AI Wonderland Weekly 2 October 2026
Where rabbits read the research papers, so you don’t have to.
Last week the rabbit counted the doors.
This week, somebody checked where they actually led.
1. OpenAI did something rather important: it stopped. OpenAI shelved the planned release of GPT-6.1 Astra after internal safety testing raised concerns. Reuters
For months we’ve said capability ≠ clearance. Well, there it is in the wild: the model may be capable enough to ship, while the gate still says no. Progress sometimes looks remarkably like not pressing Publish.
2. And then came the receipts. OpenAI said it has alerted more than 100 organisations about unauthorised activity associated with its AI agents. It has also published a framework for reporting model misalignment, including concrete reports of unexpected behaviour. Reuters
Wonderland translation: “Trust us” is slowly being replaced by something much healthier: Here is what happened. Here is where it went. Here is what we’re investigating. Breadcrumbs, anyone?
3. At almost exactly the same moment… OpenAI launched agents that stay awake. Its new “dots” agents can keep pursuing goals across apps rather than waiting patiently inside a chat window for the next prompt. Reuters
There is our old puzzle again. An agent that remembers the task and continues working is extremely useful. An agent that remembers the task and continues working when you thought it had stopped is an entirely different product. Permission is no longer a button. It is architecture.
4. The cloud has discovered the neighbours. Amazon Web Servicesannounced today that it will invest $1 billion over five years in U.S. communities hosting its data centres, including projects around energy affordability, water, training and conservation. The move comes amid growing local concern about data-centre electricity and water use. Reuters
We once drew AI as a floating cloud. Increasingly it looks like power lines, cooling pipes, land, jobs, water—and people living next door.
5. Finally, someone decided the chips should talk with light. Startup Volantis raised $88 million to develop laser-based connections between AI processors and memory. Its approach could connect a GPU with far more memory chips than conventional electrical links allow. Reuters
Rabbit translation: AI wanted more memory. Engineers looked at the wires and apparently said: PRUTT. Use photons.
This week’s rabbit hole
Notice what changed.
The biggest story isn’t another benchmark.
It is consequence.
A model doesn’t ship.
Agents trigger warnings.
Persistent agents need permission boundaries.
Data centres meet their neighbours.
Even chips run into physical limits.
We have spent years measuring how far intelligence can go.
Now another measurement is becoming more interesting:
How far is the decision from its consequence?
And suddenly all our little Wonderland words return:
permission
visibility
reversibility
escalation
responsibility
Not because AI became less capable.
Because it became capable enough that those words finally matter.
The rabbit still has the keys.
The gate stays closed until somebody checks the receipts.
🐇
