Overview
Today’s feed was about agency and reach - robots and agents adapting to the real world, chips and phones putting AI closer to the edge, and design tools that cut context switches for people doing the work. We also saw models writing games and engines, fresh debates in autonomy and mobile distribution, and a dash of science and nostalgia to keep our sense of scale in check.
The big picture
Robots and mobile agents are inching toward broad competence, helped by smarter models and practical tricks that run on real hardware.
DeepMind’s robot adapts to a squishy stress ball
Hannah Fry hands a robot a stress ball it has never seen, the system feels it out, then figures out how to place it into an odd pear-shaped container. DeepMind’s setup mixes planning with an inner monologue and a vision-language-action policy for execution across different robots. Impressive, and still a reminder there is a long tail of problems and a need for a step-change in data efficiency. 🔗 Post link
Android Use gives agents hands on phones
An open-source library routes through Android’s accessibility tree, so agents can drive native apps without heavy vision. The demo runs a WhatsApp photo through a trucking workflow for invoicing and payment, with big speed and cost gains on modest hardware. Field ops will like this. 🔗 Post link
On the roads and in pockets, hardware and distribution set the stage for what AI can do and who gets to use it first.
Rivian’s RAP1 autonomy chip
Rivian unveiled RAP1, a TSMC 5 nm multi‑chip module with an in‑house neural engine rated at 800+ sparse INT8 TOPS, forming a Gen 3 Autonomy Computer. A dual setup hits 1600 TOPS, promising 4x Gen 2 performance and better efficiency. LiDAR lands on R2 from late 2026, a call that drew debate from vision‑only fans. 🔗 Post link
Gemini lands on iPhone and iPad
Google’s built‑in Gemini experience is coming to Apple devices, which puts Google’s AI in front of hundreds of millions by default. Bulls see fewer reasons to install rivals and a win‑win for both companies. 🔗 Post link
Models are getting better at producing finished, interactive software - from games to graphics engines - often in a single run.
GPT‑5.2 builds a 3D graphics engine in one go
Pietro Schirano shows a model generating a full 3D engine in a single file, with interactive controls and 4K export. It is a clean, concrete example of rising capability in reasoning and code. 🔗 Post link
GPT 5.2 Pro vs GPT‑5 Thinking for a racing game
Side‑by‑side browser games highlight gains in physics, drifting, UI touches like lap timers and ghost replays. Both were prompted to mimic GTA2’s top‑down feel using plain JS and Canvas, no frameworks or assets. 🔗 Post link
Satya’s “chain of debate” research app
Satya Nadella demoed a Thanksgiving‑built app that pits models against each other with decision frameworks like LLM Council and Microsoft’s DXO, applied to picking India’s all‑time Test XI. The method aims to reduce brittle answers and is on a path to Copilot. 🔗 Post link
OpenAI’s 10‑year easter eggs
From tees and caps to a tongue‑in‑cheek “GARLIC 5.2 OZ” ad for the December model release, an AGI placeholder email, a power‑ranking jab in settings, draggable archives, and Froge hints. It is culture, branding, and a wink at the audience that built the hype. 🔗 Post link
Design is moving back towards craft and play, with tools catching up to how people actually work.
Figma’s new AI image editing in the browser
Erase, isolate, and expand inside Figma using Gemini 3, so designers can remove objects, keep subjects, or extend backgrounds without leaving the file. It trims tool‑switching and piles pressure on legacy editors. 🔗 Post link
Figma Make for government teams
Figma brought its AI prototyper to a FedRAMP‑authorised Gov plan, pitching quicker prototypes, secure collaboration, and earlier citizen testing for modern public services. 🔗 Post link
“The future is for the UI artists”
Microsoft’s Charles Patterson says flat design helped teams ship faster, but the pendulum is swinging toward richer motion and tactile feedback. His morphing button clip captures the mood. 🔗 Post link
Reflective UIs that match the room
A prototype UI mirrors scene colours and light through the camera for a more ambient feel. Replies flag battery and privacy trade‑offs, yet interest is clear from designers chasing context‑aware interfaces. 🔗 Post link
New products and personal finance also featured, with a founder’s-eye view.
Mercury Personal launches
Immad Maroof says the new $240 per year service changed how his family handles money, with joint accounts, controlled cards, high‑yield savings, investments, and free wires under one roof. 🔗 Post link
Science and engineering kept our curiosity alive, from cosmic timings to clever propellers and slippery ice.
The speed of light, to scale
Dr James O’Donoghue’s animation shows light loop Earth in 0.13 seconds, reach the Sun in 8 minutes, and take years to hit the nearest star. It is a brisk reminder of the universe’s size. 🔗 Post link
Voith‑Schneider propellers explained
A vertical‑axis drive with continuously varying blade pitch gives ships instant thrust in any direction, like a helicopter rotor underwater. Great for tugs and ferries where precision matters. 🔗 Post link
Why ice is slippery
Quanta trails a look at the thin watery layer on ice, and what creates it - pressure, friction, pre‑melting, or something else - with Daniel Bonn’s microscopic test bed as a focal point. 🔗 Post link
Quick hit experiments showed how fast ideas can turn into software.
Counting chin‑ups with prompts and a script
A rapid computer vision demo detects bar, hands, and head using Roboflow Rapid and SAM 3 by text prompts, then a short Python script counts reps when a head crosses the bar. Feels like a weekend app in the making. 🔗 Post link
And to close, a classic series poised for another run.
Prince of Persia remakes could sing
A fan argues the Sands of Time trilogy is a goldmine if remade with the care shown in Resident Evil remasters, with the first remake due in 2026. Nostalgia meets modern expectations. 🔗 Post link
Why it matters
Agency is moving off the lab bench. DeepMind’s robot and Android Use point to systems that reason, sense, and act in open‑ended settings, which is what you need for homes, warehouses, and field work. That progress still hinges on data, hardware, and clever shortcuts.
Rivian’s silicon and Google’s distribution show how the next gains will come from hardware stacks and default placement. If Gemini shows up on iPhones and custom chips show up in cars, daily use follows, and so do new behaviours.
On the creator side, Figma’s browser tools and Make for Gov cut tedious handoffs and help non‑experts prototype quickly. Pair that with models that can produce workable engines and games, and the bar to ship narrows. Design culture is moving back toward craft - motion, texture, and context - which should make software feel more human when used with care.
Finally, the science posts are a useful counterweight. Whether it is the time light takes to cross a room, a propeller that can shove a ferry sideways, or ice that is slippery for reasons we still test, they keep our mental model of the world honest while the software races ahead.
