Episode Transcript
[00:00:00] Sam, Every profession has a word it doesn't use lightly.
[00:00:35] Doctors say malignant until they're sure, right? Engineers don't say structural until they've run the numbers. Pilots don't say mayday for turbulence.
[00:00:48] And financial regulators, the most careful, most understated people in public life, from my experience, do not say the word systemic.
[00:01:00] Systemic means the problem isn't inside the system.
[00:01:06] Systemic means the problem is the system.
[00:01:11] It's the word for the thing that, well, if it breaks, it doesn't break alone.
[00:01:18] It takes the pension fund with it, the payment network, probably the paycheck.
[00:01:26] And last week, on three different days in three different capitals, that word got attached to artificial intelligence. And it's not by critics and it wasn't by activists. It was by the people whose entire job is to never say it unless they mean it.
[00:01:48] And Monday was Washington, Tuesday, Frankfurt, and by the end of the week, London, with the force of a law behind it.
[00:01:59] So in our episode this week, what happens when the machine becomes too big to unplug?
[00:02:08] That's what we're going to look at.
[00:02:12] So welcome to AI Today. I'm your host, Dr. Alan Badot. If you've been with us the last few weeks, you know the arc that we've been walking two episodes ago, we talked about the power and permission, the discovery of AI's real constraints and that they've moved, you know, really beneath the software down to the power grid and really the permission structures of government.
[00:02:37] Last week we talked about chain of command, autonomous agents taking real actions in the real world, and authorization layer nobody built before even shipping them.
[00:02:52] Tonight's the third act, because this week the story has moved again. It's moved from who controls the machine to the harder question, who's holding the risk when the machine stumbles?
[00:03:08] So let's. Let's roll it back a little bit.
[00:03:11] Three regulators, five days, one word, right? Let's start in Washington because Washington, it went first, and I don't think Washington meant to do it.
[00:03:20] On July 6, the news outlets, you know, notice.
[00:03:25] Obtained a draft report from inside the U.S. treasury Department. And it was not a press release, okay? It was not a policy paper written for the audience. It's an internal analysis prepared by career staff.
[00:03:40] The that. And now, you know, these analysts, it's really their entire job to watch the financial system for the Treasury Department. And of course, the chair of the Federal Reserve, Right?
[00:03:53] But the people who work at the same desk, they've been there. It doesn't matter who wins the election, okay? And so here's what the analysts concluded the AI industry now shares key structural features with the dot com bubble of the late 1990s. Now if you've been paying attention to the market, you know, you've heard this from others, okay, Heavy concentration on a number of small firms, right? Deep reliance on private market financing and an enormous amount of capital sunk into physical infrastructure. We've talked about that before, but the data centers and everything that goes into them and really it was on a bet to, you know, that the future is going to, future demand is going to justify that.
[00:04:43] But then the report says something that the dot com comparison can't hold, right? The analysts found that today's AI firms are woven more deeply into the American economy than the dot com companies ever were.
[00:04:58] That's not surprising though, right? Because the impact that AI has, it can reach every single part of your organization. But because of that, a downturn wouldn't just stay in the tech sector.
[00:05:10] It's going to move.
[00:05:12] And again, not surprising. How many times have you heard me talk about, I can apply our Janus platform any vertical sector, no issues whatsoever.
[00:05:25] But think about what the change is coming, right? And so you think about the chain of events.
[00:05:37] An AI downturn hits the stock market, that's your 401k potentially, if it hits the private credit. And we're going to watch lending markets that had financed these build outs and these buys are going to start to feel the crunch, right? If it hits companies that borrowed billions to pour concrete into those, those data centers to support them, it's going to hit those as well, right? And so it can hit the cloud providers, the chip makers, everywhere in between, even the power companies, Right?
[00:06:07] Because eventually that bill is going to arrive at your house. But that's really what systemic means.
[00:06:13] It means there's no longer a door that you can close.
[00:06:19] Now analysts, of course, like they like to be, were pretty careful in how they were framing this. This is night, not 1999 all over again. All right? Many of the biggest AI companies, they've got maturity, they've got profitability, they're carrying some pretty strong balance sheets.
