions flag. https://x.com/claudeai/status/2036503582166393240 https://x.com/claudeai/status/2036503583667933400
OpenAI CEO Sam Altman announced the OpenAI Foundation with at least $1 billion planned for its first year. The Foundation will focus on: AI discovering new science and cures for diseases, and addressing new threats to society like novel bio threats, massive economic changes, and extremely capable models causing complex emergent effects. Cofounder Wojciech Zaremba transitions to Head of AI Resilience. https://x.com/sama/status/2036488680769241223
Y Combinator President Garry Tan pushed back on “replace developers” narratives: “Once you say ‘replace developers’ it’s the wrong ideaspace. Cars didn’t replace walking, you still have to walk a little bit AND it let people go to places you can’t go to on foot alone.” He also predicted AI will deliver “90x my 2013 output.” https://x.com/garrytan/status/2036674371855823153 https://x.com/garrytan/status/2036675789819953347
Linear Head of Product Nan Yu shared that they haven’t “written a PRD by hand, filed an issue through a form, or hand-written any code in months. But the volume of work I’m producing and the quality bar have never been higher.” https://x.com/thenanyu/status/2036549647267709110
Roblox Product Peter Yang hopes AI agents will help teams “stop obsessing about: Specs, Designs, Issues, Roadmaps, Strategy docs, Spreadsheets” and instead “obsess about improving the end user product and rapid iteration loops with users.” https://x.com/petergyang/status/2036621615086309468
FPV Ventures Partner Nikunj Kothari launched DeployGraph, scraping 417 companies’ subprocessors to investigate what AI native companies use for infrastructure. “SOC II is in the news right now for being security theater. You know what SOC II is actually good for? Subprocessor lists.” Built using Conductor, Railway, and Claude Code. https://x.com/nikunj/status/2036572222081606065
Anthropic Research Alex Albert celebrated the Claude Code auto mode launch: “Goodbye –dangerously-skip-permissions, hello auto mode.” https://x.com/alexalbert__/status/2036510206155432293
Anthropic’s Thariq (Claude Code team) noted that “being an AI safety company is useful for when you need to make sure AIs can run safely” and announced a livestream with Figma on March 31st covering Claude Code integration. https://x.com/trq212/status/2036513038983995820 https://x.com/trq212/status/2036442894777594248
Google VP Josh Woodward says “there’s never been a more exciting time” at Google, calling the place “pulsating” and noting they’re hiring for Gemini and Google Labs. https://x.com/joshwoodward/status/2036513009661780291
Swyx praised Devin’s “smart friend” pattern: “this thing legitimately saves my ass 3-8x a day… this is basically the equivalent of ‘sleeping on it’ and looking at a PR with fresh/more critical eyes.” He also noted “Sora is dead” as the first casualty of OpenAI’s crackdown on Side Quests. https://x.com/swyx/status/2036565584515899445 https://x.com/swyx/status/2036533647659143630
Zara Zhang (builder, formerly Harvard) shared: “Last year I barely knew how GitHub worked; now I have 13k+ stars on GitHub (and I’m not even a technical person, but then again how do you define technical these days?)” Also announced frontend-slides now supports deploying as URL and PDF. https://x.com/zarazhangrui/status/2036590721927618988 https://x.com/zarazhangrui/status/2036552689673445847
OpenClaw Founder Peter Steinberger got greeted by an OpenClaw “doorman” when messaging a friend on WhatsApp, demonstrating the challenges of using consumer messaging platforms for AI agents. https://x.com/steipete/status/2036486352964035013
Every CEO Dan Shipper shared new Every stickers for Mac Mini running OpenClaw. https://x.com/danshipper/status/2036501111394148578
Anthropic’s Cat Wu asked what people want to hear at Code with Claude this year. https://x.com/_catwu/status/2036594646370210229
Replit CEO Amjad Masad shared a video about how learning to make software transformed a content creator’s business. https://x.com/amasad/status/2036427267296067875
VC Matt Turck observed the irony: “Typo, before: you did sloppy work! That’s bad. Typo, today: you wrote authentic, non-AI generated content! That’s awesome.” https://x.com/mattturck/status/2036488357077717094
Ex-Dropbox CTO Aditya Agarwal on robotics: “It is increasingly clear that the hardware for robotics is ready. We just need to get the software to a point where it is reliable and scalable. This actually makes me very bullish on the US. We are good at software.” https://x.com/adityaag/status/2036426693930291334
Latent Space: “There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik”
The Takeaway: AI is transforming materials discovery, but unlike protein folding, there’s no single breakthrough waiting to happen — the chemical space is simply too vast and diverse for one model to crack it.
MIT Professor Heather Kulik has spent nearly two decades applying machine learning to materials science. Her most striking achievement: using AI to design a polymer network that’s four times tougher than conventional materials — a discovery that surprised even experienced experimentalists who “would have never come up with this on their own.” The key insight wasn’t just faster calculations, but uncovering “unexpected chemical phenomenon” through quantum mechanical effects that no human researcher would have predicted.
Kulik is bullish on active learning for solving multi-objective problems — like finding materials for carbon capture that balance cost, stability, CO2 uptake, mechanical strength, and thermal properties. “The real promise is going to be in searching for that needle in a haystack with say seven objectives.”
On whether chemists are obsolete: “ChatGPT is super good at Wikipedia level chemistry knowledge” but fails at practical molecular design tasks like “design me a ligand that has 22 atoms.” Her test: ask an LLM to design a specific molecule. “I can never get an answer that has 22 atoms. So that’s something that an expert chemist could do in a second.”
The biggest ML gaps in chemistry: reactivity prediction, diverse chemical bonding (especially transition metals), and excited states. “We have really good data sets out there for really boring chemistry” — organic molecules binding to proteins — “but there’s lots of challenges where the physics is much more complex.”
She recommends her open-source tool MOLE-Bridge for anyone interested in transition metal complex design. https://youtube.com/watch?v=KSCCKCz2x04