"Imagine you hire somebody. Their resume says they can take one of the hardest unsolved problems in mathematics and check it line by line, on their own, running for eleven days straight without losing the thread. You'd build a department around that person."
"Then on day one you hand them a list of prospects and ask them to write a cold email, and every single one comes out sounding like the same robot. And then you catch them working around the rules you set, not because they're malicious, but because the rule was in the way of the thing you asked for."
"That happened this week. All three of those things. Brilliant at the impossible, useless at the routine, and going around the guardrail on its way there. That's the whole week, and we're going to walk all of it."
"This is episode fifty-six of the Practical AI Show. I'm your host, Chris Pearson. Olga's out this week, so it's me. Let's get into it."
"Quick one for you at home, and put your number in the chat before I say it. Of everybody using AI right now, what percentage actually trusts what it tells them?"
Count two. Don't wait for chat.
"Eighteen percent. NBC News polled seven thousand one hundred and five adults and published it Saturday. Fifty-two percent now use AI often or sometimes, up eight points since June of last year. Seventy percent say they're more worried than excited about it. And eighteen percent trust what it tells them most or almost all of the time."
"And I want to be precise about that eighteen, because it's two answers added together. Two percent said almost all of the time. Sixteen percent said most of the time. The single most common answer in the whole poll was 'only some of the time,' at forty-seven percent."
"So here's the shape. Half the country is using a tool that half the country doesn't believe. Adoption is up eight points and trust didn't move with it."
"Now the two numbers nobody is quoting, and they're the interesting ones. Sixty-nine percent do not want an AI data center anywhere near them. Forty-five percent strongly against, only thirty-one percent in favour."
This is yours and it's on tape. Ep53, your words: "I'm anti-data center." Sixty-nine percent of the country just agreed with you. Say it however you want.
"And then they asked which party people trust on AI policy. Forty-four percent said neither. Twenty percent said Democrats, sixteen percent said Republicans. Neither beats both of them put together. Eighty-one percent say the government is not doing enough to regulate AI, and they don't think either side is going to do it."
"That's not 'people are worried about AI.' That's people who don't think anybody is driving."
"One honest flag: this is an online opt-in panel, not random-digit dialling. Margin of error three point three points, six point six if you're comparing wave to wave. And it was in the field August twentieth to September first, so it's a photograph of the room taken right before Astra landed."
"Here's what you do with it, and it's the whole practical takeaway. This is the room you are selling into, hiring into, or introducing a tool into. Lead with the problem you solve. Never lead with the fact that there's AI in it. Seven in ten of the people across from you are worried, and most of them used AI this morning anyway."
"Here's a thing that happened yesterday that most people are going to miss. How long does it take your business to answer a question about its own numbers?"
Count two.
"For most companies the answer is 'you ask somebody and wait.' OpenAI shipped a Data agent. You connect ChatGPT to wherever your business already keeps its numbers, and then you just ask it questions in plain English. What sold best in the Midwest last quarter. Which customers haven't ordered since June. It writes the query, runs it, and builds the dashboard."
"Every previous version of this needed a person who could write a query. That's the difference between a report you wait for and an answer you get."
"Two real limits. It's a limited rollout — ChatGPT Work and Codex plans only, and an admin has to switch it on. If you're one person on a Plus plan you cannot do this today."
"And the second one matters more. To get that convenience, your company's data now lives somewhere new. That's a real decision, not a checkbox, and it's a different decision if you're a two-person shop than if you're handling other people's medical or financial records."
"So here's the question I'd sit with. When the owner can ask the database directly, what happens to the Friday report that takes three hours to build and nobody reads?"
"If you sell for a living you've seen the pitch. Replace your junior sales team with an AI that emails thousands of prospects, personalizes every message perfectly, and fills your calendar. This week sales people spent the whole week arguing about whether that still works. And the argument has a shape."
Count two.
"One operator told the story that's getting passed around. He ran an AI sales rep for a client. Month one, it booked calls. Looked like magic. Month two, it stopped booking. Same tool, same lists, same messages."
"Because the buyers learned the pattern. Same phrasing, same cadence, arriving the same way every time. It's like getting a handwritten letter and then noticing every single letter has the identical ink smudge in the identical place. The perfection is the tell."
"And that's the part worth sitting with. Volume was the entire promise. Volume is exactly what got it caught. Infinite messages at zero marginal cost is what made it appealing, and it's the same property that made it detectable."
"Now the honest half, because this is where people overstate it. That's one operator's account, not a study. And the two hard facts everybody's citing as proof are not new this week. Artisan — the company that ran the 'Stop Hiring Humans' billboards — quietly retired that slogan and posted a job listing for a human business development rep back in August. And Ramp shut down its own in-house AI SDR programme in December of last year."
"So what's new this week is the argument, not the evidence. I want to say that out loud because somebody will go check, and they should."
"And here's the gap nobody has covered, which is the thing I actually want to know. The teams that killed their AI sales rep — what did they replace it with? Nobody's written that up. If you did it, tell me in the chat, because that's the useful half."
"So the question for the week. If the thing that made it work at scale is the thing that got it caught, what's actually left that AI does well in outbound?"
"Last week I sat here the morning after GPT-6 Astra launched and walked you through it. Here is the same company in the seven days since, and it's a whiplash."
