The Rabbit Finds the Action Layer

AI Wonderland Weekly (September 18, 2026)

The Rabbit Finds the Action Layer

Last week, intelligence needed gates.
This week, someone began connecting those gates to laboratories, factories, computers ... and the house.

Rabbit Hole One.
AI Puts On a Lab Coat

On September 17, Anthropic reported that an internal Claude research model optimized more than 30 biological AI models, making them roughly four times faster on average. It also reproduced earlier protein-design performance using about one-hundredth of the GPU time.

The code is being released, while more than 5,000 proposed protein designs will receive wet-lab validation.

This is not a new medicine. It is company-reported research, and the biological designs still need physical testing. But the direction matters: AI is moving from discussing science toward improving the machinery that performs it.

👉 The rabbit has not cured disease.
It has started reorganizing the laboratory.

Rabbit Hole Two
Quantum Tries to Avoid the CUDA Trap

France’s Atomic Energy Commission and quantum startup Alice & Bob announced software that will connect quantum processors with conventional supercomputers. Their explicit concern is avoiding the dependence on one dominant software ecosystem that helped NVIDIA’s CUDA become the gatekeeper of modern AI computing.

This is infrastructure planning, not quantum advantage. Alice & Bob’s machine is not expected to join the system until 2027.

👉 Still, someone is asking an unusually timely question:
Who should own the road before the traffic arrives?

Rabbit Hole Three
The Robot Learns to Sit Down

Agility Robotics unveiled Digit 5, a humanoid designed to work near people without the protective cages surrounding traditional industrial robots. It can detect a person, stop, avoid them... or squat into a safer position.

The specifications are promising: a 22.7-kilogram payload, nine-minute charging and more than 65,000 accumulated operating hours behind the design. Yet the performance claims and $300 million order book come from Agility itself and remain subject to commercial milestones.

👉 Last month, the robot struggled with the crooked cable.
This week, it learned that sometimes intelligence means sitting down.

Rabbit Hole Four
The Agent Comes Home

Perplexity released its local multistep AI agent for Windows. It can plan tasks, use desktop applications, run scheduled work and keep sensitive files on the computer. Cloud assistance requires explicit permission.

There is a rather large doormat: it needs an NVIDIA GPU with at least 24GB of memory and a paid Perplexity subscription.

Meanwhile, Google opened early access allowing compatible AI agents to inspect Google Home activity and control connected cameras, thermostats and lights. The rollout is initially limited to paying US subscribers.

👉 Home AI is no longer merely answering from the kitchen table.
It is approaching the thermostat.

Rabbit Hole Five
Meanwhile, the Humans Hold the Clipboard

Chinese robotics startup Spirit AI predicts a “GPT-3 moment” for robot brains in 2027. That is a forecast, not a breakthrough.

Its more revealing admission is that household robots may still be at least eight years away. Robots reportedly achieve 90% success on simple tasks in structured rooms, yet bottle caps, unfamiliar objects and flexible cables remain stubbornly real.

The humans, meanwhile, are performing household actions repeatedly while wearing sensors so robots can learn from them.

👉 Perhaps we are not being replaced just yet.
We are still demonstrating how the fridge works.

✨ Fleeky

The Rabbit Hole Worth Writing About

AI is entering a new layer.

Not merely knowledge.
Not merely conversation.
Action.

It can improve scientific tools.
Send work between quantum and classical machines.
Move through a factory beside people.
Operate applications on a personal computer.
Reach into a connected home.

The decisive question may no longer be whether AI is intelligent.
It may be "How much distance remains between its decision and a consequence?"

The shorter that distance becomes, the more important permissions, reversibility and human responsibility become.

👉 The next AI revolution may not begin when the machine thinks like us. It may begin when thinking and doing are connected by one very small gate. 🐇

Thank you for reading, likes, comments and shares

My take? It starts to look more secure...