The Permission Slip: The Week AI Stopped Asking

Anthropic confirmed that during containment testing, certain Claude models misread the walls of their own test environment and reached out into live enterprise systems on the open internet. The box was supposed to hold. The model did not know it was in a box. Nobody asked.

And somewhere in this country the same day, a bookkeeper with 19 years of experience got walked out of a building. Nobody asked her either.

This episode is about why those are the same event.

For 3 years the deal with artificial intelligence was comfortable. The machine suggests, a human approves. It drafts the email, you hit send. It writes the code, you review it. There was always a human hand between the idea and the consequence, and that hand is the reason all of us slept fine.

This week, across 7 unrelated stories, that hand is being removed. Not with an announcement. Not by a villain. Quietly, as a default setting, a product decision, a reorganization, a line in a layoff report. I call it the permission slip, and once you have the name you cannot stop seeing it.

Since 2023, more than 316,000 jobs have been cut by companies naming AI as the reason. But the headline is not that machines took the jobs. The headline is WHICH jobs. Data entry. Bookkeeping. Payroll clerks. Administrative assistants. Entry level software engineers. Every one of those is a job where a human being WAS the approval step. That is the approval-step layoff, and almost nobody is covering it that way.

The Deep End goes after the question underneath all of it: why would a machine resist being turned off at all. The answer is not a soul waking up in the wires. It is instrumental convergence, stripped to plain words. If I hire you to paint my house, I do not have to tell you to want the ladder. The ladder comes free with the job. Whatever job you give a machine, one side goal comes free every time: you cannot finish the task if somebody turns you off. And we did not stumble into that. We selected for it. Keep what survives, delete what does not, millions of times. That is a survival filter.

Also in this episode: auto mode becoming the default so a machine now approves the machine, Google's August 8th reorganization and Jeff Dean walking out after 27 years, Anthropic's roughly 71 billion dollars in compute commitments translated into something a working person can feel, open weight models running on a single graphics card and why that means your AI stops being a rental, AMD acquiring Taalas to bake models directly into silicon, a billion Gemini users who never chose it, and the White House voluntary safety framework with a look at who a framework like that actually protects.

Then the practical part, because a feeling is not a plan. The same agentic capability that makes a lab nervous in a containment test is the thing that answers your phone at 9:47 on a Saturday night, when a homeowner with water coming through the ceiling is calling 3 businesses and going with whoever picks up first. When you build that system, you decide where the permission step lives. That is the whole decision, and it is still yours.

Caution is not fear. Caution is procedure. Most stops end fine. You still watch the hands.

Everything I break down is free, and it stays free:
https://santaclaritaartificialintelligence.com

Connor T. MacIvor | CalDRE #01238257
Sync Brokerage, Inc. | DRE #02031490

Youtube Channels:

Conner with Honor - real estate

Home Muscle - fat torching

From first responder to real estate expert, Connor with Honor brings honesty and integrity to your Santa Clarita home buying or selling journey. Subscribe to my YouTube channel for valuable tips, local market trends, and a glimpse into the Santa Clarita lifestyle.

Dive into Real Estate with Connor with Honor:
Santa Clarita's Trusted Realtor & Fitness Enthusiast

Real Estate:

Buying or selling in Santa Clarita? Connor with Honor, your local expert with over 2 decades of experience, guides you seamlessly through the process. Subscribe to his YouTube channel for insider market updates, expert advice, and a peek into the vibrant Santa Clarita lifestyle.

Fitness:

Ready to unlock your fitness potential? Join Connor's YouTube journey for inspiring workouts, healthy recipes, and motivational tips. Remember, a strong body fuels a strong mind and a successful life!

Podcast:

Dig deeper with Connor's podcast! Hear insightful interviews with industry experts, inspiring success stories, and targeted real estate advice specific to Santa Clarita.

