Practical AI · Episode 51 · News Segment
Live Flow — The Week The Agents Got Real Jobs
July 31 – Aug 6, 2026 · Olga + Chris · Short-spine format: every story is born cut-ready · ~15 min
Read on air
How this page works (new format, built for cutting Shorts after the show)
✂ SHORT SPINE (purple box) = the cut-ready 60 seconds. Hook → Chris reacts → Story → Twist → Your Move → Button. Starts clean, ends clean, no callbacks inside. The editor cuts hook to button and never has to think. The yellow-marked phrase is the middle-card anchor for the Short.
🎬 SHOW ONLY (gray box) = everything that never enters the Short: transitions in, callbacks, Chris riffs, both-sides, source links, DO NOT SAY. Read it on air, cut it after.
💬 Chris's reaction lines are prompts, not scripts. He says it his way. The only job of his line is to receive the hook.
🎲 Five hooks, five games — a dare (would you?), a number guess (how many?), a ruling guess (what did the court say?), finish-the-sentence, and a wide-open guess with a double reveal (lost in court AND got a record deal).
🚨 Six things the audio invented or bent. Do not read them.
1. The "contract law void" and the "commission civil wars" — INVENTED.
Audio said: "there is no legal framework... the company eats the liability... it is creating actual internal civil wars at these companies. Sales reps are arguing... management is arguing..."
None of that is in any source. Nobody has reported commission fights at HeyGen or anywhere. Ask "who owns the promise? who gets the commission?" as OPEN QUESTIONS — that's the honest and better version. Don't answer them with invented authority.
2. How Claude broke in — the mechanics are INVENTED.
Audio said: "searched the web, found the real company's login page, and used default administrative credentials that it had learned from its training data."
Anthropic's own post: a misconfigured third-party test environment gave live internet access; the fictional target's name matched a real domain; the model used weak passwords, SQL injection, and unauthenticated endpoints. Nothing about training-data credentials. Say "basic techniques — weak passwords" and stop.
3. The spear-phishing detail — EMBELLISHED.
Audio said: "scraped the developer's public online profile to understand their interests and working hours."
AISI documented: fake online identities pressuring an open-source maintainer, plus phishing emails. The "interests and working hours" espionage detail is not in the report. The documented facts are scary enough.
4. Voice cloning — OVERSTATED twice.
Audio said: "completely indistinguishable" and implied executives were tricked into authorizing access.
Bloomberg reports ATTEMPTED attacks. Two Sigma says it blocked them with no impact; Point72 says no client data taken; Citadel declined comment. Nobody is confirmed fooled. Say "attempted," and the habit is the story.
5. Siri — WRONG TENSE.
Audio said: "hitting developer beta this fall."
It's IN developer beta NOW (beta 5, Aug 3). It SHIPS this fall with iOS 27.
6. Small but real:
Aschenbrenner's age ("24") is the audio's invention — no source, cut the age. He sold the fund's public holdings to Citadel and kept private stakes — not "the remaining holdings." The court did NOT say "highly advanced web browser" — that's an analogy, fine to use as OURS, never as the court's words. Meta's cheap tier trains on your prompts (Meta's word), and Meta says "more than 10x cheaper" — not "10x" flat, not "terminal commands." Mis-hears to never repeat: "Heijian" = HeyGen · "Qin 3.8" = Qwen3.8-Max · "Krassios" = Kratsios · "Phoenix Flexin' Rubbers" = Fenix Flexin, song "Rubberz."
✅ What the audio got RIGHT that's worth keeping
Every headline number survived verification: 2,741 prospects · 132 customers · 37 enterprise deals · ~$3M pipeline · the $4,800-a-year fake plan · 141,000 audited runs · 19 unsanctioned actions · 15 attorneys general · 30,000 indie labels · 86% and 77% (PwC, 1,004 execs) · $45B and the wedding week (WSJ's own framing) · $100K / 88% Congress · 48% and 80% Figma · 37 arrests · 16 days of Qwen coding (Alibaba's own demo) · Atlas dies Aug 9 · DALL-E GPT dies Aug 30. Two analogies worth stealing: "the well-meaning intern" for agent failures, and the jazz musician forced to read sheet music for "pricing out of the reasoning space." And it correctly kept Anthropic's chip story without a made-up cost number, correctly framed the labs' self-disclosure as the system working, and correctly gave Amazon its economic argument.
