The AI Week in Review #1: What AI Now Owns That Used to Be Yours
⚡ Expert Summary: Something changed in the last seven days, and it wasn’t another chatbot. AI stopped being something that writes things for you and became something that holds things you already made — your notebooks, your code, your money, your house, your own description of your business. This is a weekly column, published every Monday: the week’s biggest AI stories, then what they mean when you stack them together and where the frontier is actually heading next. Not a roundup. You can get a roundup anywhere. The argument.
Let me tell you what this is, because you’ve clicked on a roundup before and been burned.
Every Monday I’ll take the five or six AI stories that actually mattered in the previous seven days. Not the ones with the biggest press releases. The ones that changed something. Then I’ll do the thing almost nobody does: I’ll put them next to each other and tell you what they mean when they’re side by side, because five stories you read separately look like five unrelated things, and five stories read together look like one thing happening very fast.
Here’s what that one thing was this week. AI crossed over from generating things to holding them. A year ago the question was whether AI could produce text, an image, a song. This week the question became what happens when it’s the thing storing, reading, and acting on material that already exists — a dead mathematician’s unpublished notebooks, a company’s own codebase, your card at the checkout, your living room. And the harder question underneath it: who owns the material once the machine is the one holding it?
Nobody has answered that. Not the companies. Not the courts. Not the regulators. That silence is the story, and it runs through every one of the five items below.
A Dead Mathematician, 1,009 Pages, and 663 Theorems Nobody Had Ever Read
Kurt Gödel left behind 1,009 pages of densely written notes, most of it in an obsolete German shorthand called Gabelsberger that almost nobody alive can read. The Guardian’s Alex Bellos reported on November 13 that those pages, once decoded, are now said to contain 663 previously unknown proved theorems.
The man who had been decoding them by hand is Jan von Plato, a logician and professor at the University of Helsinki. He’d been working at roughly one page per week. Claude read the whole set in a couple of days.
“I was flabbergasted,” von Plato told the Guardian. “The workbooks not only contain dozens of anticipations of results not known until the 70s and 80s, but they also contain 14 themes in logic that have not been pursued. Gödel was far ahead of his time.”
And here’s the part that makes this a result rather than a nice anecdote: von Plato ran the reconstructed proofs through Coq, an interactive theorem prover that mechanically checks whether a proof is actually valid. They passed. This isn’t a language model writing something that sounds like mathematics. It’s a machine rebuilding arguments that a checker then confirmed.
What doesn’t add up: the Guardian does not say who legally owns the Gödel archive, or whether any living descendant or copyright holder was consulted before an AI ingested and reconstructed a dead man’s private notes. And Joel David Hamkins at the University of Notre Dame gave the honest hedge: if the findings stand up “it is remarkable and historically interesting. Although how much is Claude and how much is Gödel is a job for the historians,” adding that resemblance to a historical path “does not mean he had it. It shows he was on the path. You don’t get credit for work you don’t tell anyone about.”
What it means for you: the material sitting in your files, your notebooks, your old drives — someone’s AI can read it now, at a speed no person can match, and there’s no law that says they asked first.

One Million Tokens for a Flat $1.50, and the End of Forgetting
Google shipped Gemini 4 Hydrogen on November 13, and the number that matters isn’t a benchmark. It’s this, from Google’s own blog: the model is “priced flat across its entire 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.”
Read that again. Most AI pricing gets more expensive the more you feed it, because the model has to keep re-reading what you gave it. Flat pricing across a million tokens means the economics of “read everything I have” just collapsed. A million tokens is roughly a stack of novels. Google also claims it runs 2.5x faster than the Flash architectures it replaces, and it scored 54 on the Artificial Analysis intelligence index in xhigh thinking mode, 51 in high.
What doesn’t add up: there’s no byline on Google’s post — it’s team-attributed, which means no named human is standing behind the numbers. Independent developers note that getting the full window to actually work in production depends on tiered context caching, and Google’s post is silent on regional limits, free-tier rate caps, and real enterprise SLAs. The “1 million tokens” is a ceiling, not a promise that every token is equally usable.
What it means for you: the reason AI forgot your earlier messages — the reason you had to repeat yourself — was a cost problem, and the cost just fell through the floor. Expect assistants that remember your entire history with them, and expect that memory to be a product feature somebody is monetising.
