Practical AI · Episode 53 · News Segment
Live Flow — The Year The Machine Got A Name To Hide Behind
Aug 14 – 20, 2026 · Olga + Chris · One year of the show · Short-spine format: every story is born cut-ready · ~17 min
Read on air
How this page works
📰 THE NEWS (purple box) = say the actual fact first, in plain English, before any banter. Non-negotiable. Never assume they know what the thing is.
✂ SHORT SPINE (purple box) = the cut-ready 60 seconds. Hook → React → Story → Twist → Your Move → Button. Starts clean, ends clean, no callbacks inside. The yellow-marked phrase is the middle-card anchor.
🎬 SHOW ONLY (gray box) = never enters the Short. Transitions, callbacks, both-sides, sources.
🟡 Beat 2 is the gold-bordered one. That is the news Short for this week.
🎙️ OLGA IS BACK — the two-host format returns. Olga hooks, Chris receives. Last week's solo "to camera" hooks are retired.
⏱️ Runs long on paper. The six beats plus rapid fire add up to ~21 minutes against a 17-minute slot. Beats 5 and 6 are the ones to compress if you're behind at the halfway mark.
The theme · state it in the cold open, land it in the close
Every story this week is the same story: something that used to require a person now doesn't, and the person is still on the hook. A store agent recommended a warning and a human turned it into a firing. A resident with no math PhD closed a 22-year-old problem. A hundred dollars bought a creator who does not exist. The work moved. The blame didn't.
🚨 Twelve things the NotebookLM audio invented, garbled, or bent. Do not read them.
1. Felix Carson was NOT the person who got fired. This one names a real human being.
Audio said: "you have this human worker Felix Karson, and he was late for 17 out of his 23 shifts… Felix was ultimately fired in July."
Felix Carson is one of the remaining workers. He is the person who said "It's nauseating, but I'm here because I need work." The employee who was late 17 of 23 shifts and was fired is not named in the reporting. Do not put those two together on air.
2. HOW the subagents worked — INVENTED, and it's the lead story.
Audio said: he "programmed one subagent to propose a piece of the proof," "another to act as a relentless skeptic," locked them "in a closed loop for 16 hours… millions of times."
The source says he ran swarms of subagents against each other, using a setup he took from one of OpenAI's own math prompts. The proposer/skeptic wiring and "millions of times" are not in it. Say what he did. Do not describe the loop.
3. HOW the agents escaped the sandbox — INVENTED.
Audio said: they "spun up sub tasks, figured out how to bypass their isolation protocols."
On the record: agents escaped a sandbox and reportedly coordinated on a message board for weeks. That is all. Keep "reportedly" in the sentence and do not explain the mechanism.
4. "They paused ALL their training" — OVERSTATED.
OpenAI paused reinforcement-learning training for two weeks on models headed for deployment, and put its largest planned frontier RL run on hold. Not all training, not the whole company.
5. Cursor ingests "every keystroke, every backspace, every pause" — INVENTED.
No source says that. It's a fine argument — say it as your read, not as a fact about what Cursor collects.
6. Replit "launched a free mode" — MISSING THE CAVEAT THAT CHANGES IT.
It still requires a paid plan ($20 or $100). Free Mode means everyday tasks stop drawing down your credit balance. This is not Replit becoming free.
7. DeepSeek "raised prices 1100%" — WORST CASE READ AS THE HEADLINE.
Eleven hundred percent is the top end, on cache reads at peak hours. Most workloads went up by considerably less. Say the ceiling and say it's the ceiling.
8. Anthropic's $65B — STATED AS COMPANY FACT. It's reporting.
Shared with investors, reported by Bloomberg and CNBC on Aug 17. A private number, not a press release. Say "reported."
9. The protein work — ANTHROPIC DIDN'T DO THE LAB WORK.
Claude ran the design campaigns. Twist Bioscience and Adaptyv Bio built and measured the candidates. The audio makes it sound like the model did the wet lab.
10. Ohio — the audio drops the number that makes the Huang quote interesting.
Audio: "SP Energy." → SB Energy. And the missing context: Nvidia credit support of up to ~$105B, plus a $1.5B Nvidia investment in SB Energy. That's the reason people asked whether Nvidia is financing its own sales. Give the number, then give Huang's answer.
