Live Flow · Ep43 · Conversational Show Script

0:00 – 0:45 The Hook — the 27-year-old time bomb

Cold open · Memorize

A couple of weeks ago we told you about an AI that found a flaw hiding in the internet's foundations since 1999. 27 years, missed by thousands of brilliant engineers, flagged by an AI in minutes. That bug was just the opening number. This week Anthropic released the full scorecard, and it's staggering.

"Welcome to Practical AI, Episode 43. We are unpacking one historic week, the week leading up to May 28th, 2026."
"And the headline is simple. The era of AI as a tool is over. This isn't a faster spell-checker. It's an autonomous workforce you can hire."
"We go from a 27-year-old bug, to a $50 billion check, to a 42,000-word letter from the Pope. Let's get into it."

0:45 – 3:30 Mythos — 10,000 flaws in 30 days

📰 The news — say this first

We covered Anthropic's security AI, Claude Mythos, when it found that 27-year-old bug two weeks ago. This week the full results landed: across 50 companies in 30 days, it found more than 10,000 serious software flaws.

The proof, not the hype

The proof the era changed isn't a keynote. It's in the trenches of software engineering. Anthropic gave its vulnerability model, Claude Mythos, to 50 organizations for 30 days. It flagged 10,000+ high or critical flaws across 1,000+ projects.

"Cloudflare alone pointed it at their own systems and it found 2,000 bugs. Independent security firms confirmed over 90% were real."
"And that 27-year-old OpenBSD bug we showed you two weeks ago? It's in this batch, alongside thousands more. For anyone not living in server rooms, OpenBSD is famous for being hyper-secure. It underpins a huge chunk of the secure internet."
"So how does a community that prides itself on security miss something for a generation, and an AI spots it instantly?"
"It's how they read. A human reads code sequentially, holding it in their head. A deep flaw is line 14 in one file breaking line 8,000 in another. No human can hold an entire codebase at once."
"But the AI ingests the whole library into its context window and maps it all at the same time. The unit of work goes from a human reading one file for an hour, to a machine analyzing millions of interactions in minutes."
"For decades we relied on security through obscurity. The bet was that the code was too boring for a skilled hacker to spend six months digging. That economics just shattered. A scan like this costs a few hundred dollars."

50 orgs · 30 days 10,000+ critical flaws Cloudflare · 2,000 bugs 27-yr-old OpenBSD flaw

Why this matters · For your audience

The question is no longer "is my software secure?" It's "when does an AI find the flaw that's been hiding for decades?" Attackers have the same tool now. But defenders can finally scan and patch their whole codebase overnight. It's officially machines fighting machines.

3:30 – 6:00 UltraCode — the 3am digital Thunderdome

📰 The news — say this first

Anthropic shipped a new Claude Code feature called dynamic workflows, switched on with a mode called "ultracode." You ask for a workflow and Claude spins up a fleet of AI agents (up to ~1,000) that build and check each other's work. It does in a night what used to take a team months. Released this week with their new model, Opus 4.8.

The analogy that lands

You type "create a workflow," and the AI spins up a fleet of agents. About 16 at once, hundreds across a job. It's like hiring 16 senior engineers at 3am, locking them in a room, and starting a digital Thunderdome.

"And the agents aren't just splitting a to-do list. Half are paid to build the solution. The other half are paid to ruthlessly tear that code apart. They argue, they break it, they iterate, until the result is bulletproof. No human mediates."
"The source says it turns a 3-month corporate migration into an overnight job. It's a simulated engineering team running at silicon speed."
"But there's a real cost under the hood. Hundreds of agents arguing means an astronomical number of API calls. The industry calls it the token burn rate."
"They don't need coffee or sleep. But they'll absolutely burn through the company credit card to get that 3-month job done by morning."
"The ceiling on what one person can build just skyrocketed. If you can afford the token burn, you've got an entire engineering department and a security team at your fingertips."

The big one · Pairs with Mythos

In a single week, one company shipped an AI that replaces a security firm and an AI that replaces an engineering team. That's the whole show in one line: AI is now a team, not a tool.

6:00 – 9:00 Anthropic — when a software vendor becomes a superpower

📰 The news — say this first

Anthropic raised $50 billion (its biggest round ever, and the second-largest private round in history behind only OpenAI's $122B in March), opened an office in Seoul, joined the rollout of the Pope's new 42,000-word letter on AI, and tightened its own safety rules. All in one week.

The frame

If one company can replace a security firm and an engineering team in the same week, it stops being a software vendor. It becomes a foundational pillar of the global economy. The numbers this week prove it.

"Anthropic raised $50 billion in a Series H. That's the second-biggest private funding round in history. Only OpenAI's $122 billion back in March was bigger. And it makes Anthropic one of the most valuable private companies on the planet."
"And at the same moment, they open an office in Seoul and name a country lead, because South Koreans are adopting Claude at 3.5 times the expected rate for their population."
"Then there's the Vatican. Pope Leo released a 42,000-word encyclical, one of the highest forms of papal communication, on protecting the human person in the age of AI. And Anthropic's researcher Chris Olah took part in the presentation."
"While the same company quietly updates its own safety policy to version 3.3, tightening the guardrails around chemical and biological weapons."
"So I'm trying to reconcile this. One company takes $50 billion from investors who demand growth, while rewriting its own rules so the product can't build a bioweapon. We're asking a startup to do the job of the United Nations."
"That's general-purpose technology. When your system touches everything from an overnight code migration to mapping biological threats, the company that builds it touches everything too. They're governing themselves, because they move faster than governments can legislate."

