Ep48 Live Flow (private)

00:00 · Cold openSix and a half years of government security work. Done in twenty hours.

📰 The news — say this first

On July 6, Anthropic published a case study with the government of Alberta, Canada. Around 50 Claude Code agents working in parallel scanned roughly 466 million lines of code across 1,280 applications and 3,400 repositories in about 20 hours, then generated fixes and tests. Alberta estimates the same job by hand: about six and a half years.

Why open here

Picture the IT worker staring at that backlog on a Friday. A swarm of AI agents clears it before the weekend is over. That's what "AI agents" actually means when you point them at a real, boring, enormous job.

Picture an IT worker in Alberta's government staring at a security backlog they've estimated at six and a half years of work. Decades of patched-together software. Then fifty AI agents clear the whole scan in twenty hours. Before the weekend is over.
And the case study is specific about how. It ran a two-stage routine: first a rules engine flags known patterns in each repository, then Claude reviews every flag and cites the exact file and line, so an engineer can go verify it. Where a system had no automated tests, Claude wrote the tests first.
I still want to push back on the optimism, because this is like hiring fifty of the world's fastest interns. They type at lightning speed, but nobody on that team remembers why the firewall was built that way back in 2012. If you don't check their work, a mistake gets hardwired into 466 million lines at light speed.
Which is exactly why nothing shipped on its own. Anthropic's own write-up says every patch was reviewed and approved by the Ministry's engineers before it went anywhere. The agents do the years of grunt work. The humans still sign off.
And keep the salt handy: this is Anthropic's case study about Anthropic's tool. The six-and-a-half-years number is Alberta's estimate, not an independent audit.
Still, the unlock here isn't a smarter answer. It's years of work you'd never start, done in a day. That changes what a small team can take on.

ANTHROPIC × ALBERTA · JULY 6466M LINES · 1,280 APPS · 3,400 REPOS~50 AGENTS · 20 HOURS"6.5 YEARS" = ALBERTA'S ESTIMATE

The question to sit with · for your audience

If AI can compress years into a day, what's the job in your world you wrote off as too big to ever start?

⚠ Producer note (backstage, not on air) — VERIFIED CORRECTION

The prep audio invented a mechanism: agents working in "isolated sandbox environments on separate branches." Not in the case study — CUT. What IS verified: the two-stage scan (rules engine → Claude reviews flags, cites file and line), tests-written-first, and "before any patch shipped, it was reviewed and approved by the team." Also: audio transcription mis-hears "cloud code" — it's Claude Code.

~03:30Three new coding AIs in two weeks. Betting on one brand is now the risky move.

📰 The news — say this first

Three companies shipped new coding AIs in fourteen days. June 30: Anthropic's Sonnet 5 became its new, cheaper default. July 8: Elon Musk's xAI released Grok 4.5, its strongest model yet, trained alongside the popular coding tool Cursor and built right into it. July 9: Meta launched Muse Spark plus its first paid developer API, openly chasing Anthropic, OpenAI, and Google.

The frame that lands

"Best AI for the job" is a moving target that shifts weekly now. Loyalty to the task, not the brand.

Fourteen days, three top-tier coding models. Sonnet 5 on June 30. Grok 4.5 on July 8, trained alongside Cursor and built right into it. And Meta the very next day with Muse Spark and their first paid developer API.
For anyone who doesn't code every day: Cursor is one of the most popular AI coding tools out there. Grok showing up inside the tool people already use matters more than any leaderboard score.
Two weeks ago on this show we talked about the death of model loyalty, when Sakana shipped that Fugu router. These two weeks are the proof. Betting your whole workflow on one brand is the risky move now, not the safe one.
One caution: every one of these launches says "our strongest model yet." That's the company's own claim. Meta's is days old. The real verdict takes a few weeks of people actually using them.
So the skill isn't mastering one model. It's building a workflow flexible enough to switch the second something better drops.

SONNET 5 · JUNE 30GROK 4.5 · JULY 8 · INSIDE CURSORMUSE SPARK + PAID API · JULY 9

Try this week · for your audience

Take one real task you did with your usual AI and run it through one of the new models. Judge by your own result, not the launch post.