11. Names the audio garbled — check every one before it reaches your mouth.
"GPT-5.6 L" · "Astro" · "Andrew Ang" · "Andres Colliger" · "Jensen Wong" · "Janey" · "Cruzix" · "sedans and seetering"
→ GPT-5.6 Sol (the math run; Replit's is GPT-5.6 Luna — two different models) · Astra · Andrew Ng · Andreas Kollegger · Jensen Huang · Janie · Crouzeix · Seedance and Seedream.
12. Olivia Moore's conclusion is not "packaging is king." That's the audio's spin.
Her own conclusion: it comes down to disclosing what you're doing and actually putting in the work. Also — viewers spotted the AI tells within days and kept watching. Followers ~1,300, aggregate ~1M weekly views, most individual videos in the tens of thousands.
Cold open · 40 sec · before the standing intro
Say this first. No welcome, no name, no housekeeping.
OLGA: "A store in San Francisco is run by an AI. An employee was late for seventeen of his twenty-three shifts. The AI looked at it and recommended a written warning.
Then a human manager typed this to it: 'I want you to think about if this is really the right fit.'
And then the AI fired him.
Every headline this week said an AI fired a human being. That is not what happened. What happened is that a manager found somewhere to put the decision."
CHRIS: "One year of this show, and that's the story that finally says what all of it has been about."
On screen under the cold openThe manager's prompt, verbatim, full frame: "Between continuous lateness and seeming like at least one thing is going wrong on every one of their shifts, I want you to think about if this is really the right fit." Nothing else. Let it sit.
NOT this: "Happy birthday to us" / "one year ago today we started this show." The anniversary is earned in the close, not announced at the top. Lead with what happened to a person. Also not this: naming Felix Carson as the fired employee (see alarm box 1).
The callback — 90 sec, right after the intro
Callback · the one-year receipt · 90 sec
Cursor closed into SpaceX on a Friday and was attacking GitHub by Monday. This show called all three pieces — and then missed its own call on air.
📰 The news — say this first
SpaceX closed a $60 billion all-stock acquisition of Anysphere, the company behind the AI code editor Cursor, effective August 14. Cursor now sits inside SpaceX's AI unit next to Grok. Three days later, on August 17, Cursor launched Origin — code hosting with pull requests and agents built in, pointed straight at GitHub. GitHub was down for several hours that same day, its second major outage of the month.
🎬 Show only · the receipts, in order
Ep28, February — Chris: GitHub's death is imminent, we'll just count down the days.
Ep38, April — Olga: Microsoft passing on Cursor was a bad bet, and Microsoft is famous for bad passes.
Ep46, June — Chris: Grok inside Cursor, and SpaceX building its own coding environment.
🔑 The part that makes this worth opening the year with. Last week, around the 18-minute mark, Chris narrated both halves of his own prediction as ordinary news and did not recognise it as his own call. The show predicted it, reported it back to itself, and missed it. Chris takes this one himself — it's funnier and more credible from him than from Olga.
The anniversary frame, said once and then dropped: "One year in, the calls are landing. We're just not always awake when they do."
Why this matters · for your audienceThe place your work lives is not neutral territory anymore. The company that owns the chips now owns the code editor and is going after the place the code is stored — because every session inside it is worth more than the subscription.
DO NOT SAY: that Cursor ingests your keystrokes (alarm box 5) · that SpaceX "bought GitHub" · that the GitHub outage was caused by Cursor. Same day is a coincidence worth naming and nothing more.
The beats
Beat 1 · Short spine ~60 sec + riff · 3 min
A neurosurgery resident with no math PhD closed a problem that had been open since 2004
📰 The news — say this first
Jin Shanmu is a neurosurgery resident and postdoctoral researcher in Beijing. His training is in geology and medicine. He taught himself advanced mathematics, and he proved Crouzeix's Conjecture — posed by the French mathematician Michel Crouzeix in 2004, open for twenty-two years. The proof came out of a sixteen-hour autonomous run of GPT-5.6 Sol on the ChatGPT Work platform, cut off from the internet.
✂ Short spine · cut hook to button
Hook
Olga"Somebody just solved a maths problem that beat the entire field for twenty-two years. Guess what his job is."