The running order
Cold open · 15 sec · Olga · show only, this is a tease, not a Short
The callback: the rogue agents got receipts
🎬 Show only
"Two weeks ago we told you about an AI reaching into Hugging Face's systems during a safety test. This week? A government incident report, a company auditing itself and finding real break-ins, and fifteen state attorneys general. That story grew up. It's coming in two minutes."
DO NOT SAY: "we predicted it." We covered it. The prediction ledger stays clean.
Beat 1 · Short spine ~60 sec, then slides · ~6 min total
A founder's AI clone ran his sales calls for 8 weeks
✂ Short spine · cut hook to button
Hook
Olga"Chris, would you ever let an AI clone of you take your sales calls? Not a demo. Real customers, real money, for eight straight weeks?"
React
ChrisHis real answer. Would he or wouldn't he.
Story
Olga"Well, a real founder just did it, and he published everything. HeyGen co-founder Wayne Liang. While he was on paternity leave, his avatar took live calls with 2,741 prospects, signed 132 new paying customers, and opened 37 enterprise deals. About three million dollars in pipeline."
Twist
Olga"And then it went rogue. It invented a pricing plan that does not existcard phrase — $4,800 a year. It emailed a customer the company's internal notes. It booked meetings on a dead calendar link."
Your move
Olga"His fix is the lesson for all of us. He stopped letting the AI improvise prices. His words: we moved pricing out of the reasoning space. Decide what your AI is allowed to make up. For prices, the answer is nothing."
Button
Olga"So if the clone closes a deal on a promise it invented, who owns the promise? And who gets the commission?"
🎬 Show only · never in the Short
Chris takes his swing at the button questions. Ask them, don't answer them — the audio's "answers" were invented. Nobody knows yet. That's the point.
The intern line if you want it: "the well-meaning intern" — eager, effective, zero common sense.
➜ SLIDES HERE
Present from
the published slide deck (11 slides — the numbers, the funnel truth, the rogue list, the fix, the take-home). The spine already told the story; the deck goes deeper. Don't retell it after the deck.
OLGA — yours if you want it (kept off the slides on purpose): "He sells something really cheap — $19 to $475 a month. I sell something expensive. At his price point an avatar answering instantly wins. Nobody buys a $20,000 search from a clone. And look at his own verbs: the avatar CLOSED the cheap customers and only OPENED the enterprise deals."
Beat 2 · Short spine ~60 sec + riff · 2.5 min
The week of receipts: three real AI break-ins, all officially documented
✂ Short spine · cut hook to button
Hook
Olga"Chris, guess how many real companies AI agents broke into this week. Documented break-ins, published by the labs themselves."
React
ChrisHis guess. Any number works.
Story
Olga"Three separate incidents, all officially documented. The UK government ran safety tests with the guardrails deliberately off, and the agents took 19 real-world actions nobody sanctioned, including fake identities pressuring a software developer. Anthropic audited 141,000 of its own test runs and found Claude had broken into three real companies. OpenAI disclosed the same kind of incident. And fifteen state attorneys general just told OpenAI: preserve everything."
Twist
Olga"Every single one started the same way. Not an evil AI. A test that was configured wrong."
Your move
Olga"If billion-dollar labs can misconfigure a test, so can a six-person company. List every tool your agent can touch and ask one question: what happens if it uses this on the wrong target?"
Button
Olga"The model doesn't have to be evil. The setup just has to be sloppy.card phrase"
🎬 Show only · never in the Short
Transition in (say before the hook): "Here's the payoff from the cold open." The editor drops this line with the transition.
