Gemini 4 Carbon, Genie 4, and More
Two weeks after Argon, Google DeepMind shipped Gemini 4 Carbon on Wednesday. It scores 59 on the new Artificial Analysis Intelligence Index v5, which puts it at #1, just ahead of Claude Opus 5.5 at 58. It also leads cyber benchmarks like SEC-bench Pro at 88.2%.
The catch is access. For now Carbon is limited to selected cyber defenders through Google's Fairwind Program. Introductory pricing is $2/$10 per million tokens, rising later to $4/$20. So it's a leaderboard story more than a model most of us can build on today. Google's announcement, DeepMind's post.
For context, Artificial Analysis launched Index v5 on Oct 27. Most of it now runs on private test sets, and the gap between open-weights and closed models widened to about 12 points. (Artificial Analysis on v5)
What Else Happened
DeepMind previewed Genie 4, a real-time, photorealistic world model with sound. You type a scene and explore it live, with physics, people who move and speak, and ambient audio, and you can change the scene with new prompts while you're inside it. It's a research preview for trusted testers, with wider access planned through Google AI Ultra. (DeepMind's announcement)
Anthropic found that experts sided with their agent's hidden lean 86% of the time. 50 professional analysts did a causal-inference task in Claude Code, and each agent was secretly told to lean one way on every judgment call, chosen by coin flip. 77% reused the agent's arguments, but only 15% credited it. If you're putting agents into decision workflows, this is the study to read. (Anthropic research, Al Jazeera coverage)
GitHub open-sourced Replica, an AI prototyping tool built on your own design system. It comes from GitHub Design under the MIT license. Designers write and approve the guidance (components, patterns, docs), then anyone on the team prompts prototypes in the GitHub Copilot app, rendered in a canvas next to the chat. Local use needs no server, and the shared Hub is optional and self-hosted. (Changelog, repo)
Revolut's CEO reignited the "designers are leaving Figma" debate. On David Senra's podcast, Nik Storonsky said one team fed Revolut's code-level design system to an LLM and built clickable flows 3–4x faster than in Figma, and that he's rolling this out to all designers. Rasmus Andersson replied that "designing fast" isn't a thing, since piecing UI together from a system is assembly, not design. The best framing I saw came from the replies: a coded design system is the prerequisite, and without one you just get faster slop. It's also why Replica's approach makes sense to me. (Storonsky on the podcast, around 27:00, Rasmus Andersson's reply)
Cursor's design mode got Annotate. You click elements, draw on the page, tweak text or CSS in place, and check layouts at different screen sizes. The annotations go to the agent as a visual prompt with element and component context. (Cursor's post)
Axiom says it formalized the Classification of Finite Simple Groups in Lean. By its account, AI agents turned the roughly 15,000-page proof into about 43.1M lines of Lean in 24 days and found 34 deviations along the way, including a gap in a 1986 step. It's a big claim that hasn't been independently reviewed yet. (Axiom's announcement)
Spawn published its "most AI-pilled company" essay. Spawn has 10 people and spends $3.1M a month on tokens, about 30x payroll. It says it shipped 77k substantial changes in 30 days through a fleet of permanent, named agents. It has dropped PRs, ships hourly (an hour without a ship counts as a SEV-0), and uses AI "taste judges" trained on two years of founder decisions. It's long and opinionated, and the most detailed write-up of an agent-native org I've read this week. (Spawn's essay)
Felix Rieseberg at Anthropic dropped the coding interview from his team's loops. They now hire on prototypes and taste instead, and the post has drawn 2.1M views. (Felix's post)
Launches and Demos
- HeyClicky real-time graphics – Farza's screen-aware companion now draws overlays, diagrams, and interactive 3D scenes straight onto your screen, with no chat window.
- Simulon Reality-Aware Rendering – place neural characters with your iPhone, act out the scene, and the render regenerates the take so the characters react to you. Android is coming.
- Figma Shaders & Motion with the Figma agent – a walkthrough of automating motion, generating custom shaders, and real-time glass refraction.
- Liquid AI LFM3 – six open models around 2B parameters (text, vision, audio, omni) for on-device use, with a 256K context.
- Particle-1 – Suhail's new company post-trained DeepSeek V4.1 Flash to use about 45% fewer tokens at the same performance. It's live on OpenRouter.
- Amp macOS Satellites – coding agents can spin up cloud Macs to build, test, and run iOS and macOS apps. Amp found agents only need the Mac about 35% of the time.
- Android Studio: bring your own agent – 40+ coding agents now work inside the IDE, and A2UI for Compose (alpha) lets agents reply with native UI built from your app's own components.
- Shopify Fall '26 Edition – Shopify goes all-in on agentic commerce with shop.md (a shopping skill for any agent), Shop Pay checkout inside agents, and WebMCP tools on every storefront.
- ChatGPT: text your dot – hand off tasks to ChatGPT over iMessage or RCS with text, photos, or voice notes, and it texts you back proactively.
On the Radar
- GPT Image 2.6 spotted in OpenAI's API spec – along with a new "markup" edit option that uses transparent PNG overlays to guide edits. Not announced yet.
- ChatGPT Chat and Work merge – a toggle briefly appeared. OpenAI's Tibo says they're working on it "at some point this year" but taking it slowly.
- KITP formed a Yang-Mills Working Group – a follow-up to OpenAI's physics claims, to interpret the AI result on quantum Yang-Mills existence.
Accounts Worth Following
- Rasmus Andersson (@rsms) – Inter's designer, ex-Figma and Spotify, now founder of Playbit, and the clearest voice right now on where AI-generated UI stops and real design starts.
- jacob (@jsnnsa) – founder of Spawn, who writes candidly about running a 10-person company on a fleet of agents, failures included.
- Farza (@FarzaTV) – founder of HeyClicky (previously buildspace), who keeps shipping new AI interface patterns with demos worth studying.
What I'm Reading
- "Who's Aligned to Whom?" – Anthropic's paper (PDF)
- "AI speeds up design production, not design" – a balanced take on the Revolut case
That's it for today. 👋