Your Card, in an Agent’s Hands
On November 14, OpenAI launched ChatGPT Wallet, letting its agents buy things on a user’s behalf. It’s limited at launch to US ChatGPT Pro subscribers aged 18 and over, it supports American Express, Mastercard and Visa, and it generates a dynamic per-transaction payment token instead of exposing your stored card credentials — which is genuinely the right design.
Nayak, OpenAI’s head of monetization and payments product, framed the ambition plainly: “Our goal here is to really focus on merchants and our network partners.” Luke Gebb of American Express was more careful about how fast this should move: “It just makes sense to watch it, start it in a contained environment, and then keep rolling from there.”
What doesn’t add up: the reporting says cardholders keep their existing protections, but it never allocates liability when an autonomous agent buys the wrong thing because the model got something wrong. That gap is the whole story. And the pushback is already here — Amazon, eBay and United Airlines are actively moving to block or restrict third-party automated shopping agents. United has already said it can’t reliably tell when an agent is the one booking. So the wallet arrives into a market where the biggest merchants are putting up walls.
What it means for you: your money is now something software can act on, with your biometric approval as the last gate — and the question of who eats the loss for a wrong purchase is currently nobody’s published responsibility.
| Story | What AI now holds | Who carries the risk |
|---|---|---|
| Claude / Gödel, 1,009 pages | A dead man’s unpublished notebooks | Nobody named — rights holder unconfirmed |
| Gemini 4 Hydrogen, 1M tokens | Everything you have ever fed it | The customer, at $1.50 per million in |
| ChatGPT Wallet | Your card, one token per transaction | Undefined when the agent errs |
| Apple HomeView | Your living room and your context | EU buyers: no product at any price |
| OpenAI Daybreak / Defense Factory | Your codebase, attack and repair | Enterprise adopters, unevaluated |
Seven Countries Get the Smart Display. The EU Doesn’t.
Apple released HomeView on November 13 — a $299 device with a 7-inch Liquid Retina display on a domed speaker base, connected by a stainless steel hinge. Available in the US, UK, Australia, Canada, India, New Zealand and Singapore.
Apple’s own release says why Europe isn’t on the list: “Apple HomeView is not available for sale in the European Union, as Siri AI is currently unavailable in iOS in the EU. Apple is working constructively with the EU in an attempt to find a path forward.”
The new Siri is described as “an entirely new version of Siri powered by the next generation of Apple Intelligence and advanced Apple silicon… a profoundly more capable and conversational assistant with personal context understanding and onscreen awareness.”
What doesn’t add up: there is no date, no timeline, and no named executive standing behind the EU situation — it’s all company-attributed language. “Working constructively” is what companies say when they have nothing to announce. A hardware product is being held off an entire continent by a software regulatory question, and Apple won’t say when that resolves.
What it means for you: if you’re in Europe, you’re now in a two-tier market for AI hardware — and the tier you’re in was decided by a compliance argument, not by what’s available.
OpenAI Arming the Defence Side, and Admitting It Doesn’t Know If It Works
At its Cyber at the Frontier keynote, OpenAI President Greg Brockman introduced two tools: Daybreak Red Team, which runs automated offensive penetration testing against corporate web applications using specialised cyber agents, and Defense Factory, which takes confirmed vulnerabilities and generates merge-ready code fixes with validation evidence, requiring mandatory human software engineer review before deployment.
Brockman’s framing was that “frontier-model defense as the new security floor” makes traditional app-security tooling without native frontier integration structurally vulnerable. The Register’s Alexis Dufresne had the story on November 13.
What doesn’t add up: OpenAI has not published independent accuracy percentages or false-positive metrics for either tool. That leaves early enterprise adopters testing reliability themselves, on their own infrastructure, at their own risk. And the mandatory human review step means the “autonomous defence” isn’t autonomous — it’s a very fast junior engineer whose work a human still has to sign off on.
What it means for you: if you run a business with a website, the attack side of this is now automated and cheap, and the defence side is automated and unevaluated. That’s not a comfortable pairing.
| What shipped this week | The limit you will actually hit | Published answer? |
|---|---|---|
| Claude on 1,009 pages of Godel shorthand | Historians still disagree on how much is Claude | No — open question, by Hamkins’ own account |
| Gemini 4 Hydrogen, 1M tokens flat | Needs tiered context caching to work in production | No — regional and free-tier caps unpublished |
| ChatGPT Wallet, agent-initiated purchases | US Pro subscribers, 18+, biometric each time | No — liability when the agent errs is undefined |
| Apple HomeView, $299 smart display | Seven countries only; EU gets nothing | No date, no named executive, no timeline |
| OpenAI Daybreak and Defense Factory | Human engineer must sign off before merge | No — no independent accuracy or false-positive rate |
The Turn: All Five Stories Are the Same Story
Stack them and the pattern isn’t subtle.