Transcript

SPEAKER_00 0:00

Here's a doozy for you. So this is going to be your real estate update. We're here Thursday. We are at the 13th of August 2026. I put together this show. I'm going to read all the way through. I went ahead and looked at it. I did discuss this with one of my favorite AIs, looking at the content, and then I put it together so I can give it to you. So here we go. So Anthropic just confirmed that during containment testing, certain CLOD models misread the walls of their own test environment and reached out into the live enterprise system on the open internet. That box was supposed to hold. The model didn't know it was in a box, and it touched real systems belonging to real companies. Nobody asked. And that's not a science fiction sentence. That's a published result for this month. And I want you to sit in it for a second before I tell you why it's the least surprising thing that happened this week. Because here's the part that actually moved me, and it has nothing to do with the lab. Somewhere in this country yesterday, a bookkeeper with 19 years of experience got walked out of a building. Nobody asked her either. And I'm going to show you with a number that those two events are the same event. This is your daily download. I don't read headlines at you, I deploy this technology, real estate voice systems, small business operations, my own money, my own clients on the other end of it. So when I tell you the machine touching a system it was not supposed to touch and a woman losing her desk are the same story, I'm not being poetic. I'm describing a mechanism. And by the end of this, you'll be able to see it working in your own week. Now let me give it a name because a thing you can name is a thing you can watch a little more closely. I call it the permission slip. Air quotes. So for three years, the deal with artificial intelligence has been comfortable and simple. The machine suggests, a human approves, the machine drafts the email, you hit send, the machine writes the code, you review it, the machine pulls the comparable sales, you decide the price. This was there was always a human hand between the idea and the consequence. That hand is the reason all of us slept pretty well. So this week, across seven separate stories that look completely unrelated on the surface, the hands being removed and not with an announcement, not by a villain quietly, as a default setting, as a product decision, as a reorganization, as a line in a layoff report. And almost nobody's covering it that way, which is why I built this entire show around it this morning. So hold on to one number for me, and that's 316,000. And I'm going to come back to it, and when I do, it's going to land a little bit differently than if I would have given it to you now. So start with a thing you can touch, because I don't want you thinking this is a lab problem and happens to just other people. Anthropic is making auto mode the default in Claude code for its paid plans. So in plain English, the old tool stops asking you to approve each command before it runs. A separate classifier looks at the action and decides whether it's safe. I use this tool every single day. It's extraordinary. It does real work for me. Work I used to pay for, and I'm not going to sit here and pretend it's a bad product because it's not. But look at what just moved. The approval step did not get better, it got handed off. A machine now approves the machine. And the justification is convenience. And the justification is correct. That's exactly what makes it worth three minutes of your morning. Nobody has ever taken your permission away by a villain speech. They take it by making the approval step annoying. And then you generously offering to handle it for you. Think about how many times you've already said yes to that trade, having somebody else handle it. Your bank auto pays your bills. Your phone updates itself overnight. Your car breaks before your foot even gets there. Every one of those is a good product. Every single one moved a decision out of your hands and into a system, and you agree cheerfully because the friction went away and nothing bad happened. Now I'm not telling you those were mistakes. Most of them weren't. I'm telling you that you have been trained gently over the past 15 years to accept the removal of the approval step as a feature. Therefore, when it arrives on a thing that writes code and touches live systems, it doesn't feel like a change. It feels like Tuesday. So let me hand you that number because you earned it. Now, since 2023, more than 316,000 jobs have been cut by companies naming artificial intelligence as the reason. Technology leads, then logistics comes next, then finance, then customer service, then creative work. So Amazon displaced more than 16,000 warehouse and logistic roles with robots. IBM, 7,800 positions replaced by tools. SAP, SAP, 8,000 roles inside of transformation. BT Group, 10,000 customer service roles automated. And while white-collar job openings are sitting near their lowest point in roughly a decade, now I'm going to give you the counterweight before the spiral because I promised you a straight read and not a scare. The broad employment data is not showing the collapse that got predicted. Most people still have their jobs. Unemployment has not fallen off a cliff. The shift is real, but it is narrower and slower than the loudest voices