Callback (Rule: frame as payoff, not fresh)

Ep47 covered Sakana's Fugu router as "the death of model loyalty." This beat is the payoff: say "we called this two weeks ago," don't re-explain Fugu.

~06:30OpenAI checked the industry's favorite AI test. A third of it was broken.

📰 The news — say this first

On July 8, OpenAI published an audit of SWE-Bench Pro, one of the coding benchmarks the whole industry cites to prove whose AI is best. OpenAI found about 30 percent of the tasks are broken and formally withdrew its own earlier recommendation to use the benchmark.

The frame that lands

So many "our AI is the best coder" claims ride on this scoreboard. The scoreboard just failed its own test.

When a press release says "our AI is the smartest coder in the world," this benchmark is often the test they're pointing at. OpenAI audited it and found roughly thirty percent of the tasks broken. Then they pulled their own recommendation to use it.
And "broken" is very concrete. Real example from the audit: the task's prompt shows an example with one leading space, but the hidden grading test demands two. The AI solves the actual problem and gets marked wrong over a single space character.
So a correct answer fails because the rubric wanted an exact match. And about a third of the test is like that. Every confident ranking built on it is shakier than it looks.
Give OpenAI credit, pulling your own recommendation is rare and honest. But notice who's grading: a leading AI company auditing the test that scores its own rivals. Useful. Not a neutral referee.
For everyone picking tools: this is your permission to stop taking benchmark bragging at face value. Run your own real task through two AIs and judge which one actually helped. That's your benchmark.

OPENAI AUDIT · JULY 8~30% OF TASKS BROKENRECOMMENDATION WITHDRAWN

Why this matters · for your audience

Your own task is the only benchmark that counts. Same prompt, two models, side by side. Five minutes, better than any leaderboard.

Producer note (backstage, not on air)

Mechanism VERIFIED against OpenAI's own post ("Separating signal from noise in coding evaluations"): overly strict tests, underspecified prompts, low-coverage tests. The one-space example is real (OpenLibrary task). Say "OpenAI found," and keep the not-a-neutral-referee side in.

~09:30A government switched a top AI off. Then back on.

📰 The news — say this first

A US export directive, citing national security, forced Anthropic to pull its Fable 5 model offline worldwide in June. The government lifted the controls on June 30, and Fable 5 came back globally on July 1 after Anthropic added what it calls an improved safety classifier. Meanwhile: Sonnet 5 became Anthropic's cheaper default (intro pricing through August 31), California put Claude in state government at 50 percent off, and the EU delayed its high-risk AI rules to late 2027 and 2028.

The frame that lands

The on/off switch stopped being hypothetical. If a government can freeze your primary tool overnight, you don't own your tools. You're renting a utility.

Plain version of what happened: in June, a US directive citing national security ordered Anthropic to suspend access to Fable 5 for any foreign national. Anywhere. Including Anthropic's own employees.
And here's why that meant everyone lost it. You can't check passports in real time on an API. So the only way to comply was to switch the model off for all customers, worldwide, overnight.
Both sides of this one. The government's stated concern was a reported jailbreak, a technique that could push the model into dangerous cybersecurity territory. That's their reason, on the record.
And Anthropic complied but pushed back publicly. They said the evidence was verbal, that the jailbreak was narrow and not universal, that similar capability exists in other public models, and they called for a clearer legal process. Both of those things are true at once.
Then the comeback: Anthropic trained a new safety classifier it says blocks that technique in over 99 percent of cases, the government lifted the order June 30, and the model was back worldwide July 1. The trade-off: it now blocks more harmless coding requests too. Someone decided what your tool will and won't do for you.
And governments aren't just regulating, they're buying. California just put Claude into every state agency, city, and county at half price, with free training. Sonnet 5 became the cheaper default, promo pricing ends August 31, so plan around it. And the EU pushed its high-risk rules back to 2027 and 2028.
Last episode our whole theme was AI becoming invisible infrastructure, like electricity. Well, the utility now has a government switch. So the question: whose tool is it really?