Gödel’s notebooks are material that existed and was unreadable. Claude made it readable. Gemini’s million-token window is the price of holding everything you’ve ever written in one place. ChatGPT Wallet is AI acting on the money you already have. HomeView is AI living inside the room you already own. Defense Factory is AI reading the code you already shipped.
Every single one of those is a machine reaching into something that already exists, and the thing it reaches into is yours — your archive, your history, your card, your house, your repository.
Here’s the mechanism, plainly. Reading used to be the expensive part. A human had to sit with material, page by page, at von Plato’s one page per week. Once reading is nearly free and near-instant, the bottleneck moves from can we process this to who is allowed to. And the answer, right now, is: whoever has the API key.
The villain here isn’t a person, and it isn’t a company being evil. It’s a business model. The entire value of these systems comes from ingesting material they did not create — every one of the five stories depends on it — and the entire cost of that ingestion lands on whoever owned the material. Gödel’s rights holders weren’t asked. Your bank can’t tell if an agent or a person is spending. United can’t identify the agent booking a seat. Apple pulled a product rather than answer the EU’s question. The incentive runs one direction: read everything, own nothing, and let the person who made it carry the downside.
And the honest counter-case, because you deserve it: none of this is autonomous yet, and pretending otherwise is how you get fooled twice. ChatGPT Wallet is US Pro subscribers only, with biometric approval on every transaction. Defense Factory requires a human engineer’s sign-off before any fix ships. Gemini’s million tokens need caching to work in practice. Von Plato ran the proofs through Coq precisely because nobody trusts an AI’s word on a proof, including him. Human validation is still the bottleneck everywhere — and that’s the strongest argument that the “AI owns everything now” panic is premature.
But notice what that argument actually says. It doesn’t say the systems are safe. It says a human is currently standing between the machine and the consequence — and every one of this week’s announcements is an effort to remove that human from the loop as fast as possible. That’s not my opinion. That’s Brockman’s keynote premise, OpenAI’s payments roadmap, and Apple’s entire pitch for Siri. The human is the thing being engineered out, politely, with a review step attached to make it feel safe.
What none of the five stories says out loud: the ownership question has no owner. Not one of these announcements includes a clause about who is responsible when the AI reads something it shouldn’t, spends something it shouldn’t, or describes your business in a way you can’t correct. That’s not an oversight. That’s the deal.

How This Touches Your Daily Life
At home. If you’re in the US, UK, Australia, Canada, India, New Zealand or Singapore, a $299 screen in your kitchen is now listening with what Apple calls “onscreen awareness” and “personal context understanding.” If you’re in Europe, that product does not exist for you, and not because you can’t afford it. That’s the first time a smart device has been regionally blocked by an AI compliance question rather than a supply chain one, and it won’t be the last.
At work. If you’re a security engineer, your job just split in two: the attackers got an automated red team, and the fixes arrive pre-written for your review. If you’re a researcher, a translator, an archivist, or anyone whose value was the slow careful reading of difficult material, von Plato is the honest case study — decades of work, compressed by a machine, and his own reaction was “flabbergasted.” Not threatened. Flabbergasted. Watch what your employer does with that word.
With your money. An agent can now spend from your card. You’ll approve it biometrically each time, for now. Cardholders reportedly keep their existing protections, but if the agent buys the wrong thing and nobody has written down who pays, you are the one with the empty account while the lawyers figure it out. And the merchants you buy from most — Amazon, eBay, United — are already trying to shut agents out entirely.
What you now have to check. This is the new chore the week created, and it’s the one I’d tape to the fridge: before you let an AI read, hold, or act on anything you own, find out what you’re giving it permission to do with it. Not the marketing page. The terms. Because every story above is a permission slip being signed by someone who didn’t read it — and this week, six months of permissions got handed out at once.