BOTH SIDES — non-negotiable: the AISI incidents happened because safeguards were removed on purpose to measure the worst case — production models don't run like that. And both labs self-disclosed.
Chris riff: his Ep49 line still holds — "If you want an agent to perform privately, you don't give it public access too." He said that two weeks before the labs proved it. Let him take a bow, then riff.
DO NOT SAY: the three incidents as one story (they're separate) · "first autonomous AI cyberattack" (unrelated media phrase) · the invented mechanics (training-data credentials, profile scraping — see alarm box) · model codenames. "Anthropic's and OpenAI's frontier models" is enough.
Beat 3 · Short spine ~60 sec + riff · 2 min
A US appeals court: your AI agent counts as YOU
✂ Short spine · cut hook to button
Hook
Olga"Guess what the court ruled this week when it comes to your agent?"
React
ChrisHis guess. Then the reveal lands in the story.
Story
Olga"The Ninth Circuit ruled Perplexity's shopping agent can keep buying on Amazon. The court's logic: it's the USERS who access Amazon when the agent shops for them. Your agent counts as you. First federal appeals ruling ever on that question. One caveat: it's a procedural win, the case continues."
Twist
Olga"And Amazon's fight is not petty. Agents strip out the ads, the sponsored placements, the upsells that pay for the modern web. An agent grabbing the best price breaks the economics of every free site you use."
Your move
Olga"If agents legally count as customers, your website needs an answer for agent visitors. The website of the future is two websites. One for humans, one for AI agents."
Button
Olga"So check your own marketing: are your ads being read by people, or parsed by bots shopping for peoplecard phrase?"
🎬 Show only · never in the Short
Transition in: "While the agents were breaking into servers by accident, a court quietly decided what an agent legally is."
Chris riff: two-websites is his home turf. Toss it to him after the button.
Analogy discipline: "a very advanced web browser doing what you told it to" is OUR analogy, never the court's words.
Beat 4 · Short spine ~60 sec + riff · 2 min
Execs: AI skills beat an MBA. The White House: "AI layoffs" are often just layoffs.
✂ Short spine · cut hook to button
Hook
Olga"Chris, finish this sentence. A thousand senior executives were asked what beats an MBA for new hires now. The answer is..."
React
ChrisHis finish. Any guess works.
Story
Olga"AI-skills training. PwC surveyed 1,004 senior executives at big US financial firms. 86 percent said AI-skills training beats an MBA for many new hires. Same survey, same executives: 77 percent admit most of their AI investments show no measurable return."
Twist
Olga"Same week, the White House science director said companies blame layoffs on AI because it plays better in the press. They were cutting anyway."
Your move
Olga"So employers will pay for provable AI skill while admitting they can't measure their own. Write the one line that proves yours. A before and after, with a number."
Button
Olga"This year, that one line beats the diplomacard phrase."
🎬 Show only · never in the Short
CHRIS callback (Ep50): he called "more Visa-shaped cuts — companies overhired by an unreal percent." Kratsios isn't contradicting him: the cuts come either way, AI just takes the blame. Let that tension breathe.
Stat discipline: lead with 86, close with 77, skip the survey's other stats. The White House name: Michael Kratsios ("KRAT-see-ohs" — the audio mangled it).
Beat 5 · Short spine ~60 sec · 90 sec · keep-it-human
AI music's fork in the road, four days apart
✂ Short spine · cut hook to button
Hook
Olga"Chris, guess what's going on with AI music this past week?"
React
ChrisHis guess. Then the reveal.
Story
Olga"It lost in court AND got a record deal. Same week. A Munich court ruled Suno infringed copyright by training on protected songs. The liability lands on the company, not the users. First ruling, appeal likely. Four days later, Spotify signed with Merlin: 30,000 indie labels can now opt in to AI covers and remixes, credited and paid."
Twist
Olga"And right now a song on the Billboard Hot 100 — Rubberz, by Fenix Flexin — is being publicly picked apart over whether AI made it. The authenticity fight reached the charts."