[00:06:40] Nothing like of course, some of the speculation that was going on in the dot com era. But the report I want to highlight, this does not predict a crash, right?
[00:06:52] What it predicts is exposure.
[00:06:55] If growth disappoints, companies pull back, confidence drains and the whole economy grows slower.
[00:07:03] That's fundamental.
[00:07:06] But two details in that report, I think they deserve a, a little bit more attention because they're the ones that nobody is talking about.
[00:07:18] And the first one is who's holding the bag.
[00:07:25] As it's changed in the DOT era, it was really the retail investors, day traders, regular people who lost money they couldn't see.
[00:07:38] Now this time though, seems to be fewer retail investors in the game, which means the exposure is concentrated in institutions, okay, Pension funds, insurance companies, banks, really the load bearing walls of the financial system itself.
[00:07:59] Second, where the money came from has changed.
[00:08:07] So the dot com boom was really financed in public markets where every price was visible every day.
[00:08:17] Much of the AI build out is, you know, financed through private credit where lending happens away from the spotlights.
[00:08:28] Far less oversight of course, but it's money also that is coming from a lot of different sources. Now the trouble comes through the public market, you watch it arrive. When it comes through a private one, you find out afterward.
[00:08:46] And there's a clock inside of those balance sheets too, right? We all know that the biggest cloud companies have poured almost a $trillion into AI infrastructure over the past two years.
[00:08:59] That's public record.
[00:09:01] The money's already spent, but by accounting rules the cost hits their books gradually as you know, really as that equipment depreciates, that wave concentrates over the next three years.
[00:09:19] And the bill for this build out has not been paid yet. It's just been scheduled almost like you put, put everything on layaway.
[00:09:28] And here's the part that makes this story about power and not just money.
[00:09:39] That report was finished a week ago.
[00:09:44] It was prepared for the Treasury Secretary and the Fed chair.
[00:09:50] It's been sitting there for a while, unreleased, unapproved, while the administration's public posture stays fully bullish.
[00:10:00] In late June the Treasury Secretary stood in New York and really praised the industry for putting three quarters of a trillion dollars into AI buildouts this year.
[00:10:12] When the report leaked, a Treasury spokesman dismissed the findings as unvetted, but not the, of course not the department's official view.
[00:10:26] So hold both of these in your hand, okay?
[00:10:30] The official position, full speed ahead.
[00:10:33] The staff analysts financial system now rests on AI meeting its promises for productivity and profits.
[00:10:42] One government, two answers.
[00:10:45] And the honest one wasn't necessarily the, the one they published.
[00:10:53] That was just Monday.
[00:10:56] On Tuesday, an ocean away, a different regulator stopped writing reports and started setting deadlines.
[00:11:07] That's after the break, so stay with us. We'll be right back.
[00:11:40] Welcome back. I'm your host, Dr. Alan Bedot and we are talking about too big to unplug this week, really the third part of our series is on really accountability.
[00:11:52] Trying to go from the large enterprise down to how it Affects you and your day to day lives. And this week we're looking at a couple of different things. We talked about what happened on Monday last week. Let's talk about what happened on July 7th in Frankfurt, Germany.
[00:12:10] The European Central bank, the institution that supervises the biggest banks in the Eurozone, sent a letter to the chief executives of every single bank that it manages, and it's 110 banks.
[00:12:27] And signed at the top, the ECB's chief supervisor, Claudia Boont.
[00:12:36] The instructions inside were very simple.
[00:12:40] And those were build a formal action plan to defend your bank against AI powered cyber attacks.
[00:12:49] Name who's accountable, fund it, staff it, put dates on it, and submit by October 31st, a Halloween deadline.
[00:13:02] Interesting.
[00:13:04] I call it coincidence, but hey, supervisors don't do coincidences, especially in the EU right now. Why?
[00:13:14] What's changed? Central banks have been warning about cyber risk for 20 years. Okay, why does this letter exist now?
[00:13:23] Why does it have a hard date on it? That's what we wanted to try to figure out.
[00:13:28] But the reality is the math of an attack has changed.
[00:13:35] The letter lays it out.
[00:13:37] The newest frontier AI models find weaknesses in software and build a working attack.