$50B Series H · 2nd-biggest ever Seoul office · 3.5x adoption Vatican · 42,000-word encyclical Safety policy v3.3

9:00 – 11:00 Taiwan — the atoms behind the algorithms

📰 The news — say this first

Nvidia announced it's spending $150 billion a year in Taiwan and breaking ground on a new campus. AMD added more than $10 billion. Almost all of the world's AI chips are made on one island.

The pivot

You can have $50 billion in the bank and brilliant algorithms, but you can't conjure agents out of thin air. There's a physical bottleneck to this invisible revolution. And it sits on one island.

"Nvidia's Jensen Huang announced they're now spending $150 billion a year across the Taiwanese ecosystem. Annually. They're breaking ground on a new campus called Constellation. Huang called Taiwan the epicenter of the AI revolution."
"And AMD followed right behind with more than $10 billion of their own across the Taiwan chip ecosystem."
"Every autonomous agent in that Thunderdome runs on cutting-edge silicon made predominantly on a single island. It's an unprecedented concentration of physical risk."
"The whole industry is tethering its future to one center of gravity, because the demand for compute is basically infinite right now."

11:00 – 14:00 The math race — AI that grades its own work

📰 The news — say this first

Google DeepMind's AI solved 9 unsolved math problems (two stuck for 56 years) and verified its own proofs with no human checking. A day after OpenAI claimed a similar win that still needed human experts to confirm it.

Why it matters

More compute doesn't help if the AI just hallucinates faster. The only way to use that infinite power is if the machine can verify its own reasoning. That's exactly what happened in math this week.

"First, OpenAI claimed a breakthrough, disproving an 80-year-old math conjecture. But they had to lean on human experts to verify it. Then, a day later, Google DeepMind announces AlphaProof Nexus solved 9 open Erdős problems. Two of them unsolved by humans for 56 years."
"And the real story is how. DeepMind's proofs were machine-verified, automatically, with no human in the loop."
"Walk people through that, because it's the key."
"They paired a creative language model with a formal logic solver. The language model is the creative student tossing out ideas. The logic solver is the ruthless professor. It grades every step, and if there's one flaw, it rejects it and makes the model try again. Thousands of loops, until the proof is flawless."
"And here's what I love. Demis Hassabis, the CEO who just oversaw this, comes out and tempers the hype. He says this is still not AGI. He calls it the foothills of the singularity, and puts true AGI about 4 years out."
"4 years is still fast. But once a machine can formally verify its own logic, it graduates from a text predictor to an autonomous reasoning engine. That's the same mechanism powering the coding agents."

DeepMind · 9 Erdős problems 2 unsolved for 56 years Machine-verified, no human Hassabis · "~4 years to AGI"

14:00 – 16:30 Down to your desk — proteins and PowerPoint

📰 The news — say this first

Zuckerberg's Biohub released an AI that mapped 7 billion proteins and can invent brand-new ones to fight disease. And OpenAI put ChatGPT inside PowerPoint while Google put Gemini inside Adobe, Canva, and CapCut.

From the lab to your laptop

This same autonomous reasoning is hitting the highest peaks of human health and the most mundane office tasks. Same week. Same shift.

"The Chan Zuckerberg Biohub launched an AI world model that mapped nearly 7 billion proteins."
"And a world model is not a database. A database is a list of known proteins. A world model learned the grammar and physics of how proteins fold. Just like ChatGPT learned English to write new sentences, this learned biology to invent proteins that have never existed."
"They're compressing years of drug discovery into hours. They've already designed novel protein binders aimed at cancer and immune diseases."
"And at the other end, the mundane. OpenAI shipped a ChatGPT add-in for PowerPoint that builds the slides natively. Google wired Gemini straight into Adobe, Canva, and CapCut."
"The friction is gone. You're not opening a browser to ask an AI. The reasoning engine is now an invisible layer inside the software you already use."

16:30 – 18:30 The free internet starts to crack

📰 The news — say this first

Meta started charging for its apps: $3.99 a month for exclusive features on Instagram and Facebook, and $2.99 for WhatsApp.

The jarring one

At the exact moment AI embeds into every tool to make life easier, the platforms we've used free for a decade are putting up paywalls. That's not a coincidence.

"Meta just rolled out paid Plus tiers. $3.99 a month for Instagram and Facebook, $2.99 for WhatsApp. The causality is undeniable."
"You cannot run a world where billions of people prompt AI agents all day on ad revenue alone. The compute cost is forcing the end of the free internet."
"Consumers are finally feeling the ripple of those $150 billion ecosystem investments. The whole digital world is being rewired for autonomous intelligence, and the cost is trickling down to you."
"We're moving from an internet of human content paid for by ads, to an internet of machine reasoning paid for by subscriptions."

For your audience

The free ride is ending because intelligence costs real money to run. Expect more of your everyday tools to start charging. The question is whether what you get back is worth it.

18:30 – 19:15 Summary — the scope of one week

"To trace the scope of this one week:"
"An autonomous AI found a 27-year-old time bomb that humans missed for a generation."
"That capability turned Anthropic into a $50 billion global center of gravity, rewriting bioweapon guardrails and sitting with the Vatican."
"$150 billion poured into Taiwan just to keep the silicon flowing."
"DeepMind taught machines to verify their own 56-year-old math proofs."
"And the same shift slipped into your PowerPoint while it started charging you for Instagram."

★ The Provocative Close · 19:15 – 20:00

We saw AI argue with itself in a Thunderdome, breaking its own code and verifying its own work with no human in the room. We saw DeepMind's machines grade their own math.

So what happens when we point that same self-verifying intelligence at designing its own hardware? Or, more pressingly, at rewriting its own safety guardrails?

If the machine can generate the idea and ruthlessly grade the idea without us, are we still the builders of this technology, or have we become the budget providers for an intelligence learning to build itself?

Keep questioning. Now let's get into the demo.

Ep43 Live Flow