FABLE 5 BACK · JULY 1 · WORLDWIDECONTROLS LIFTED JUNE 30SONNET 5 DEFAULT · PROMO ENDS AUG 31CALIFORNIA 50% OFF · JUNE 29EU RULES → LATE 2027–2028

Why this matters · for your audience

AI policy stopped being abstract. You can now see it in the tools you use. One more reason not to build your whole workflow on a single model.

⚠ Producer note (backstage, not on air) — VERIFIED CORRECTION

The prep audio invented the legal mechanism: "compute thresholds" classifying the model as "dual-use like weaponry or aerospace." Fabricated — CUT. Per Anthropic's own posts: a targeted export directive citing national security over a reported narrow, non-universal jailbreak; Anthropic disputed it as overly broad while complying. Also: "stronger safety filter" is a press paraphrase — Anthropic's actual words are "improved safety classifier." Say "Anthropic says." Do NOT explain the freeze with the dual-use/weaponry frame on air.

~13:30 · human hookA genuinely good, free AI image maker just landed inside Instagram and WhatsApp.

📰 The news — say this first

On July 7, Meta launched Muse Image, its first in-house AI image generator. Free for basic use and built into Instagram Stories, WhatsApp, and Meta AI chat. You edit with sketches and plain-language notes, and it searches the web to get details right: a QR code that actually scans, a chart that actually adds up.

The frame that lands

No new app, no new login, no prompt engineering. The lowest-friction way most people will ever start using AI images is now inside something they open every day.

The delivery is the real story here. They didn't launch a website you have to go visit. If you're on Instagram or WhatsApp, a capable image tool is now sitting inside an app you already open every day. A post, a flyer, a thumbnail.
And the accuracy piece is the sleeper feature. It checks the web while it generates, so the QR code on your flyer actually scans and the chart actually adds up, instead of the AI guessing at squiggly lines.
Now the wrinkle. It can pull public Instagram accounts into generated images just from a handle in the prompt. There is an opt-out, but the default is on. That's the part drawing real scrutiny.
And free from Meta usually means your usage helps train the next version. That's the price tag on "free."
Which leaves the question: when a polished image is free and instant for billions of people, what makes yours worth looking at?

MUSE IMAGE · JULY 7FREE BASIC · IG / WHATSAPP / META AIPUBLIC-PROFILE PULL · OPT-OUT, DEFAULT ON

Try this week · for your audience

Make one real thing with it: a flyer, a story graphic, a thumbnail, right inside WhatsApp or Instagram. See what it gets right and where it still needs you.

~16:00 · cultureAn AI "actress" just landed a real movie role.

📰 The news — say this first

Announced around July 6: an AI-generated performer named Tilly Norwood was cast to star in a feature film. A fully synthetic performer in a role a working actor could have had.

The frame that lands

This is the AI story people argue about at dinner. Not benchmarks. A synthetic performer competing for a human's job, right now.

This isn't a revived legacy voice or a background effect. It's a fully synthetic lead, cast in a real feature film, in a role a living actor could have had.
And you can see the studio math. A performer who never needs a second take, never has a scheduling conflict, never renegotiates a contract. That's exactly why actors are alarmed.
Both sides, honestly. Supporters say it's a new creative tool, a character, not a person taking a paycheck. Critics say a synthetic performer taking a real role is precisely the line people said they didn't want crossed.
For anyone who earns a living being a human on camera, or on a page, this stopped being hypothetical this week.
So in a feed already full of AI faces: what's the value of being an actual human, and how do you make that value obvious?

TILLY NORWOOD · ~JULY 6FULLY SYNTHETIC LEAD · FEATURE FILM

~18:00 · cultureNetflix brought Gene Wilder's voice back. His family said yes.

📰 The news — say this first

For a new Wonka-world reality series, Netflix used AI to recreate the voice of Gene Wilder, the original Willy Wonka, as the narrator. His widow, Karen Wilder, publicly approved it, calling it a celebration of his imagination. Announced June 30, the 55th anniversary of the 1971 film. Wilder passed in 2016.

The frame that lands

A beloved voice, gone since 2016, speaking again. Touching and unsettling at the same time. Consent settles the legal question, not the human one.