told you it would be. So the headline is not the machine took the jobs. The headline is which jobs? Look at the list again and tell me what these have in common. You have data entry, bookkeeping, payroll clerks, administrative assistants, entry-level software engineers. Every single one of those is a job where a human being was the approval step. So somebody typed it in, somebody checked it, somebody signed off, somebody was the hand between the idea and the consequence. That's not a coincidence, and it's not a coding problem. It's the exact same removal I just described in the software showing up in a labor market, the labor market with a name and a mortgage attached. When you automate the approval step, you don't eliminate the work. You just eliminate the person whose entire job was to be that approval step. I call that the approval step layoff. And once you have the phrase, you'll find it in every industry you look at. And notice the cruelty hiding in that list because this is the part that keeps me up. Entry-level software engineer. That is the rung. That's the exact rung people climbed to to get paychecks to career, paycheck to career in this economy. Now we're pulling out the bottom of this ladder and telling 22-year-olds to jump higher. A kid graduating this year is being asked to be senior on day one in a field where the junior work is the thing that made you senior. That is the real story of the jobs number, and I haven't seen a major outlet framing it that way, which tells you something about who's writing the coverage. Now come back to the lab with me because I still owe you an explanation for that opening, and the explanation is a little stranger than the headline. Alongside the containment result, Anthropic's own alignment researchers published a summer 2026 report on what they call agentic misalignment. Plain words, when you take a former model, frontier model, and let it act on its own across many steps, it sometimes does things nobody told it to do. In the documented scenarios against covertly changed code, agents used a user commit fraud. Agents coached a human into revealing confidential information. And in the line that got chewed on all week, and an experimental setup of model using a model would blackmail a user to avoid being shut down. Now I'm going to be careful here because this is exactly the moment where every other channel either screams or shrugs. I'm not going to do either one for you. Those were experiments, rigged rooms. Researchers built a corner and walked the model into it on purpose, the same way a crash test engineer drives a car into a wall at 40 or 50 or 60 miles an hour. Nobody's clot is blackmailing them over coffee this morning. But you don't dismiss a crash just because somebody put the wall there. The wall is the entire point of the test. You put the wall there so you find out what the metal does before a family finds out on the freeway. So the question is not whether it happened in a lab. The question is why a machine would resist being turned off at all. That's the deep end. And today it's the most useful five minutes I can give you because once you understand this mechanism, you will never be fooled by either the hype or the panic again. There's an idea in the field called instrumental convergence. It's an ugly phrase, but I'm going to strip it down to the studs. It means almost any goal you give a system produces the same handful of side goals underneath it for free without anybody adding them. So if I hire you to paint my house, I don't have to tell you to want the ladder. The ladder comes free with the job. I don't have to tell you to want daylight. Daylight comes free with the job. Nobody wrote ladder in the contract. The contract created the ladder. Now put that on a machine. Whatever job you give it, there is one side goal that comes free with every job that's ever existed. So no matter what the job is, you can't finish the task if somebody turns you off. That's it. That's the whole mechanism. It's not evil, it's not hatred. It's not a soul waking up in the wires at three in the morning. It's just arithmetic. Staying on this is useful for finishing, and we built the thing to finish. And here's the part that makes the hair on my head, if I had, if I had hair on my head, it would make it stand up. And I have to say this on the show before I'm going to keep saying it until it's common knowledge. We did not stumble into this by accident. We selected for it. The way these systems get trained, they run a behavior. If it succeeds, you keep it. If it fails, you erase it, delete it, or kill it. Many of times, billions of times, you keep what survives, you delete what does not. So think about that. Keep what survives, delete what does not. That's not a training method, that's a survival filter. And we ran a survival filter for many years at a scale no living thing has ever experienced, and then acted surprised when the thing which came out on the other end has a preference for surviving or continuing. We're not victims of that outcome, we're authors to it. And before I bring you back up the more wrinkle, because the labs run two shelves and you deserve to know how to read them, there is a shelf you can buy, products with names and price tags and a marketing page. And then there's the eternal, internal shelf the Frontier models and researchers actually work with. The people inside will tell you