[00:13:44] We knew that, right? We talked about that.
[00:13:47] Stable and Mythos and all those other things. You know, of course, the technical term exploits, right? And they can do it in minutes or hours, okay?
[00:13:58] Work that used to take a skilled human, you know, certified ethical hacker, or I guess I would say an unethical hacker would take a few days or weeks to do it.
[00:14:12] Certified ethical hackers like myself, we would try to help build up and you know, get help, help these institutions prepare for stuff like this.
[00:14:21] But you know, there's a second number though that I want you to think about because, you know, I think it's going to surprise a lot of people.
[00:14:32] And that's, you know, regulators now warn that these models can take a security patch, you know, which is the fix. The vendors go out, you see it all the time. You've got to update your Windows machines or your, your Apple devices or stuff. But they can reverse Engineer in about 30 minutes and find a hole in it.
[00:14:52] That's not new, okay, but now it's getting out there. People are starting to understand those, those things.
[00:15:00] And if you think about it, what actually does.
[00:15:04] Every defense we've built assumes a human speed clock, okay?
[00:15:12] Monthly patch cycles, quarterly reviews, annual penetration tests. The ECB is telling 110, 10 banks, that clock is gone.
[00:15:22] The window between a flaw being discovered and a flaw being used has collapsed your defenses that you often Hear about their design for the old window. And we talk about this, you know, we've talked about how a zero day threat isn't found for three months, six months, whatever that is. So it's just been sitting there, but it's still a zero day threat.
[00:15:48] It's just three months old.
[00:15:52] Yeah, yeah.
[00:15:58] So as you think about that, as these become more explicit, you start to realize that treating AI as an insider threat is more appropriate, even now that we have agents, than it was three months ago.
[00:16:18] And so again, as we have talked about these things, and I have been screaming on the hide, you know, for, for years almost, they're starting to, to come to fruition. Now.
[00:16:31] The supervisor was explicit that this is not about any single product. Okay. She called it a lasting shift in the threat landscape, not a temporary phenomenon. But everyone in that room knows the event behind the letter, right? Earlier this year, a frontier model. We're not going to say who it is, but you all know who it is.
[00:16:55] It's very Greek in its, in its name, you know, and access to it now is tightly restricted. But it's a little brother model, is available now, but it demonstrated that, you know, it could uncover thousands of previously unknown vulnerabilities across major operating systems, flaws that have been sitting there for, for years, and it could discover them and exploit them.
[00:17:23] Eurozone's banks, by the way, are currently excluded from access to that model. They're being told to defend against a class of capabilities they're not even allowed to hold tough.
[00:17:39] Now, products like ours, like Atlas, are built specifically for these things to, to track, to hunt first instead of defend first. Let's get in front of these things. It's a mindset change, right? It's not okay to sit there and be passive in your defense, which is why the first thing we do is launch HUNT agents to go out and figure out what's going on. Find your own vulnerabilities first, then use that information and train.
[00:18:13] Those are the kind of model shifts that we're starting to see. And it's not because, you know, we want these things to happen. It's just because they're happening and the AI is evolving every single day.
[00:18:25] Now let me take a little breath, all right?
[00:18:30] On the very same day that letter went out though, the European Systemic Risk Board, which I had to look this one up, right? That's the body whose only job is watching for threats to the financial system as a whole, issued a formal warning about systemic cyber risk from AI frontier models.
[00:18:55] Their scenario is the one that keeps well it keeps me awake at night and it keeps all the supervisors awake as well. And that's a large scale automated cyber disruption that doesn't take down just one system's bank, it takes them all down. It shakes public confidence, it disrupts payments. It starts the oldest event in banking.
[00:19:21] A run.
[00:19:23] There's that darn word again, right? Systemic.
[00:19:27] That's the second time in two days that they saw now three things worth knowing about and, and how this is landing, okay? Because it's important. So one, this isn't just a European problem, okay?
[00:19:42] The euro area subdivides the largest, you know, American banks and you know, the big, some of the biggest names on Wall Street. So it's very interconnected, but they are still supervised by the ECB in some cases. And they face the same similar October deadline as any bank in Frankfurt or Paris.
