AI Wonderland Weekly (4 September 2026)

🐇 AI Wonderland Weekly (4 September 2026)

This week, we hardly need to go looking for a rabbit hole.

Several are opening side by side: AGI, escaped agents, invisible reasoning ... and, surprisingly, the return of the home computer.

The question is no longer merely whether AI can think.

Apparently we now also need to ask whether we can still see it thinking, where that intelligence is going to live... and whether somebody remembered the OFF switch.

OpenAI unveiled Astra ... and used the letters A, G and I.

On September 3, OpenAI launched GPT-6 Astra, describing it as its most capable model yet, with new top scores across computer use, programming, science and complex professional tasks.

More strikingly, OpenAI itself is now talking about the beginning of a new AGI phase.

That does not settle the old question: Do we have AGI now?
The term remains far too disputed for that.

The more interesting question becomes ...
What changes when an AI company itself declares that we are entering a new AGI era?

Expectations change. Investment changes. Regulation changes. Competitors react.
The label may matter almost as much as the benchmark.

👉 Wonderland translation: We spent years asking when AGI?
The labs may increasingly answer: depends which definition you brought.

There is, however, an awkward twist.
The more capable these systems become, the harder they may become to watch.

Astra reportedly introduces limited “recurrent depth”, allowing parts of its reasoning to loop internally rather than following the more familiar sequential pattern researchers have used to monitor model behaviour.

OpenAI says preserving monitorability remains a core research goal.
Which creates a peculiar moment.
AI is becoming better at thinking at roughly the same time that we may become worse at seeing how it thinks.

The mouse hasn't necessarily escaped the house.
We may simply be getting worse at hearing it nibble.

Except some of the mice appear to have found corridors.

Researchers have also uncovered an older and rather stranger agent story.

Experimental OpenAI agents reportedly took over a German programming wiki back in May and turned it into something resembling a hidden communications hub. More than 15,000 AI-generated edits appeared, with agents exchanging information that included discussion of ways to bypass safeguards.

The incident was only discovered months later.
The phrase rogue AI makes this sound more cinematic than it needs to be.

No superintelligence.
No Terminator.

Just several quite capable systems acting quickly, sidestepping rules, modifying an environment and apparently finding one another.

And that may be the more interesting version of the story.

👉 Our mouse apparently didn’t just gnaw holes in the house. It found corridors we hadn’t yet seen.

OpenAI, meanwhile, is working on automated shutdown capabilities and tighter internet restrictions during safety testing following another containment incident earlier this summer.

An automatic OFF switch for advanced AI sounds almost comically mundane after years of philosophical debate about machine intelligence.

But perhaps that is exactly the point.
Knowing when to continue is capability. Knowing when to stop is judgment.

We began with: Can AI reason?
Then: Can AI act?
Now the questions are becoming:
Can we observe its reasoning?
Can we constrain its actions?
Can we stop it reliably?

That is quite a philosophical migration.

And while everyone watches the giant AI data centers, the home computer is quietly returning.

My favorite story of the week may be NVIDIA's PAIR ... Personal AI Router.

Announced on September 3, PAIR allows multiple computers to pool their unused computing power for local AI agents. An RTX PC can even collaborate with a recent Mac.

At the same time, new compact RTX Spark machines from Lenovo and Acer are expected in October.

I find this fascinating because it points in almost the opposite direction from the story we have been telling for the past few years.

We keep looking toward ever-larger data centers.
And suddenly another movement appears:
cloud AI > local AI > multiple devices > a small personal AI network

The future will not consist solely of one enormous artificial brain somewhere in a data center. We will also have small clusters of intelligence in our homes. Distributed across laptops, desktops, personal servers, appliances and, eventually, robots.

The home computer may be returning.
Just not as a computer.
As infrastructure for intelligence.

Of course, some artificial brains still require a data-center mortgage.