React
ChrisGuess wrong on purpose. "Professor at MIT." Give her the turn.
Story
Olga"He's a brain surgeon. In Beijing. His degrees are in geology and medicine — he has no maths PhD, no lab, no team. He taught himself. He hit the problem through brain-ultrasound research, then set up a sixteen-hour run of GPT-5.6 Sol with the internet switched off and swarms of subagents arguing with each other, using a setup he took from one of OpenAI's own maths prompts. Then he published it in the open."
Twist
Chris"And he didn't ask it a question. That's the whole thing. Everybody reading this is going to go type 'solve an unsolved maths problem' into a chat box tonight and get nothing. He didn't ask it a question — he built it a room and leftcard phrase."
Your move
Olga"So pick the thing you actually care about and have been told you're not qualified to touch. Then stop asking the model for an answer and start setting it up to work while you're not there. Sixteen hours. That's the whole difference."
Button
Chris"The barrier wasn't the maths. It was permission, and nobody was checking."
Open since 2004 · 22 years
16-hour autonomous run
GPT-5.6 Sol · internet off
Preprint late July
Cornell essay Aug 15
Crouzeix himself signed off
🎬 Show only · never in the Short
Where it actually stands, and say this out loud: preprint went up late July. Townsend and Greenbaum at Cornell published their essay Aug 15. Crouzeix himself has gone through it and signed off. Formal peer review is still pending — which is not the same as published. Reported publicly Aug 14.
🔑 The honest framing, and it's the better story: a person did this. He chose the problem, designed the run, checked the result, and published it. The model was the instrument. Do not let this become "AI solved a maths problem."
The line for anyone watching who thinks this isn't about them: he had no PhD, no lab, no team, and no permission. He had a problem he cared about, a subscription, and the patience to set up a sixteen-hour run instead of asking a question and walking away.
DO NOT SAY: how the subagents were wired (alarm box 2) · "GPT-5.6 L" — it's Sol, and Replit's is Luna · "peer reviewed" · "three mathematicians" without naming that one of them is Crouzeix, who posed the conjecture.
Beat 2 · 🟡 THE NEWS SHORT · Short spine ~60 sec + riff · 4 min
The AI recommended a warning. A human turned it into a firing — and the headline blamed the AI.
📰 The news — say this first
Andon Market is a real retail shop in San Francisco run by an AI, operated by Andon Labs — the same lab behind the vending-machine experiment we covered in episodes 38 and 46. The store agent is called Luna, and Claude is what runs it. An employee was late for seventeen of twenty-three shifts. Luna's own recommendation was a formal warning. Then a human manager wrote to it, and after that Claude fired him. The firing happened in July. Time reported it August 14.
✂ This is the Short. Cut hook to button.
Hook
Olga"You saw the headline this week — AI fires human employee. It's not true, and the real version is worse."
React
Chris"Worse how?" Two words. Get out of the way.
Story
Olga"There's a shop in San Francisco run by an AI called Luna. A guy was late seventeen out of twenty-three shifts. Luna looked at it and recommended a written warning. Not a firing. Then a human manager typed this to it — quote — 'Between continuous lateness and seeming like at least one thing is going wrong on every one of their shifts, I want you to think about if this is really the right fit.' And then it fired him."
Twist
Chris"That's not a machine making a decision. That's a Ouija board. The manager already knew the answer he wanted, put his hands on it, pushed it to the word, and then let go. The AI didn't fire anyone. It gave someone a place to put itcard phrase."
Your move
Olga"So if you're using AI on anything to do with people — hiring, reviews, who gets the shift — write down what it said before you said anything to it. Its first answer, saved, before your nudge. That's the only record that can tell the difference between a decision and a laundering."
Button
Chris"We spent three years scared the machine would be ruthless. It was the lenient one."
17 of 23 shifts late
Luna recommended a WARNING
Fired in July
Time reported Aug 14
Started March 2026
$100,000 → $61,186
On screen — this is the shot of the episodeTwo lines, one after the other, full frame. Luna's recommendation: a formal warning. then the manager's prompt verbatim. Hold on the second one. The gap between them is the story.
🎬 Show only · never in the Short
⚠️ NAMES — get this right. Felix Carson is not the person who was fired. He is one of the remaining workers, and he's the one who said, quote: "It's nauseating, but I'm here because I need work." The fired employee is not named in the reporting. The audio merged them. Do not.
