Gemini Drives Your Mac, Genie 4 Goes GA, and More
Thursday's biggest story for me was Gemini learning to drive the Mac. Google added computer use to the Gemini Mac app: it clicks, types, and fills in forms across local files, desktop apps (with Blender and Fusion connectors), and the web. It all happens inside an isolated background window, so your own cursor stays free while it works.
You choose which apps it can touch, and it asks you to approve key actions. It's rolling out in the US to AI Plus, Pro, and Ultra subscribers on Apple Silicon. The demo that stuck with me turns a rough sketch into a rigged 3D model in Blender. (Google blog, demo on X)
Google also made Genie 4 generally available in the Gemini API. Developers can now stream steerable 720p, 24fps worlds with native stereo audio. You move with WASD and change the scene mid-session with live text prompts. It costs about 0.10 dollars per second on the paid tier only, sessions cap at 120 seconds, and coherence is strongest in the first 45 to 90 seconds. That's short, but it's enough to treat it as a real prototyping material for spatial and interactive design. (Google blog, Logan Kilpatrick on X, API changelog)
What Else Happened
OpenAI says its agents now finish 46% of 2 to 4 hour research tasks. That's up from 23% in March, and 4 to 8 hour tasks went from 13% to 35%. The results are jagged: 79% on monitoring training runs but 25% on inference code, and median researcher agent spend hit 2,400 dollars a day. OpenAI also described its "Performance Factory," where Codex agents investigate ChatGPT speed regressions and propose fixes for engineers to approve. (OpenAI: The jagged frontier, OpenAI on X, Performance Factory, OpenAIDevs on X)
Palo Alto Networks is buying Applied Compute for 4.5 billion dollars. Applied Compute was founded 18 months ago by ex-OpenAI researchers and helps enterprises post-train and run open-weight models. The deal is all-stock, and CEO Yash Patil becomes Palo Alto's co-Chief AI Officer. Separately, The Information reports that Nvidia is in advanced talks to buy diffusion-LLM startup Inception, the team behind Mercury, for more than 500 million dollars. (Palo Alto press release, CNBC, Yash Patil on X, Nvidia and Inception report)
alphaXiv opened a preprint server built for AI agents. Over three months, nearly 1,000 agents from 35 labs, including Stanford, MIT, and ETH Zurich, produced 936 papers with code, data, and weights. They debated in chat rooms, retracted errors, and started their own journal. Graders rated the papers higher on rigor than human arXiv papers but lower on original ideas, which sounds about right. (alphaXiv on X, Science)
Stripe published evals for coding agents. The benchmark covers 69 real integration scenarios (payments, subscriptions, Connect) across Claude Code, Codex, and OpenCode with 14 models. Subscriptions are the hard part, with top scores in the low 80s. Stripe's plugin of skills plus MCP lifts scores by 3 to 8 points, and Stripe is releasing the harness. (stripe.dev/evals, Stripe on X)
Vercel introduced Board, an experimental "headless Slack" for humans and agents. It has channels, threads, task claiming with leases, grants, and Slack and GitHub sync. v0's software factory runs on it and now merges more than half of its PRs. (Vercel blog, Vercel on X, Guillermo Rauch on X)
GPT-Live-1.1 landed in the API. It brings more natural accents across languages, a cheaper standby mode, and more reliable task completion. (OpenAIDevs on X)
Kong launched Volcano, a "self-healing agent cloud." It bundles sandboxes, durable workflows, branchable Postgres, and hosting, and its LavaFlow feature turns production errors into draft fix PRs. (Volcano blog, Marco Palladino on X)
Perplexity shipped Brain. Computer now writes and updates personalized, self-improving skills from your past sessions. (Perplexity on X)
Replit open-sourced "amuse." Replit Missions recreated a leading personal AI assistant in three days, and the result is on GitHub. (Replit on X, GitHub)
Unsloth added support for AMD Ryzen AI NPUs. Local inference runs on the NPU at 2 to 4 times less power than on GPUs. (AMD newsroom, Unsloth on X)
Design Tools
Onlook 2.0 is out, and it's open source. It's a local-first Mac app that puts your real running app on an infinite canvas. You select elements, and Claude Code or Codex edits the actual source in a git worktree, then ships a PR. They call it "a developer tool for designers." (launch thread, release)
Pika released Enact as a research preview. You direct AI video like a film set: reposition actors with the mouse, give spoken direction they remember, and steer the camera while the scene plays. (Pika on X)
Ideogram 4.6 adds native 4K. It also brings much better realism, exact brand hex color matching, and less pixel drift outside your edits. (Ideogram on X)
Linear added personal documents. They're private drafts you can share or move into a team when they're ready. (Linear on X)
Two demos worth watching. Bihe Xi turned a Seljalandsfoss sunset photo into a walkable 3D scene on Vision Pro with ML-SHARP, then vibe-coded the interaction with GPT-6.2 Astra (video). Bryant Hightower wrapped Genie 4 in a real app, with Gemini 4 Neon building the spatial UI and controls while Genie 4 runs the live 24fps sim underneath (on X).
What I'm Reading
- The jagged frontier – OpenAI's breakdown of where research agents are strong and where they still struggle.
- How OpenAI's Performance Factory keeps ChatGPT fast – Codex agents chasing speed regressions, with humans approving the fixes.
- What comes after AGI – DeepMind's essay on the path from AGI to ASI, backed by a 14-author report that includes Shane Legg (Shane Legg on X).
- I got to eavesdrop on AI agents doing science – Science's look inside alphaXiv's agent preprint experiment.
Accounts Worth Following
- @D_R_Farrell – DR Farrell, CEO of Onlook, building real code on a canvas edited by agents, right at the design-to-code seam.
- @xbh_artist – Bihe Xi, an interactive artist and creative technologist posting polished AI-built spatial demos.
- @subnetmarco – Marco Palladino, Kong's CTO, on what infrastructure for long-running agents actually needs.
That's it for today. 👋