AI Wonderland Weekly – 23 October 2026
The Rabbit Discovers That Everyone Has Hired an Assistant
Where rabbits read the research papers, so you don’t have to.
Last week, the rabbit followed the footprints.
This week, it discovered an entire workforce making them.
AI is learning to read DNA, investigate brains, design images, supervise other AI systems and find holes in the internet.
Apparently, the rabbit hole now requires a personnel department.
1. Claude Duet arrives before the coffee.
On October 21, Anthropic introduced Claude Duet, a proactive assistant that monitors connected applications, prepares daily briefings and helps complete work across different tools.
It can even prepare financial forecasts and build applications. Sensitive actions still require approval. For now, access is limited to a US research-preview waitlist.
Rabbit translation: We used to ask AI what needed doing.
Now AI arrives with a list.
👉 The rabbit wonders when it will start asking why we haven’t finished yesterday’s tasks. Reuters
2. Thousands of AI agents investigated a mouse brain. In 32 minutes.
This one deserves a second look.
On October 21, OpenAI, PerturbAI and Scripps Research reported using thousands of AI agents to analyse existing experimental data from millions of mouse brain cells.
The agents connected genetic changes with scientific literature, uncovering promising research questions related to neurological disease. The analysis took just 32 minutes.
These are research leads, not confirmed treatments.
Rabbit translation: One scientist reads a paper. A thousand AI assistants read the library. The scientist still decides which experiments are worth doing.
👉 And perhaps asks who drank all the coffee. OpenAI
3. Hugging Face wants to chat with your DNA.
On October 22, Hugging Face unveiled Carbon-X, an open model that understands both natural language and DNA sequences.
Researchers can ask questions about genetic variations and biological functions without relying entirely on expensive proprietary systems.
It can also run on researchers’ own infrastructure.
Rabbit translation: We taught AI to speak English. Then mathematics. Now we’re teaching it to read the alphabet of life.
A, C, G and T.
Four letters. Rather complicated conversations.
👉 And no, understanding DNA doesn’t mean we can simply rewrite biology. Hugging Face
4. Canva decided to build its own imagination machine.
On October 22, Canva announced Craft, its first publicly available image-generation model developed in-house.
Here’s the interesting twist.
Instead of immediately generating an image, Craft first plans its composition. Then it renders.The result is intended to offer more predictable layouts, readable text and editable designs.
Rabbit translation: AI image generation is discovering something artists have known for centuries. Sometimes you should sketch before painting.
👉 The rabbit approves. Especially when the final image contains the correct number of paws. Canva
5. NVIDIA built a playground with fences.
On October 21, NVIDIA announced OpenShell, a reference architecture designed to contain, monitor and enforce security policies for autonomous AI agents.
The goal is to give agents useful capabilities without granting unlimited freedom.
This is an architectural proposal and technology platform, not proof that autonomous agents are now perfectly safe.
Rabbit translation: The agents wanted a bigger playground. The engineers agreed. Then installed gates, cameras and an emergency stop.
👉 Somebody has finally remembered that autonomy and permission are different things. NVIDIA Newsroom
6. AI found the internet’s cracks. The repair crew is exhausted.
And now for the less comfortable news.
On October 23, Reuters reported that AI-assisted vulnerability discovery is accelerating dramatically.
Security researchers are uncovering weaknesses in widely used software faster than organisations can always repair them. Some websites have temporarily gone offline as operators scramble to protect their systems. Experts disagree about whether the growing number of discoveries will necessarily translate into more successful attacks.
Rabbit translation: AI found the cracks in the castle walls. Unfortunately, the builders and burglars received the same map.
👉 The race is no longer just about finding weaknesses. It’s about repairing them in time. Reuters
This week’s rabbit hole
Notice something?
We have assistants helping assistants.
Models interpreting biology.
Machines planning creative work.
Agents examining other agents.
And security systems struggling to keep pace with what intelligence can discover.
The interesting question is changing. Not simply: How intelligent can AI become?
But: How do we organise intelligence when it becomes distributed across thousands of systems, tools and decisions?
Because more intelligence doesn’t automatically mean more understanding.
More agents don’t automatically mean better coordination.
And more discoveries don’t automatically mean better outcomes.
Perhaps the next frontier isn’t another enormous brain.
Perhaps it’s learning how different kinds of intelligence can cooperate without losing sight of their purpose.
The rabbit closes its notebook.
Looks at the growing workforce.
And asks one final question.
“Wonderful. But who’s supervising the supervisors?”
🐇
Same rabbit. More helpers. Bigger questions.
This week’s takeaway: AI is becoming less of a single assistant and more of an interconnected workforce. The challenge is making that workforce useful, understandable and accountable.