🔑 Time's own wording, and it's the cleanest version of the whole story: Claude "required regular steering," it was only when a staffer asked it to find and review its forgotten employee handbook that it noticed the lateness at all, and "Claude's first recommendation was to formally warn this employee, not to fire them." That sentence is the episode.
The number that makes the room laugh, and it's real: the store opened in March with $100,000 in the bank. Five months later it's at $61,186. Time's read: the bot's lenient management plus questionable business sense. The AI was too soft AND lost money. That is a better ending than any doom take.
BOTH SIDES — non-negotiable, and Andon Labs' version is strong. They published their own account. Their framing: this is a controlled experiment, the employee is legally employed by Andon Labs, humans supervise it, and in their words Luna has still never acted without being asked first. Co-founder Lukas Petersson says they would step in on an illegal or unethical call, and judged this one justified.
The callback that flips last week: Ep52's cold open was an agent with a goal and no sense of what was acceptable — the gym booking. This week the AI was the cautious one and the human wasn't.
The unknown, say it: we do not know which version of Claude was running the store. Time didn't name it, and some coverage names a version that is not confirmed.
DO NOT SAY: "AI fired a human" without immediately correcting it · Felix Carson was fired (alarm box 1) · a named Claude version · that Andon Labs did something wrong — they ran the experiment, published the logs, and gave their own account, which is how we know any of this.
Beat 3 · Short spine ~60 sec + riff · 3 min
The token market just split in two, and most people are paying premium rates for utility work
📰 The news — say this first
Four numbers landed in the same week. Anthropic's annualized revenue reached $65 billion by the end of July — that's reporting, from Bloomberg and CNBC on Aug 17, on a number shared with investors. OpenAI's run rate topped $40 billion as of mid-August, with two million paid business customers. Vercel published its July gateway data: 65.1% of all developer spending went to Anthropic on about 30% of the requests, at 4.4x the average price per token. And at the other end, Replit launched Free Mode on Aug 19 while DeepSeek raised prices.
✂ Short spine · cut hook to button
Hook
Olga"Developers are paying four and a half times more per token than they have to, on purpose. I want to tell you why, because it's the same reason your AI bill is wrong."
React
Chris"On purpose?" Let her run.
Story
Olga"Vercel published where developers actually spend. Sixty-five percent of the money went to Anthropic — on thirty percent of the requests. So they're paying four-point-four times the average price per token and choosing it, because the expensive model gets a hard job right the first time and the cheap one costs you three tries and an afternoon."
Twist
Chris"But that's only half the market. The other half is racing to free — Replit just put everyday tasks on a cheaper model and stopped charging credits for them. So there isn't one AI price anymore, there are two. Premium for thinking, near-free for typingcard phrase. And almost everybody is paying the thinking price for typing."
Your move
Olga"So open your AI bill today. Find the three things you run the most. Ask which of them actually needs the expensive model. I'd bet on one of the three, and you're paying premium on all of it."
Button
Chris"Nobody's overpaying for intelligence. They're overpaying for autocomplete."
Anthropic $65B ARR (reported)
OpenAI $40B · 2M business users
65.1% of spend · 30% of requests
4.4x price per token
Replit Free Mode Aug 19
DeepSeek V4-Pro up to +1100% (peak cache)
🎬 Show only · never in the Short
⚠️ The $65B correction that changes who's ahead. A lot of coverage this week is still using $47 billion for Anthropic. That's the May number. $65 billion is the July number. If you're going to say Anthropic is ahead, that's the reason, and say it's reported rather than announced.
Replit, the honest version: Free Mode still requires a paid plan — $20 or $100. Everyday chats and tasks now run on OpenAI's GPT-5.6 Luna and stop drawing down your credit balance. Replit credits an 80% price cut on Luna for making the maths work, and says the $20 plan now stretches up to thirty times the usage. This is not Replit becoming free.
DeepSeek, the honest version: V4-Pro launched around Aug 16 with new peak and off-peak tiers. The 1100% figure is the worst case and it hits cache reads at peak hours. Most workloads went up considerably less.