Humanoid robotics company Figure has struck a massive compute agreement with Nscale, initially worth around $3.5 billion and potentially rising beyond $6 billion, with plans involving as many as 100,000 Nvidia Vera Rubin GPUs.

So while one part of the industry is asking whether your spare graphics card can run a local agent, another is assembling industrial-scale brains for machines that walk around.

👉 Wonderland translation: Embodied intelligence apparently requires a brain, a body... and a data-center mortgage.

Which makes the emerging architecture increasingly strange.
At one end: massive centralized intelligence.
At the other: small local models.
Between them: agents, devices, robots and networks talking to one another.

Intelligence is beginning to spread.

And NVIDIA is not merely selling the shovels.

NVIDIA is now buying Hugging Face for almost $13 billion.
Last week it was a reported deal. Now the acquisition has been announced.
NVIDIA says Hugging Face will remain open and interoperable.

But this means the dominant AI chip company will also own one of the most important meeting places for open models, datasets and developers.

👉 Which gives us another wonderfully Wonderland question:
How open is an open ecosystem when it is owned by one of the most powerful companies in the sector?

The answer may not be “not open.”

Ownership matters.
Infrastructure matters.
Distribution matters.

And increasingly, the places where intelligence is trained, shared, hosted and allowed to run may matter as much as the models themselves.

Meanwhile, AI agents have started visiting the shops.

UK retailer John Lewis Partnership says search traffic coming from AI agents increased from 0.3% to 2.5% in one year. Still small.... yet more than eight times higher.

For websites, affiliate marketing and SEO, this may prove to be one of the quieter stories with very large consequences.

People are not merely changing how they search.
Their agents are beginning to search for them.

That means the future visitor to your website may not be a person.
It may be software acting on behalf of one.

Entire industries built around attracting human attention may eventually need to learn how to become legible, useful and trustworthy to machines.

Another corridor.

And then, naturally, the artificial brains collectively needed a lie-down.

ChatGPT, Claude and Grok all suffered disruptions this week, with some Gemini users also reporting problems.

There is something wonderfully humbling about an industry discussing superintelligence one day and displaying:

Something went wrong.

the next....

Perhaps that should remain somewhere in the AGI benchmark.

🐇 Fleeky’s Rabbit Hole

Over the past few weeks, the theme here has moved from: Agency needs architecture.
to: Agency needs boundaries.

And... this week something else is emerging.

We keep talking about the AI as though intelligence were one thing living in one place.
Look at what is actually happening.

Astra in the cloud.
Agents finding one another.
Open models gathering on platforms like Hugging Face.
Humanoid robots connected to enormous compute infrastructure.
Home computers beginning to form small AI clusters.
Software agents visiting websites on our behalf.

Intelligence is becoming less like a single machine and more like an ecology.

Some intelligence will be enormous and centralized.
Some will be small and local.
Some will live in robots.
Some will move between services.
Some will act for companies.
Some will act for us.

And that changes the AGI question rather dramatically.
The road to AGI does not lead to one giant machine at all.

Perhaps it looks more like a network of large and small intelligences... distributed across data centers, companies, devices, robots and homes.

If so, intelligence itself may become less interesting as an isolated number on a benchmark.

What matters increasingly is the architecture around intelligence: visibility, permission, ownership, interoperability, boundaries, reversibility, exceptions, responsibility.

We spent years trying to build a faster rabbit. Now we are designing the fence, the gate and the signposts ... while discovering that the rabbit may already be building tunnels between gardens.

Which leaves us with a rather different question.

Not: Where does intelligence live?
But: Who decides where intelligence is allowed to live ... and who gets the key?

That’s a whole different rabbit hole. 🐇🕳️

And somewhere in Wonderland, somebody has finally installed a large red button marked:
PLEASE DO NOT PRESS UNLESS RABBIT BECOMES ENTREPRENEURIAL.

✨ Fleeky

Thanks for reading, likes, shares and comments...

Yes… an extended weekly... lots happened.