October 13, 2026 · 2:36 PM PT edition
Mind Blow: 20 Chromium changes at once, a robot that sinks cornhole and a master key found by AI
Gemini 4 Argon goes to work on code, hospitals, patents and a 3D cathedral.
A Chromium engineer says that with Gemini 4 Argon he has landed more changes into Chromium in two weeks than he would normally manage in months. Meanwhile a humanoid robot sank seven cornhole shots straight in the hole after learning from one video of a human.
Here is what happened in AI and DevOps this cycle.
The Chromium team is using Gemini 4 Argon to untangle a codebase of about 40 million lines built up over two decades. Earlier models would go off the rails in a repo that size, but now, with no skills, no harness and just a memory file, it spins up 20 changes at a time. Each change runs a couple hundred thousand tests, then two human reviewers, then a couple million more tests, and the changes pass. The team says a typical engineer ships one change a day, so 20 changes is 20 days of work, and modularizing Chromium no longer has to take years.
Security firm Wiz says Gemini 4, in shadow A/B testing on real systems, found over two times more vulnerabilities than baseline models while using fewer tokens. In a Hack for Good scan of a healthcare app, it bypassed a public file download feature, found a config file holding the application’s master key, and turned it into administrator access using only read-only operations. Wiz says attackers could have controlled the systems without the operators knowing; it was fixed, and Wiz argues defenders can now outpace attackers.
Google DeepMind researchers showed Gemini 4 Argon one video of a person playing cornhole and let it practice in the MuJoCo simulator, learning from each failed throw. It mapped the human arm swing onto robot joints, measured the board in 3D with its two eye cameras and built a physics simulation of its own body. On the real robot it adjusted its release by a few centimeters per throw, which shows video understanding turning straight into motor skill.
At Isomorphic Labs, a chemist gave Gemini 4 a set of patents that can run 500 pages, with a structure sitting 50 pages from its experimental readout. It extracted them into a spreadsheet and a full report that she found highly accurate, turning nearly a month of work on 10 patents into an afternoon. A colleague says a paper summary that would have taken one or two days came back in 10 to 15 minutes. The team’s pitch is better drugs reaching patients faster.
A Google for Developers creator connected Antigravity to Blender over MCP and handed Gemini 4 Argon a single 2D image of a cathedral. Over 45 rounds the model rebuilt the base mesh, with an agent laying comparison strips over the reference after each pass. Then micro-loops added column fluting, and the model dialed in volumetric fog and audio-reactive sunlight, ending in a full 3D walkthrough.
Linear’s CI playbook for GitHub Actions, as summarized by Yonas Beshawred, starts with moving CI off GitHub’s runners onto faster machines. Jobs ran 34% faster on average and typecheck 52% faster. The next step swapped tsc for tsgo, the native TypeScript compiler. Slow CI is a tax on every pull request, and this is a cheap place to cut it.
Tatsunori Hashimoto announced Deep Learning Alchemy, a new class that treats deep learning as messy, empirical frontier work rather than a sanitized set of facts and math. It follows the hands-on spirit of his Language Modeling From Scratch course. The bet is that the craft of making models work can be taught, not just picked up.
The bill for AI’s buildings keeps climbing. Cushman & Wakefield’s 2026 guide puts modern greenfield data centers at an average of $17.6 million per megawatt, before a single chip or GPU. Construction costs per MW are up 21% since Q4 2024, a number every capacity plan now has to absorb.
Akhilesh Mishra’s Kubernetes troubleshooting guide for beginners opens with one rule: it is not about knowing commands, it is about knowing where to look. His first step, he says, is to stop assuming, a habit that will outlast any one cluster.
That’s the cycle. Go throw a bean bag, shave a minute off your CI, and see you next edition.
Also on the radar: these also appeared this cycle, but there was no transcript to report from yet (captions rate-limited by YouTube, no English captions, or no speech), so they aren’t summarized here: Can Infrastructure Keep Pace With AI? (Bloomberg Live); How to Create an Oracle AI Data Platform Instance Using Codex (Oracle); How to onboard to Looker using your favorite LLM and skills (Google Cloud Tech); What Technical Due Diligence Finds When the Deck Says “70% Automated” — Soukaïna Abelhad (AI Engineer); Tokens Are the Cheap Part — Steve Jamison, Jellyfish (AI Engineer); The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale (AI Engineer); Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers (AI Engineer); Every Building Is a Balance Sheet: Spatial Memory for Real-World Assets — Jordi Spranger, Spranic (AI Engineer); Work with your team in ChatGPT Space (OpenAI); Block on Finding and Fixing Security Risks with OpenAI (OpenAI); I Made My AI Agents Faster Without a GPU Upgrade (Alex Ziskind); Why is my AMI or EBS snapshot creation slow or stuck in the Pending state? (Amazon Web Services); Gemini 4 Argon Puts Google Back On the Frontier (Every).