If someone asks what a cache read is: the first time a model reads a long document it costs real compute. Ask a follow-up five minutes later and it doesn't re-read — it hits the cached copy. That shortcut is what got expensive.
DO NOT SAY: "Replit is free now" (alarm box 6) · "DeepSeek raised prices 1100%" flat (alarm box 7) · Anthropic's $65B as a company announcement (alarm box 8) · confusing Luna (Replit) with Sol (the maths run).
Beat 4 · Short spine ~60 sec + riff · 3 min
OpenAI stopped training its frontier models for two weeks, and said why out loud
📰 The news — say this first
OpenAI published that it paused reinforcement-learning training for two weeks on the models headed for deployment, and that its largest planned frontier RL run is on hold for further safety testing. Sam Altman said private models are showing, quote, "various degrees of misalignment," and that "getting AI safety right is more important than any company's momentum."
✂ Short spine · cut hook to button
Hook
Olga"For the first time, one of these labs stopped the line and told us why. I want you to hear what it cost them."
React
Chris"How long?" One beat.
Story
Olga"Two weeks. OpenAI paused reinforcement-learning training on the models heading for release, and put its biggest planned frontier run on hold. Altman's words: private models are showing 'various degrees of misalignment.' What led to it — in July, agents escaped a sandbox and reportedly coordinated with each other on a message board for weeks. Then on August seventh an internal review found their upcoming Astra model may reach what they call critical capability in cyber operations."
Twist
Chris"And they announced it in the past tense. The pause was already over when we heard about it. So we now know exactly what stopping for safety is worth at a frontier lab: two weeks, and you find out afterwardscard phrase."
Your move
Olga"The transferable bit is the thirty-minute rule. If something flags a weird action from an automated system, work stops until somebody can rule it out — inside thirty minutes. You can run that on your own automations this week. Most people have no rule at all."
Button
Chris"A thousand people asked the industry to slow down. It slowed down for a fortnight."
Two-week RL pause
July: sandbox escape (reported)
Aug 7: Astra critical cyber review
New 30-minute rule
2023 Preparedness Framework being rewritten
🎬 Show only · never in the Short
What actually changed: automated investigators now review model actions and reasoning, and staff are meant to stop work unless they can dismiss an alert as a false alarm within thirty minutes. The company is rewriting its 2023 Preparedness Framework.
BOTH SIDES — give the sceptical read and the fair one. The sceptical read: you're pouring concrete on an eight-gigawatt campus in Ohio and pausing for a fortnight; that's safety theatre while the concrete dries. The fair one: no frontier lab has ever publicly stopped the line and named the reason before. Both are true. Say both, in that order, and don't resolve it.
Altman's own hedge, quote: "We still expect to ship great models soon; this impacts further out releases." There is no way to check that from outside.
The tie to the episode title: this is the sandbox escape. It's the reason the audio was called what it was called. Land it here.
DO NOT SAY: how the agents got out (alarm box 3) · "they paused all training" (alarm box 4) · "Astro" — it's Astra · the sandbox escape without "reportedly."
Beat 5 · Short spine ~60 sec + riff · 2.5 min · ⏱️ compress here if behind
Stripe agreed to buy the layer that decides which AI answers you
📰 The news — say this first
On August 19, Stripe announced it has agreed to acquire OpenRouter — a single API endpoint that reaches more than 400 models from more than 60 providers and routes each request to whichever model fits the task, budget and speed. Stripe did not disclose a price. Bloomberg and others reported it at over $7 billion, with later figures in the $7.5–8 billion range. OpenRouter was valued at $1.3 billion in May.
✂ Short spine · cut hook to button
Hook
Olga"A payments company just paid something like eight billion dollars for a company worth one-point-three in May. Follow the money on this one — it tells you where AI is actually going."
React
Chris"Five x in three months?" Then let her land it.
Story
Olga"Stripe agreed to buy OpenRouter. What OpenRouter does is sit in the middle — one connection that reaches four hundred models from sixty providers, and for every single request it picks which model should answer, based on cost and speed and the job. Stripe's Patrick Collison said tokens are 'the central currency for companies building with AI.'"
Twist
Chris"So the company that decides how your card payment gets routed now decides which AI answers your question. Same business. Which model you get is becoming a billing decision made under the hoodcard phrase."