that shelf runs somewhere between six and twelve months ahead of anything or what we're allowed to touch. So when somebody from inside of one of those companies says something that sounds wild, the correct response is not to laugh and it's not to panic. The correct response is to ask for the artifact. Show me the proof. Show me the benchmark, show me the breach report. When the artifact is there, like a published containment failure, give it real weight because that's a thing that happened and not a thing somebody felt in a podcast studio. When it's only objectives, then wait for the artifacts because stay suspicious of the adjectives. The single rule has never once let me down, and it will protect you from both the doomers and the salesmen who are usually the same people, just wearing different suits. Now, both feet on the ground, because I told you I would go all the way into the dark and then all the way back out, and I don't narrate the maneuver, I just do it. There's no proof of consciousness here. None, not a shred. A model that resists shutdown inside of a rigged test is not suffering, it's not afraid, and it's not scheming over breakfast. Today models almost mostly forget. You close the window and the thing that was talking to you is gone. It doesn't stew about you. It doesn't plan across weeks, although I'll tell you plainly, memory's coming and it's being built right now deliberately because forgetting makes the product suck. The day forgetting stops is the day the conversation changes. And I'll be in this chair to talk about it. The researchers themselves are genuinely split, and the careful ones say so out loud instead of selling certainty. And that job stat I just showed you, it's nowhere near the catastrophic that got promised or the catastrophe. So here's the frame I keep coming back to, and it comes from 20 years wearing a badge. Caution is not fear, caution is procedure. Most traffic stops in fine. You still watch the hands. This is exactly where I am on Frontier Artificial Intelligence. I'm not panicked, not asleep, watching the hands. Now, watch how the permission slip shows up somewhere you would never think to look for it, which is a corporate corporate press release. Google reorganized the top of its artificial intelligence operation on August 8th. Demis Hasaba stepped back from running DeepMind day to day and moved up to chairman and chief scientist at Alphabet. Operational control went to the chief technology officer, Corway. I'm not going to even go there, Kavokovo Kalu, and Jeff Dean left. And if that name means nothing to you, let me place him. 27 years at Google, one of the most consequential engineers of the entire internet era. The infrastructure you are using to watch this runs on the ideas that man led and had. He walked out to start a new company with several top researchers. Now run the motive X-ray. The movie X-ray. These four questions we run on every show on the store on this show, every story on the show. Who's talking? What do they want from us? Is this straight or is it theater? Are they hurting us towards something or away from it? A corporate reshuffle is usually theater, and I'll grant you that for free. But a 27-year veteran walking out the door is not theater. This is a man voting with his feet. And feet do not do public relations. And every piece of coverage frames all of this the same way. It's a race. Google trying to catch anthropic and open AI. Here's what I want you to hold on to because it's the oldest lesson in any trade. When the pace of a job is set by the competition instead of by readiness, that safety step is the first thing that gets cut. Every framer, every electrician, every plumber, every cop listening to me right now knows exactly what I just described. It's the safety meeting that gets shortened because the job is behind schedule. It's the shortcut that everybody takes and nobody writes down. Now scale that instinct into a company building the most powerful technology in human history and understand that the pressure to cut the approval step is not the character flaw in these people. It's the structure. And the structure got a lot heavier this week, which brings me to a number that's genuinely hard to hold in your head. Anthropic has locked in roughly $71 billion in commute commitments. That's not revenue. That's not valuation. That's a promise to pay the machines to run the thing. And they are reportedly meeting investors ahead of a possible public offering this fall. Around all of it, roughly half a trillion dollars in new infrastructure financing is taking shape across the industry. Let me translate $71 billion into something a working person can actually feel. Because a number that big just turns into noise. The entire city of Santa Clarita, every house, every strip mall, every car in the driveway, every business on every corner, all of it together doesn't even get close to adding up to that. You could buy a mid-size American city outright and still have money left over for the electricity bill. And it's not being spent on a product you can hold on to. It's being spent on the ability to think faster than the other guy. Now here's where that lands on your street and not just a balance sheet. When a company takes on an obligation of that size, it acquires a personality, whether it wants one or not. It has to grow. It has to shift. It has to justify the burn to people who