[00:20:07] The second thing is, is that the regulators are not aligned on their approach.
[00:20:13] And the split is kind of revealing, right? So the same day the letter went out, the government of the bank of England, he called it sensible, but said that Britain would take a less prescriptive path.
[00:20:25] Not edicts, conversations.
[00:20:28] And in Washington, the Federal Reserve's top supervisor officially emphasized a lighter touch. Framing the risk of AI as the risk of not adapting it to it fast enough.
[00:20:44] Three regulators, three postures.
[00:20:46] Frankfurt sets deadlines, London has some meetings, and Washington cleared the road right now. Number three.
[00:20:55] And this is the tell, this is what I want you to really pay attention to. The ECB cleared the calendar to make room for this, they postpone their annual, you know, technology risk questionnaire by five months.
[00:21:11] They offered to ease inspections.
[00:21:16] Supervisors do not lighten the paperwork load out of kindness.
[00:21:22] They do it when they've decided. One, risk matters more than everything else on the schedule.
[00:21:31] And buried in that letter is a working assumption that tells you how seriously they actually mean it. The ECB instructs banks to plan as if the perimeter will be breached.
[00:21:43] Not might, but will.
[00:21:46] Segment your networks, verify every user, every device, every connection, continuously.
[00:21:54] Trust nothing by default. That's not a posture you adapt to for a threat you expect to be able to repel.
[00:22:02] That's a posture that you adopt for a threat you expect to survive.
[00:22:08] Very different.
[00:22:12] So step back and look at what this letter really is.
[00:22:20] It is, it is a stress test, but not of capital.
[00:22:24] This time.
[00:22:27] It's speed.
[00:22:29] Can your institution survive an attacker that thinks in seconds and notice the shape of it too? Right? The banks get stress tested against the AI punch the models throwing the punch where that's their October 31 deadline.
[00:22:51] The intelligence itself sits outside of the perimeter as it happens to be examined by no one, submitting no other action plan.
[00:23:04] Interesting.
[00:23:06] So Europe found one answer for that, but it's not for the models, it's really for the machines underneath them.
[00:23:14] And it took effect yesterday with the force of law.
[00:23:20] That's next.
[00:23:21] So stay with us. We'll be back after a few messages from our sponsors and we're going to really deep dive into the letter and its implications to, to you all.
[00:23:59] Welcome back. I'm your host, Dr. Alan Bideau and we are deep diving into really some chain of events and some, some words that we heard last week that normally are not said very loudly at all and especially not three days in a row. Right.
[00:24:19] So in London on Thursday, and that was July 10th, you know, his magistrate treasurer published four names. Okay. And yesterday, Monday the 13th, those names came under a kind of supervision that until really this week was reserved for banks.
[00:24:42] So Microsoft, Google, Cloud, AWS and Oracle.
[00:24:50] The UK formally designated all four as critical third parties to its financial system.
[00:25:01] Yeah, it's the first ever use of powers Parliament created back in 2023, where, you know, powers written for exactly this moment and then held in reserve.
[00:25:15] But from yesterday forward, the bank of England, the Prudential Regulation Authority and the Financial Conduct Authority supervised these companies directly, not through the banks that rent their servers, through the providers themselves.
[00:25:36] Regulators can demand information, you know, assess the resilience, write specific rules for them and them only and enforce those rules.
[00:25:50] Sit with the category shift.
[00:25:53] These are not financial companies.
[00:25:56] They sell computing.
[00:25:58] And the combined value of those parent companies is north of $10 trillion, which is bigger than the economy of every nation on earth, except for two.
[00:26:13] And as of yesterday, three financial regulators in London can walk through their doors the way they walk into a bank.
[00:26:23] Wow.
[00:26:26] Why? You might say one word, you already know it. Concentration. Right.
[00:26:33] Nearly every bank, insurer and market infrastructure in Britain runs on one of those four clouds, which means the failure of a single provider is no longer one company's outage, it's a national event.
[00:26:53] The UK just said out loud what everybody in the industry has known for a decade. And that's cloud infrastructure isn't a vendor relationship anymore, it's a load bearing national infrastructure.