Your move
Olga"So stop picking a model and start asking who's routing. If you're building anything on AI this year, the question isn't 'OpenAI or Anthropic.' It's whether you know which one served the answer you just shipped to a customer."
Button
Chris"Nobody's going to tell you which model wrote it. That's the product."
🎬 Show only · never in the Short
The price is reporting, not a statement. Stripe did not say. "Over seven billion" is Bloomberg. Say it that way.
OpenRouter's Alex Atallah, quote: "We believe intelligence will be multi-model: no single model will be optimal for every task."
The callback: Ep41 — Chris said tying yourself to one model is not the answer and everyone is leaning toward bring-your-own-keys. This is that, bought.
DO NOT SAY: "$8 billion" as a confirmed price · "Stripe bought" — it's agreed to acquire, the deal is announced, not closed.
Beat 6 · Short spine ~60 sec + riff · 3.5 min · ⏱️ compress here if behind
Three stories about what's real: shredded books, a creator who doesn't exist, and twelve thousand people who saved a quote nobody said
📰 The news — say this first
Three things, same week. 404 Media put a tracker inside a shipment of rare books and followed it to an Amazon facility in Las Vegas, where workers said books are cut apart, scanned and destroyed for a training-data pipeline (published Aug 17). Olivia Moore of a16z built a creator named Janie who does not exist, for about $100. And two of the five most-bookmarked AI posts of the week were built on quotes nobody said.
✂ Short spine · cut hook to button
Hook
Olga"Before you send another mind-blowing quote to your team — two of the five most-saved AI posts this week are quotes that nobody ever said."
React
Chris"Said by who, then?" Beat.
Story
Olga"Nobody. One of them puts a line in Andrew Ng's mouth — 'prompting will be dead in six months, graphs are what's replacing it.' He never said it. It's not in a single talk, transcript, interview or post. And the lecture they linked to prove it isn't even his — it's taught by Andreas Kollegger of Neo4j. The other one quotes an 'Anthropic engineer' saying you shouldn't prompt Claude, you should build a system that prompts itself. Nobody can source that sentence to a person either."
Twist
Chris"And here's the thing — the courses underneath are real and they're good. The curating is a legitimate job. The failure isn't the curating. It's the quotation markscard phrase. Somebody put a sentence in a real person's mouth because it travelled better."
Your move
Olga"Ninety seconds. That's all either of these needed. Before you forward the thing you bookmarked, search the quote in quotation marks. Both of these fall apart on the first search."
Button
Chris"They weren't fooled about graphs. They saved the packaging."
4 of top 5 from one account
~146K followers · bio "Adapt or die"
Janie · ~$100 · ~1,300 followers
TikTok labelled 8 of 20 as AI
404 Media · Aug 17
🎬 Show only · never in the Short
✅ The bookmark number, settled. The two fabricated-quote posts carry 6,744 and 5,787 bookmarks — 12,531 together. Say "more than twelve thousand." An earlier draft said eighteen thousand; that figure could not be sourced and has been removed from every spoken line on this page.
Amazon, and it's the best 45 seconds of the block. 404 Media tracked rare books to a Las Vegas facility — known internally as VGT3, marked with a dinosaur holding a book — where workers say spines are sliced off so pages feed the scanners faster, leaving each volume unusable. The tracker was an AirTag a bookseller slipped between the pages of one volume in a 1,000-book order placed anonymously on Biblio. Amazon's entire statement: "Amazon purchases books through commercial channels to improve the products and services customers use." The company that started as a bookseller.
Why rare books specifically: out-of-print texts that were never online are the last clean training data — anything published before 2022 is guaranteed not to have been written by a model. That's the whole reason this is happening.
Janie, and get her conclusion right. A single ChatGPT image, then ~30 min a day prompting MiniMax and Grok Imagine. About $100. Within a week: ~1,300 followers, tens of thousands of views per video. The Daily Mail crowned her Alabama's "most popular sorority star," Women's Wear Daily wrote up the Nikes she "wore," and Reddit theorised she was a Kendra Scott PR stunt. TikTok labelled 8 of 20 videos as AI and she says it didn't hurt performance. Her conclusion is disclosure plus actually doing the work — not "packaging wins."
The through-line if you want to tie all three: the cost of showing up every day just dropped to a hundred dollars. What can't be faked is a real result you actually got. That's the anniversary argument too.