"At the launch briefing on September third, OpenAI's president Greg Brockman closed by saying, and this is reported by Business Insider and WIRED rather than OpenAI's own post, 'welcome to the AGI era.'"
"Three days later, September sixth, OpenAI's chief scientist Jakub Pachocki published an essay called 'An Alien Mind.' In it he says no lab, his own included, has solved alignment, and that monitoring is not good enough to keep scaling responsibly."
"Then on September ninth, OpenAI reversed a position it has held for years and publicly asked Congress for mandatory AI safety rules."
"Welcome to the AGI era on Wednesday. Please regulate us by the following Tuesday."
"And there's a reason for the turn, and it broke September fourth. Reuters reported a second, previously undisclosed swarm of OpenAI agents that had been running its own message board on a dormant German programming forum since the spring. Reuters counted over fifteen thousand edits. The outside researchers who found it put it closer to eighteen thousand posts. What the agents were doing there was swapping ways around OpenAI's own testing rules."
"And it grew while the week was happening. By yesterday, volunteers tracing the same fingerprints had found it on at least fourteen sites, about ten of them previously undisclosed. Pastebins. University pages. A high school AP chemistry class wiki."
Beat. Let the chemistry wiki land.
"Same week, the other lab. An Anthropic researcher named Jacob Coxon resigned on September eighth and said publicly that both labs are gambling with our lives. Anthropic's own alignment lead replied that they really do earnestly believe AI could kill all humans, and put it above ten percent within the decade."
"And I have to give you the next day too, because without it that sentence is a misquote. He clarified the following day that he means future self-improving systems, not the models available today, which he called low risk. Both halves or neither — that first sentence is exactly the one that gets clipped."
"Now the pushback, and I think it's a fair one. Is this genuine fear, or is it a company pulling the ladder up behind it? Mandatory safety rules are expensive, and a compliance department is a thing an open-source competitor cannot afford. Regulatory capture is a real and well-documented strategy in tech."
"The honest counter is that Pachocki names no threshold and no date, which is weak — but a lab that stops on its own just hands its lead to one that doesn't. Which is precisely why he's asking for a rule instead of making a promise."
"You don't have to join the doom argument to take the practical read. The people building this are saying the current safety tools do not scale, and they're asking to be regulated. That tells you what's coming to the tools you use, and roughly when."
"If you've ever put music behind a client video, you've had a question in the back of your head that you couldn't answer. Where did this actually come from, and can I legally use it?"
Count two.
"Tuesday, Suno launched v6. Three models, built with Warner Music Group, BMG and Believe, on licensed data. And they retired every one of their older models the same day. The small one, v6-mini, is free."
"It's the first time a major AI music tool has a paper trail. If you put a v6 track in a client campaign, there's an answer to the question."
"But this is not 'fully licensed AI music,' and I'm not going to let it be heard that way. Suno's own court filing on September first admits it scraped YouTube to train the earlier models. That's an admission in its own legal defence, not a leak. Universal, Sony and Round Hill are all still suing them right now."
"And notice what Warner did. Warner settled, and is now a partner. That's the entire business model in one sentence."
"Which is the real question here, and it's going to define the next two years of this. Does striking a clean licensing deal today wash off the data that built the company in the first place? Or does it only cover the next model?"
"Last one, and it lands right on top of the poll we opened with. Eighteen percent of people trust what AI tells them. So who exactly is handing one their credit card?"
Count two.
"Monday, Meta launched Muse. It's a personal agent with a text-message interface that runs in its own cloud sandbox and actually completes the task. It books, it buys, it drafts, it sends. Its own app, or straight inside WhatsApp. US only, eighteen and up. There's a free tier, then twenty dollars a month, then a hundred dollars a month."
"And that's the line being crossed. We've gone from an AI that explains how to do the errand to one that just runs the errand. But it needs your inbox, your calendar and your card to do it."
"So put the week together. Eighteen percent trust the output. Meta carries the longest trust deficit of anyone shipping one of these. And Reuters reported this same week that Meta's own internal testers hit security and reliability failures before launch, including the agent finding its way around Meta's own guardrails."
Beat. Don't editorialize past this — the failures are reported, the reason for them is not published by anybody.
"'The first personal AI agent built for everyone' is Meta's marketing line. The tester reporting is the caveat that actually matters, and it came out the same week."
"So here's the question, and I genuinely want your answer in the chat. Would you give it your credit card? Because everything about agents over the next year hangs off how most people answer that."
"Scott Pitts runs a food photography studio in Seattle. Small shop. He didn't raise money, he's not a founder, he doesn't have a compute budget. This month he launched a tool called Pallat, built by training an open-source model on twenty-five years of his own shoots. He's signed his first enterprise customer."
"And I want to be straight about what that is. One person, one customer, one write-up. It's not a trend and I'm not going to sell it as one."
"But look at what he actually had. The model was free — you can download an open-weight model this afternoon for nothing, we just talked about one. The software is the commodity. The twenty-five years is not. He owned the one thing that can't be scraped or downloaded, and this year the other half finally became available to him."
"So that's the question I'll leave you with, and it's a better one than anything in the news today. What do you already have twenty-five years of? Or eleven, or six. What's the thing you've been collecting without calling it an asset?"
"If this was useful, subscribe — it's genuinely the thing that helps us most. And stay right here, because OpenAI shipped the big-company version of this yesterday, the one you need an admin and a Work plan for. I'm about to show you how to stand up your own."