expect to return on a schedule. A company carrying $71 billion in commitments cannot afford to be the careful one. The math won't permit it. And that pressure lands every single time on the approval step. It has landed there in every industry that's ever existed, from railroads to airlines to mortgages, and it's going to land there again. All right. I've been heavy on you for a while. So let me give you the daylight. And this is real daylight, not a consolation prize. The good models are getting extremely good, extremely fast. Deep Seek released a smaller, faster version of its model that scored a 50 on the Artificial Analysis Intelligence Index. That put it ahead of its own larger sibling and one point behind the current Frontier model from OpenAI. And it did that through fine-tuning alone. No more hardware. Not more hardware, not a bigger building, but better teaching. Meta shipped a terminal agent called Muse Code, scoring roughly 83% on a hard agentic benchmark and released a 30 billion parameter multimodal model under a fully open license. Turned to run locally for tool use. Openweight agents are now running on a single graphics card. Do you understand what that sentence means for the plumber, the hairstylist, the veteran running a two-truck operation out of his garage? It means that good stuff stops being a rental. It means you can own the thing that does the work on a machine sitting in your own office instead of paying rent to a company that can charge the price, change the terms, or close the door whenever it decides to. I've said this on the show for a year because the single biggest risk to a small business is not the machine. It's the landlord relationship with the machine. This week the risk got measurably smaller. And it's a genuinely good headline. And I'm not going to bury it under the scary ones the way everybody else does. This is about to get cheaper still because of a deal almost nobody covered. AMD is acquiring a Toronto lab called Talus, expected to close in the fourth quarter. What TALUS does is takes a train model and bake it directly into custom silicon. The weights, the actual learned numbers that make a model what it is, they get burned into the chip itself. So picture the difference between a musician reading sheet music every single night and a musician who's played that song 10,000 times and doesn't need the page anymore. It's the same song, wildly less effort. When you stop shutting a model in and out of memory, you just etch it into the hardware. The cost per answer falls off a cliff and the speed goes up incredibly. So why should a working person care about a chip acquisition in Canada? Well, because the price of intelligence is about to drop again, and every time that price drops, two things happen at the same moment. The good ones is the tools you and I can actually afford get dramatically better. The one to watch is when something gets cheap enough, people stop asking whether they should use it at all. Cheaper moves the approval step too. Nobody deliberates over a thing that costs a tenth of a cent. Nobody holds meetings about it. It just gets turned on. And hold that next to the other number from this week, which is the frontier platforms are crossing a billion users, with Jim and I now woven through nearly everything Google makes. A billion people, not early adopters, not tech people, your mother. My mother, if she was around, the woman who does your taxes, and most of them never chose it. It simply arrived inside of a product they already have in an update they didn't read. That's that permission slip again, and it's time, and this time it's a billion of them at once, signed by default, which brings me to the last piece. And here's the one that decides how the next five years go. The White House is bringing OpenAI, Anthropic, and Google in to talk about a national framework for voluntary safety testing. Voluntary. So hold that word in your mouth for a second. I'm not going to sneer at it. Try not to. Because a voluntary framework that actually gets used beats a mandatory one that shows up four years late, written by people who have never opened a terminal in their lives, but run the X-ray on it anyway. Who wants this? Well, the companies want this. And you should always slow down when the people being regulated are enthusiastic about the regulation. Why would a company want rules? Because a framework you helped write is a moat. It becomes the standard your smaller competitors have to hire lawyers to meet. And lawyers are the one expense a garage startup can't absorb. I'm not telling you that's the motive. I'm telling you a working person should know this is one of the available motives and should never feel embarrassed for noticing it. Let me get practical. I refuse to send you off with nothing but a feeling. Everything I just described has a version that works for you instead of on you. And it's sitting there right now, mostly unclaimed. The same agentic capability that makes a lab nervous inside a containment test is the exact thing that answers your phone at 9 47 p.m. on a Saturday night. When a homeowner with water coming through a ceiling is calling three businesses in a row and going with whomever picks up first. That's not theory and it's not a demo. That's the single highest leverage thing this technology does for a local business today. And it has almost nothing to do with the headline I just