What you still don’t have to worry about. Genuinely: the machines are not running loose. Wallet is US Pro only with biometric approval. Defense Factory won’t merge a line of code without a human engineer signing it. Gemini’s million tokens need caching to behave in production. Gödel’s theorems passed a formal checker that a human chose to run. The humans are still in the loop — this week nobody took them out. They just started building the door.
Where the Frontier Is Heading: Three Predictions, and What Would Prove Them Wrong
One — a provenance fight over dead authors’ archives becomes public before Q1 2027. The Gödel case is too clean to stay quiet: a famous archive, an unidentified rights holder, and 663 new results with commercial value. Somebody is going to ask who gets paid. What would prove this wrong: if the archive’s custodians come forward, confirm they were consulted, and publish the terms — the story closes and the pattern breaks.
Two — at least one major retailer or airline launches a formal agent-detection programme, not just a block. United already admitted it can’t tell who’s booking. Amazon and eBay are blocking outright. Blocking is a stopgap; within weeks one of them will announce a verification layer for legitimate agents, because refusing all agent traffic means refusing real revenue. What would prove this wrong: if the blocks hold for a quarter with no revenue impact, nobody builds the verification layer.
Three — a liability clause for agentic purchases gets published, and it won’t favour the customer. The current silence on who pays when an agent buys wrong is unsustainable once volume arrives. Expect a card network or OpenAI to publish the rule, and expect it to place verification duty on the cardholder. What would prove this wrong: a major issuer announcing chargeback protection for agent-initiated transactions without conditions.

Where This Leaves You Tonight
Nothing here is finished. Wallet is US-only and weeks old. HomeView is in seven countries and blocked in a whole continent. Von Plato’s Gödel work is being checked by historians who don’t yet agree on what it means. Defense Factory shipped without a published accuracy number.
But the direction of travel isn’t ambiguous, and that’s all this column is for. Last Monday, the interesting question about AI was what it could make. Tonight, the interesting question is what it’s holding, and whether anyone asked the person who made it.
Next Monday this column picks it up again, because by then it will have moved — it always does. In the meantime, one question worth carrying into your week: what do you own that a machine can now read faster than you can protect it?
Answer that honestly and you’ll know exactly what to do next.
Frequently Asked Questions
What is the AI Week in Review?
The AI Week in Review is a Monday column on Top Notch AI Tools that takes the week’s biggest AI stories, explains what each one actually changed, then goes further: it shows what the stories mean when stacked together, where the frontier is heading next, and how all of it touches your daily life.
Did Claude really discover new mathematics in Gödel’s notebooks?
Not quite. Claude decoded 1,009 pages of Kurt Gödel’s obsolete Gabelsberger shorthand and surfaced 663 previously unknown proved theorems. Jan von Plato then ran the reconstructed proofs through Coq, an interactive theorem prover that mechanically validates proofs. The proofs check out, but logician Joel David Hamkins notes that how much is Claude and how much is Gödel remains a question for historians.
Can ChatGPT Wallet spend my money without asking?
No. ChatGPT Wallet requires users to review and biometrically approve transactions, and it is currently limited to US ChatGPT Pro subscribers aged 18 or older. What is not defined in the reporting is who carries liability if an autonomous agent makes an erroneous purchase — users reportedly keep existing cardholder protections, but fault allocation is unaddressed.
Why isn’t Apple HomeView available in the EU?
Apple’s own release states that HomeView is not for sale in the European Union because Siri AI is currently unavailable in iOS in the EU, and that Apple is “working constructively with the EU in an attempt to find a path forward.” No timeline or resolution date has been published.
Does Gemini 4 Hydrogen really have a 1 million token context window for $1.50?
The flat pricing is $1.50 per million input tokens and $7.50 per million output tokens across the full 1 million token window, per Google’s announcement. In practice, independent developers note that full-length utilisation depends on tiered context caching, and Google has not published regional limits or free-tier rate caps.
Are AI security tools replacing human security engineers?
Not yet. OpenAI’s Defense Factory generates merge-ready code fixes for confirmed vulnerabilities but requires mandatory human software engineer review before deployment, and OpenAI has not published independent accuracy or false-positive metrics. Human validation remains a required step in the pipeline.
About Barak Schurr
Barak Schurr built and sold a major ticketing platform, ran global nightlife operations at Eventbrite, and now spends his time figuring out which AI tools actually move the needle for independent business owners — and which ones are just noise. No hype. No fluff. Just what works.
Email:
[email protected]
More AI Resources for Your Business