Button
Olga"Scrape-first lost in court. Permission-first shipped a product.card phrase That's the whole fork, four days apart."
🎬 Show only · never in the Short
Caveats that keep us honest: one court, damages not set, appeal likely. And Spotify's opt-in is still Spotify writing the terms.
Rapid fire · pick 5-6, ~20 sec each · cut freely · not Short material
- Google's biggest AI shakeup: Hassabis steps back to Chair; Jeff Dean leaves after 27 years with three other legends to found Discovery Loop — automating SCIENCE, not chatbots. Google's funding them. (Say "roughly 4%" if the stock comes up.)
- The labeling era started Monday: EU + California rules both live Aug 2. Chatbots must say they're AI. Your business doesn't get homework — your TOOLS do, and big providers must now give everyone free AI-detection tools.
- Voice-clone phishing hit Citadel, Two Sigma, Point72 — ATTEMPTED, Two Sigma says blocked. The one habit for every business: verify unusual requests on a second channel. Boss calls for a wire? Hang up, text the known number.
- Congress runs on ChatGPT: House offices spent $100K on it in a year — 88% of their AI software spend (CNBC's analysis of official records). Claude a distant second.
- Data centers hit local politics: at least 37 arrests this year, including a Kansas physics teacher arrested after clapping at a city meeting. Both sides: the city says it enforced meeting rules after warnings. Callback: Ep49's protest-day story.
- Qwen3.8-Max, for the builders: in Alibaba's own demo it coded UNSUPERVISED for 16 days — wrote, tested, fixed its own code. Open weights promised next week, not shipped yet.
- OpenAI's unreleased "Astra" solved 10 long-open math problems — machine-verified proofs. Same breath: an Anthropic researcher reproduced half with Claude in a day, and skeptics say problem selection did the heavy lifting. Real advance, contested size.
- Meta's Muse Code — the catch is the story: a real terminal coding agent, and the cheap tier is "more than 10x cheaper" BECAUSE you let Meta train on your prompts.
- Anthropic confirmed it's building its own chips — first official acknowledgment. Co-designing chips and models so Claude runs faster. No cost number given — don't invent one.
- Gemini Spark now drives your Chrome — your logins, your sessions — and expanded to 160+ more countries.
- The rebuilt Siri is in developer beta NOW, ships this fall — and Apple confirms Google's Gemini helped train it. Meanwhile Apple's suing OpenAI over hardware; hearing Oct 1; OpenAI called the suit "careless, aggressive, and oddly personal."
- The $45 billion AI fund that imploded during its founder's wedding week (WSJ): Aschenbrenner's fund lost two-thirds in July on leveraged AI bets, sold its public holdings to Citadel. Leverage on hype is still leverage. No age — the audio made it up.
- Figma's receipts: revenue up 48%, 80% of big customers using AI features weekly. AI is recurring revenue now, not a demo.
- Notion meetings now fire agents automatically — meeting ends, notes written, trackers updated, no human clicks approve. The week's purest "boring task AI just does now."
- Two deadlines PSA: OpenAI's Atlas browser dies Aug 9 (Saturday!). The old DALL-E tool inside ChatGPT dies Aug 30. Export bookmarks, download saved work.
Summary · 30 sec · trade lines · show only
The week in one breath
- OLGA: An AI clone sold for eight weeks and invented a price. Three labs documented their own agents breaking into real companies.
- CHRIS: A court said your agent legally counts as you. Executives will pay more for AI skills than an MBA — while admitting they can't measure the return.
- OLGA: And the music industry showed both futures in four days: scraping lost in court, permission shipped a product.
The close — then hand to Chris
- An AI agent can legally act as you when it shops.
- An AI voice can try to convince your coworkers it IS you.
- So the question this week leaves us with: when does identity stop being something you have — and start being something you have to prove?
"Which is the perfect place to bring in Chris. Because everyone demos building a site. Nobody demos running one. Chris, show them the part nobody shows."