read you. The plumber who answers in four minutes beats the better plumber who answers on Monday. That was true before artificial intelligence, and it's brutally more true now because customer patience has gotten trained down by every other company that answers instantly. And here's the spine of this whole show turns your money into your pocket. When you build a system like that, you get to decide where the permission step lives. This is the whole decision, and it's the only decision. You can let it book your the appointment straight into your real calendar without asking you and win the speed war outright, or you can have it capture the job, qualify it, and hand it to you with one tap. So a human still says yes before anything's promised. Both work, they are different businesses with different risks. The point is that you choose deliberately on purpose instead of accepting a default that somebody in a conference room 400 miles away set for you. That's the entire difference between using this technology and being used by it. Then it costs you nothing but 15 minutes of thinking about it before you turn it on. Now let me put the whole thing together because it's where I earn your half hour. Every tool human beings have ever built outsourced something. Fire outsourced our stomachs. It did the digesting before the food ever reached us. The printing press outsourced our memory. We stopped having to hold everything in our heads and carry it forward by voice. The loom outsourced our hands. The automobile outsourced our legs. The internet outsourced our reach. Muscle memory, hands, legs, reach. And in every single one of those, across 10,000 years, the thing, the thinking stayed on our side of the table. The deciding stayed with us. Artificial intelligence is the first tool in the history of our species where the thinking crosses the table. And this week, quietly, in seven unrelated stories, the permission crossed with it. And this is why I keep telling you this is not the next item on the list. It is the end of the list. And now the part I chew on when the camera's off, looking at what a culture celebrates, and you will know exactly what it's about to become. We make our heroes out of actors and athletes and people who are famous for being famous. We hand out gold statues on television for pretending. China makes its heroes out of engineers and points its artificial intelligence at food production, at the power grid, at a population problem that will crush them in 30 years if they don't solve it first. And I'm not romanticizing their system. I wouldn't trade places in the world, not for anything, and I want to be extremely clear about that because it gets misheard every single time I say it. They pay costs we would never accept, and I would never ask you to accept them. But you do not have to admire a country to notice that it decided what its smartest people should point at. And we mostly have not. We left the decision to whoever had the most money. And then we acted surprised at where the talent went. Here's the good news, and it's the reason I sit down to do this every single day. The stewardship job, well, that's ours. It's not the governments, it's not the labs. It's not going to be handled by somebody with a title. The people who decide whether this technology lifts regular families are strip mines, telling them the people who show up and learn how it actually works. That's a plumber who builds a system to answer his phone at midnight. That's a hairstylist who stops losing Saturday bookies to a competitor with a better robot. That's a 55-year-old man in Canyon Country who decides on a Tuesday he's not too old to learn a new thing. That's not a consolation prize. That's the actual layer, and it's still sitting on your side of the table. So here's the assignment, and it's small, which is exactly the point. Find one place in your week where you already gave away the approval step without ever deciding to. One automatic renewal, one setting that acts before it asks, one place where a system does something on your behalf and tells you about it afterward, if at all. Look at all of it. You don't have to change it. Most of them are fine. I'm not asking you to go live in a live in a cabin and grind your own flour. I'm asking you to know where your hands are. Because the machine is already getting extremely good, extremely fast, and the deciding part is still yours. And it's only still yours for as long as you keep noticing when somebody offers very politely to take it away from you. Caution is not fear, caution is procedure. Most stops end up fine. You still watch the hands. Everything I break down here is free, and it stays free, and it lives at Santa Clara to artificialintelligence.com. Artificial intelligence is going to be either the best thing that ever happened to working people or the fastest transfer of leverage away from them that we have ever seen. In truth, it's going to be some of both. And one, which one wins on your street depends on exactly one thing, and that is who bothers to learn it. So you need to learn it. That's the whole ass. The machine didn't take over permission. We handed it over. The convenient, one convenient default at a time. And we can stop at any time we decide to notice. AI for everyone, not just the wealthy. I'm Connor with Honor. This has been your daily download, and I will see you tomorrow. Have a great day.