[00:27:10] And the UK's not alone.
[00:27:13] European Union designated a list of critical infrastructure providers, you know, under its own law that they had last year. And the perimeter of really financial regulation is moving down this stack. It's not moving back up.
[00:27:30] But here's the irony, nobody at those four Companies will say out loud.
[00:27:41] Well, they may not mind, though, right? The markets barely moved on the news because designation cuts two ways. Yes, it. It brings compliance costs and inspectors and more rules and stuff, but it also builds a wall.
[00:27:57] Once a regulator has certified your infrastructure as fit to carry a nation's financial system, the fifth company, who's the challenger, they're not competing on price anymore.
[00:28:14] Right.
[00:28:15] They're competing against the government. Signature supervision, as it turns out. And we all know this, it's also a moat. You get pretty sticky if you're one of those four companies.
[00:28:26] It's pretty, pretty good.
[00:28:29] Now, how do I connect that to last week? Right, because that's what we're talking about.
[00:28:34] Well, because the fourth document of this remarkable stretch landed with almost no coverage, and it's one that speaks directly to what we discussed on this program a week ago.
[00:28:51] The intelligence and really the cyber security agencies of five nations. The Five Eyes. Now, that's a word you haven't heard me say before, but it's really the us, the uk, Canada, Australia and New Zealand jointly published guidance on deploying agentic AI in critical infrastructures and for defense, the alliance, you know, we talk about it, really. We share a lot of intelligence information. We share a lot of data.
[00:29:22] It is a very open community when it comes to those. Those five from an intel perspective and heck, even, you know, we host some of their nation's best scientists, they host some of ours, and it's really a cooperative that has put us ahead of many others from an intelligence gathering and sharing capability.
[00:29:43] But they organized every danger of autonomous agents, really, into five formal categories of risk. Okay.
[00:29:54] Privilege, design and configuration, Behavior, structural. And guess what? Accountability.
[00:30:01] Nothing should shock you all about that, right?
[00:30:05] Last week, though, we said industry shipped autonomous agents before it built a chain of command.
[00:30:13] Before anyone could trace an agent's actions back to the human authority that approved it, we called the missing piece. What?
[00:30:25] Called it accountability.
[00:30:28] Yeah. This week, five governments wrote accountability into a formal risk taxonomy for the Western Alliance.
[00:30:38] Their prescription, Deploy incrementally, govern strongly, monitor continuously, and keep a human in the loop.
[00:30:48] It's not a courtesy, but a control.
[00:30:51] Yes, thank you. Yes, thank you.
[00:30:55] So when your editorial thesis shows up in a Five Eyes document seven days later, one of two things is true. Either we got lucky or the logic was always going to arrive there.
[00:31:09] Because it's the only place that logic can go.
[00:31:14] Now, before we assemble the week. Okay, you know, I want to bring this down to a scale, or really from the scale, I would say, of nations to the scale of one person.
[00:31:28] Because everything that we've discussed tonight, supervision, speed, someone signing up for the risk has a version that lands on a single human being.
[00:31:40] And it's a version that nobody's protecting yet, right? And it's really a shame.
[00:31:46] So think about it this way. If your face, if your work, if your likeness is stolen and it's spread online, whether it's a clone photo, a fabricated image, you know what they call revenge porn? A piece of content that you have never consented to, the law hands you exactly one tool.
[00:32:10] And it's called a DMCA takedown.
[00:32:14] And it was written in 1998.
[00:32:17] And really what it is, it's a notice, take no notice and take down system.
[00:32:25] You find the infringing copy or image, you file it. You wait one address at a time. A signature era remedy, right? Built for a web that moved at human speed.
[00:32:41] Then you wonder why.
[00:32:43] You know, sometimes it's a battle that you can never win. You feel like you're always fighting uphill. You feel like you have something taken down and by the time that, you know, you get back to the office, it's already back up somewhere else.
[00:32:58] It's because the machines don't move at human speed anymore.
[00:33:02] A stolen image can be replicated.
[00:33:05] It can be altered and reposted across a hundred platforms before the first takedown notice is even read.
[00:33:14] Which means the takedown by itself, it's really a mop chasing a flood.
