01The thaw
The government that blacklisted Anthropic just came back.
An AI model just got so dangerous that the U.S. government — the same government that, weeks ago, told agencies to cut ties with Anthropic — quietly walked back to the table on Friday.
The administration had labeled Anthropic a supply-chain risk, arguing its restrictions could limit defense readiness, and ordered agencies to cut ties. Lawsuits followed. It looked over.
Then Anthropic previewed Mythos. And the tone changed.
Weeks ago — the blacklist
The administration was walking away.
- DesignationSupply-chain risk.
- OrderAgencies told to cut ties.
- LegalLawsuits filed in response.
- MeetingsNone scheduled.
- PostureHostile and public.
This Friday — the thaw
The administration was at the table.
- MeetingSenior White House officials sat down with Anthropic.
- AgenciesFederal agencies now preparing to use Mythos.
- AlliesAllied governments briefed and coordinating.
- TreasuryPersonally calling bank CEOs.
- Posture“Productive and constructive.”
That’s government-speak for: we need them.
02Capability, confirmed
This isn’t hype. The UK regulator confirmed it.
Mythos can autonomously find and exploit software vulnerabilities. Not suggest them. Not guide a human. Actually exploit them. The UK’s AI Security Institute confirmed it can do in minutes what takes human researchers days.
~40
Companies with early access
Minutes
vs. days for human researchers
Allies
Governments briefed before release
Banks
Treasury calling CEOs directly
So Anthropic did something unusual. They didn’t release it. They gave early access to about forty companies that run the internet — Apple, Google, Amazon — so they could patch systems first. Critical infrastructure operators got access. Allied governments got briefed.
Jim’s take
Federal agencies are now preparing to use Mythos. Other governments are coordinating. Treasury is personally calling bank executives. You don’t move like that for hype.
Mythos looks real. Real enough that a government that tried to cut Anthropic off is now asking how to work with them.
This isn’t abstract. You’re about to see a wave of security updates — your laptop, your browser, your phone. Don’t ignore them. Once a patch is released, attackers can reverse-engineer it to find the vulnerability it fixed.
04Forecasts, contradictions
Five CEOs, five forecasts, zero certainty.
Verizon’s CEO, Dan Schulman, just said the quiet part out loud: he’s publicly predicting 20 to 30 percent unemployment in the next two to five years. He’s already announced 13,000 layoffs, the biggest in the company’s history, and set aside $20 million for a “career transition fund for the age of AI.”
Schulman (Verizon)
“20–30% unemployment in the next two to five years.”
Just announced 13,000 layoffs — the largest in Verizon’s history.
Huang (Nvidia)
“A lot of people are saying we’re going to run out of jobs. It’s exactly the opposite.”
Jassy at Amazon sits in the middle: AI reshapes jobs, doesn’t eliminate them — but many roles will need fewer people.
A BCG study says up to 15% of U.S. jobs could be eliminated outright. A Quinnipiac poll shows 55% of Americans now think AI does more harm than good — up from 44% last year. Five different people, five different answers, all running multi-hundred-billion-dollar operations. None of them agree on whether we’re heading into a jobs apocalypse or a productivity boom.
Jim’s take
Most of the “AI caused our layoffs” stories aren’t actually caused by AI. They’re caused by over-hiring in 2021, higher rates, and CEOs looking for a cleaner public story than “we grew too fast.” AI is the cover, not the reason.
On the other side, the “we’ll all have better jobs” crowd is mostly selling chips or cloud services. That’s talking their book.
If you’re a professional trying to plan, don’t plan on certainty. Plan on ranges. Keep your skills portable. Keep your options open. And don’t let a CEO — on either side — tell you they know how it ends.
We’ve never had a tool like this before. Nothing in history gets this smart this fast.
05Retail, by default
Shopify just enrolled four million merchants in agent checkout.
Shopify flipped a switch. Every one of their four million merchants is now, by default, accessible to AI shopping agents — ChatGPT, Google AI Mode, Copilot, Perplexity. Merchants can opt out. But they’re opted in. By default.
4M
Merchants enrolled by default
$80B
Amazon ad revenue at stake
~15%
Projected agent share of revenue
0
Teams dedicated to agent SEO today
When a customer tells ChatGPT “buy me running shoes under 150 bucks,” the AI doesn’t see your Instagram. It doesn’t see your influencer deal. It sees three things: your price, your structured product data, your return policy. Maybe your review count. That’s it. That’s your brand to the agent — a data feed and a return window.
Ten years of marketing strategy. Twenty years of DTC brand building. Shopify just committed four million merchants to a protocol where most of that doesn’t matter nearly as much. And the merchants didn’t vote on it.
Shopify
Opened the gates. Agent checkout on by default for all four million merchants.
Betting the agents win.
Amazon
Went to federal court to block Perplexity’s agent. A judge sided with Amazon in March.
Betting it can keep control — and protect $80B in ad revenue.
Jim’s take
Agent SEO is the new SEO. If you own a brand, audit what your product data looks like to a machine this week. Not your website. Not your social feed. The actual feed — the structured data Shopify sends the agent. Because pretty soon that’s going to be 15% of your revenue.
Nobody — I mean nobody — has a team dedicated to it yet.
06The memory got big
AI’s short-term memory just became a document drawer.
Anthropic shipped Claude Opus 4.7 on Thursday. Standard with it: a 1-million-token context window for every paid user. Doesn’t matter which plan. Doesn’t matter where you’re running it. No price premium.
1M
Tokens, every paid tier
800pg
Research, in one shot
You can hand Claude an entire codebase — the whole thing, not chopped-up pieces — and ask it a question. You can hand it a quarter’s worth of legal filings. 800 pages of research. All at once. The model just reads it. And based on early testing, retrieval quality holds up across the full window.
07The platform pattern
The RAG industry just got absorbed.
An entire industry — real, venture-backed, thousands of engineers — got built in 2024 around the fact that the model couldn’t hold that much at once. Retrieval-augmented generation. Vector databases. Embedding pipelines. Billions in infrastructure. All of it exists because the model had a small window.
My take on this is a pattern we’ve all seen before. It’s the Apple pattern.
Apple lets the ecosystem build something amazing on top of their platform. A flashlight app. A clipboard manager. A screen time tracker. They let it get popular. They let developers make real money. They let the category prove the value. And then in the next iOS release, Apple absorbs the feature into the core operating system.
The startup that built it wakes up one morning and finds out their business is now a Settings toggle.
That’s literally what just happened to the vector database industry. The model absorbed the function. The startups didn’t do anything wrong. They just built on the wrong layer.
Jim’s take
Anthropic and OpenAI will keep doing this. Every capability that currently lives above the model — memory, retrieval, routing, agent orchestration — is a candidate to get pulled inside the model next.
If you’re a builder, the question is: is the thing I’m building something the model will absorb, or something the model will need? Those are very different businesses. Only one of them survives the next twelve months.