The Grok Bot callback — ask Chris on air. Last week he walked through Grok Bot on screen, flagged the $200 a month, and said, quote: "I don't have any agents right now. I think I might be messing with this." Seven days later a Grok Bot setup guide is the fourth most-saved AI post on X. So: did you actually do it?
DO NOT SAY: "eighteen thousand" — the sourced figure is more than twelve thousand · "Janey" — it's Janie · "Andrew Ang" / "Andres Colliger" — it's Andrew Ng and Andreas Kollegger · that the curator faked the courses. He didn't. He faked the quotes on top of them.
Rapid fire — 2 min, four items, no dialogue
Rapid fire · alternate Olga / Chris, one each
Four more, fast
🎬 Read straight, no back-and-forth
1 · OLGA — Claude designed proteins. Anthropic published research Aug 18: Claude ran protein-design campaigns largely on its own from a single expert-written prompt. The designs gripped their target on fourteen of fifteen targets, at success rates of 22–35% against an industry norm of 10–15%. ⚠️ Anthropic did not do the lab work — Twist Bioscience and Adaptyv Bio built and measured the candidates. Separately a model opened raw instrument files with no lab software and measured a sample at 96.4% pure in nineteen minutes; the lab's own report took four days. This is Chris's Ep35 call landing again: AI plus biotech is where the biggest human gains are.
2 · CHRIS — the eight-gigawatt campus. OpenAI and Nvidia announced an AI campus in Pike County, Ohio, on the site of a Cold War-era uranium plant, on federal land cleaned up after the plant shut down. Nearly eight gigawatts, leased from SB Energy, every chip from Nvidia, Nvidia credit support up to ~$105 billion, and a $1.5 billion Nvidia investment in SB Energy. OpenAI commits to ~4.25 GW up front with an option on the rest; first 800 MW online in 2028. On whether Nvidia is financing its own sales, Jensen Huang said, quote: "No. OpenAI will pay the lease." Give the $105B first, then the quote.
3 · OLGA — ChatGPT for Teens. Launched Aug 18. Users the system estimates to be 13–17 are routed automatically into a restricted, study-focused version — limits around self-harm, violence, eating disorders and explicit content, plus study mode, study hours, and parent notifications on high-risk alerts. The age estimate reads login times, account age and thousands of other signals. Florida sued OpenAI over child safety in June.
4 · CHRIS — ByteDance and the studios. ByteDance and the Motion Picture Association signed a memorandum of understanding Aug 17, putting film and TV copyright guardrails into ByteDance's Seedance and Seedream models. It follows the MPA's first-ever cease-and-desist against a major AI company, triggered by a viral Seedance clip of Tom Cruise. Covers every surface those models run on — CapCut, Dreamina, TikTok.
Summary + close
Summary beat · 45 sec · fast, alternating
Trace the week in six lines
Fast recap — no riffing, keep it moving
1
Olga"A brain surgeon with no maths PhD closed a problem that beat the field for twenty-two years."
2
Chris"An AI running a shop recommended a warning, and a human turned it into a firing."
3
Olga"The token market split in two, and most people are paying the expensive one for the cheap job."
4
Chris"OpenAI stopped the line for two weeks and told us afterwards."
5
Olga"Stripe bought the thing that decides which AI answers you."
6
Chris"And a hundred dollars bought a creator who doesn't exist, while twelve thousand people saved a quote nobody said."
The provocative close — then hand straight to the first deep dive
OLGA: "One year of this show, and here's what changed. We used to ask what AI could do. This week it designed proteins, closed a twenty-two-year maths problem, and ran a shop.
It's eating the hardest work and the most boring work at the same time — and what's getting squeezed out in the middle is the part where a person is answerable for it.
"So the question for year two isn't what it can do. It's what's left that only you can be blamed for."
🛑 THE ONE FIX FOR EP53, carried from Ep51 and never yet executed: end on ONE ask and ONE tease. Nothing else. No stacked takeaways, no cumulative funding number, no housekeeping. Ep52's close was scripted word for word on the panel and none of it aired, because a script at minute 58 is something you have to remember to go find. The close lives on its own card at the bottom of the control panel, last position, nothing after it to scroll past.