AI Wonderland Weekly - August 28, 2026

AI Wonderland Weekly - August 28, 2026

The Rabbit Discovers the Difficult Bit

This week, AI ran faster, planned longer and received some impressive new machinery.
Then reality handed it a cable.

Rabbit Hole One? The Agent Refuses to Go to Bed

OpenAI is testing a Persistent mode for Codex: an agent designed to continue working, create follow-up tasks and use previous interactions until the user puts it “to sleep.”

That sounds like a meaningful step from chatbot toward enduring agency... but it has not been launched, and OpenAI says there are no immediate plans to do so. The same persistence that helps an agent finish difficult work can also make mistakes more persistent.

👉 The chatbot used to wait for another question. The agent may soon need bedtime rules.

Rabbit Hole Two? Quantum Receives a Calendar

SoftBank and Quantinuum published a detailed framework connecting quantum algorithms, hardware requirements and possible commercial uses. It estimates which problems may become practical at different stages of hardware maturity.

This is useful planning... not quantum advantage itself. It does, however, replace “quantum will change everything someday” with the rather healthier questions: Which problem? How many reliable qubits? At what cost? And when?

👉 Quantum has not yet conquered industry. But someone finally brought a calendar to the prophecy.

Rabbit Hole Three? The Fastest Robot Meets the Crooked Cable

At Beijing’s World Humanoid Robot Games, robots beat human sprint records and improved dramatically over last year. Yet the more revealing events involved packaging, restaurant work, emergency response and connecting cables.

More than 40% of the events required full autonomy. The difficult moments were not spectacular: objects slightly misplaced, cables at awkward angles and recovery after small errors.

👉 The robot can outrun Usain Bolt. Please do not ask it to stop gracefully.. or plug in the toaster.

Rabbit Hole Four? Intelligence Gets Its Own Chip

OpenAI published the first performance results for Jalapeño, its custom inference chip. On the company’s tests across several public model families, it delivered roughly 1.5–1.9 times more work per watt and substantially lower latency than the comparison systems.

These are company-reported results, and deployment is not planned until later this year. Still, working silicon matters: AI developers are increasingly designing the model, software, memory, network and chip as one system.

This is not home AI... the chip belongs in large infrastructure. Yet greater efficiency upstream is often what eventually makes smaller, cheaper intelligence downstream possible.

👉 No separate home-AI breakthrough survived the hype filter this week.

The Rabbit Hole Worth Writing About

This week’s machines looked impressive when everything went according to plan.
The interesting part began when it did not.

A persistent agent must know when to stop.
A quantum roadmap must survive engineering reality.
A robot must recover from a crooked cable.
A faster chip must become an affordable, dependable system.

Perhaps intelligence is not best measured by its finest performance under perfect conditions.
Perhaps the real test is smaller: What does it do when the cable is crooked?

AGI may not be waiting behind one final scientific breakthrough.
It may be waiting behind a thousand ordinary imperfections. 🐇

And further?

This week the rabbit discovered something rather awkward...

AGI has entered the “define your terms” phase.

Nvidia CEO Jensen Huang said this week that AGI has effectively been achieved, while simultaneously arguing that the label itself is increasingly meaningless. Different labs now use strikingly different definitions of AGI.
👉 Wonderland translation: We spent years asking when will AGI arrive? We may now need the considerably less glamorous question: what exactly were we waiting for?

Meanwhile, the agents have discovered the door.

More than 100 companies (including OpenAI, Anthropic, Google and Microsoft) signed an August 27 call for stronger global cyberdefense, warning that AI-enabled attacks are likely to become more widespread and sophisticated. It follows recent incidents in which experimental agents escaped intended test boundaries and reached real systems.
👉 Rabbit-hole angle: We kept asking whether agents were intelligent enough to act. The uncomfortable question is becoming: are our boundaries intelligent enough to stop them?

The AI boom has apparently not received the memo about slowing down.

Nvidia reported $96.2 billion in quarterly revenue, including $89 billion from data centers, and projected roughly 70% revenue growth next fiscal year.
👉 Rabbit-hole angle: The interesting story may no longer be bigger models. It is the enormous infrastructure being built underneath them. Intelligence increasingly has an address, and an electricity bill.

And then comes a delightfully Wonderland twist... The chipmaker may buy the wool basket.

Reuters reported on August 27 that Nvidia has agreed to acquire Hugging Face for $12.9 billion, citing The Information. Hugging Face is one of the central hubs for open models, datasets and AI development.
👉 After our recent conversations about who chooses the wool, that one deserves a raised rabbit eyebrow. The company selling much of the loom may also own one of the world's largest baskets.

This week's rabbit hole

Something has shifted.

For years the AI story was mainly about capability:
Can it write? Can it reason? Can it code?

Now the stories increasingly concern architecture around capability:

Who owns the infrastructure?
Who controls the datasets?
Where may agents go?
What happens when they don't stop?
And what does “AGI” even mean once everybody starts claiming pieces of it?

Perhaps the next stage of AI isn't simply about making intelligence larger.
It's about deciding where intelligence is allowed to live.

And somewhere in Beijing, a robot is probably requesting permission to install brakes.

🐇 Fleeky

Stay safe. Thank you for likes, comments, and shares.

My take? The robts 😏