Put these two stories side by side. Tilly Norwood is a synthetic newcomer competing for work. This is a beloved human being brought back, with his family's yes, on the film's 55th anniversary.
And the consent matters enormously. The estate said yes, the widow called it a celebration of his imagination. Legally, that settles it.
But it doesn't settle the human part. Plenty of people find a late performer speaking new lines uneasy even with permission, and the online pushback is real.
That tension, touching and unsettling at once, is exactly why this is the human heart of the whole two weeks.
So here's the one to take home: if your family could bring your voice back after you're gone, and they wanted to, would you want them to?

ANNOUNCED JUNE 30 · 55TH ANNIVERSARYKAREN WILDER · PUBLIC APPROVALWILDER PASSED 2016

~20:00Quick hits: the tools you already use got better.

📰 The news — say this first (about 10 seconds each)

ChatGPT voice (OpenAI, July 8): it can listen and talk at the same time now, with web search and memory. Gemini on your Mac (Google, June 30): a Mac app that organizes, renames, and summarizes your local files, triggerable from your phone with the laptop closed. Gemini API background agents (Google, July 7): builder news, agents running in the background with outside tools. FireSat (Google, July 7): satellites that now spot wildfires as small as five meters across.

Four fast ones. Talking to ChatGPT stopped being a walkie-talkie. It listens and talks at the same time now, with search and memory. And Google put Gemini on the Mac: it organizes, renames, and summarizes your files, and you can trigger it from your phone while the laptop is closed.

Important detail on that one: it works on your files, it does not drive your screen. Faster and less disruptive. For builders, Google also added background agents to the Gemini API. And my favorite: FireSat satellites can now catch a fire the size of a living room, from orbit. AI pointed at something that matters.

CHATGPT VOICE · JULY 8GEMINI MAC APP · JUNE 30GEMINI API AGENTS · JULY 7FIRESAT · 5-METER FIRES · JULY 7

Producer note (backstage, not on air)

Keep these to ~10 seconds each. Gemini Mac app: say "acts on your files, doesn't drive your screen" — the audio's "file system hooks" phrasing is invented terminology, skip it.

~21:30 · NEW BEATMyth of the Week: the pocket-sized SpaceX phone was never real.

📰 The news — say this first

The headline everyone shared: "SpaceX is building a pocket-sized iPhone killer." The truth: Elon Musk called the report "utterly false" the same day it ran. A rumor, denied by the person it was about.

Why this beat exists

Half of what goes viral in AI is a newsletter's guess dressed up as news. This beat is the show's whole value proposition in 60 seconds.

This thing was everywhere. Renderings, feature lists, people on Reddit debating the battery life of a phone that does not exist.
And that's the anatomy of it: a speculative theory gets aggregated, stripped of context, aggregated again, until it reads like objective fact. Nobody upstream ever confirmed anything.
So the habit to steal from us: before you share an AI headline, check whether the company or the founder actually confirmed it. Musk denied this one the same day, in public. That one habit will save you from looking silly, and it's the whole reason a show that checks its facts is worth your time over a hype feed.

MUSK: "UTTERLY FALSE"DENIED SAME DAY THE REPORT RAN

Producer note (backstage, not on air)

This is the new beat the viewer panel promoted — a recurring "Myth of the Week" candidate. If it lands on air, make it a weekly ritual.

~22:45 · RecapTwo weeks in one breath.

Fast recap. AI agents did six and a half years of government security work in twenty hours. Three new coding models dropped in two weeks. And OpenAI found a third of the industry's favorite benchmark is broken.
A government switched a top model off and back on, so you're renting, not owning. Meta put a real image maker inside WhatsApp and Instagram. A synthetic actress got a lead role, and Gene Wilder's voice came back with his family's blessing.
The through-line: who controls the AI, who grades it, and what it changes for you. Pick tools by the task. Check the source before you share. And stay the human who decides.

Provocative close → hands to Deep Dive 1 (update handoff line once the deep dive is locked)

We just watched AI write the code, make the images, and even take an acting role. So here's the one to mull: if machines are making the art, and your personal AI agent watches it for you and hands you a five-second summary, are we engineering ourselves out of being the audience at all? Sit with that one. Let's get into today's deep dive.