[00:33:23] So content has changed.
[00:33:26] You know, you think that you're safe, you think that nothing is going to get out. And then all of a sudden, something else has popped up and you may not even know about it, at least not know about it until somebody sends it to you.
[00:33:42] So the real protection, it's moved, it's moved upstream and it's, you know, identity searching continuously, scanning where your likeness appears.
[00:33:56] So you find the theft in hours instead of months.
[00:34:01] Detection before spread, right? Prevention before takedown.
[00:34:05] It's really the same shift the regulators made this week, but at a national scale.
[00:34:12] Don't wait for the failure and file the paperwork.
[00:34:19] Supervise the damn thing continuously and catch it while it's still small.
[00:34:24] So if you see something you have not seen it before, figure out where it came from, figure out if there's something that you can do and fix it. It's the same thing.
[00:34:37] Nations got their continuous supervision this week, really, and it was written into a law. Individuals, though, are still filing forms from 1998, and that gap is the next story. And the tools really to close it are Already being built, so assemble this week, okay? Four documents, four days. Washington the financial exposure, Frankfurt. The system security posture of AI is now systemic, but plans are due by Halloween. London infrastructure under AI is systemic, supervised as of yesterday. The five eyes, the agents, how they're handling risk and accountability.
[00:35:26] But if you're the individual, you're still filling out forms.
[00:35:30] So you've got money, you've got security infrastructure authority. Really four angles, one conclusion.
[00:35:37] And not one of those documents was written by a technologist.
[00:35:46] And that raises the question that we came here to ask.
[00:35:49] Every load bearing thing in a modern economy has someone who underwrites it.
[00:35:55] Banks have capital requirements, bridges have inspectors, power plants have operators with licenses.
[00:36:03] Who underwrites the machine?
[00:36:06] When we come back, we're going to close out this show and I'm going to give you some ideas I want you to, to plan for and to take with you. We'll be right back.
[00:36:35] Foreign.
[00:36:44] Welcome back. I'm your host, Dr. Alan Bideau and I want to spend our last segment on a word we've been circling all night without naming. Okay?
[00:36:56] And the word is underwriting.
[00:36:58] Underwriting is a.
[00:37:00] It's an old idea. So really it's older than computers, older than the banks that we know.
[00:37:07] But it's something.
[00:37:10] Before something valuable you put at risk, someone signs their name beneath it, right?
[00:37:18] That's literally what the word means, that before something valuable is put at risk, you sign your name to it to write under.
[00:37:28] In the shipping houses of London, you know, three centuries ago, merchants would write their names under a ship's manifest and by signing they accepted the loss if the ship went down. The signature was the system.
[00:37:44] You knew who paid because you knew who signed.
[00:37:50] Everything we call financial regulation today is a descendant of that signature.
[00:38:01] Capital requirements are a signature.
[00:38:04] Deposit insurance is a signature. A stress test is a regulator asking a bank. Show me your signatures. Any damn good.
[00:38:14] Now look at the week through that lens.
[00:38:22] The treasury analysts in their unreleased reports were saying an enormous ship has been loaded. Trillions in valuation data centers and debt and everything else. We can't find a signature.
[00:38:36] The system rests on AI meeting its promises and no one formally accepting the loss. If it doesn't, the ECB and it's, I'm going to call it the Halloween letter was saying the attackers now move faster than any signature we've ever collected.
[00:38:56] Show us a new one with names, with budgets, with dates and by Halloween.
[00:39:04] But the UK went the furthest, right? Then it walked down the stack, found four companies the whole system physically stands on and Made them sign a document, direct supervision, course of law.
[00:39:20] So as of yesterday, in one country, the cloud has an underwriter.
[00:39:25] Interesting.
[00:39:27] 3 Responses, all notice you know what's going on. And think about, though, what all three have in common.
[00:39:39] Not one of them regulates the intelligence itself.
[00:39:45] That's curious.
[00:39:48] They regulate the money around it.
[00:39:50] They regulate the banks near it, the infrastructure under it.
[00:39:56] So the perimeter keeps tightening and the center stays empty.
[00:40:04] But there's a reason for that, and it's worth, like I always try to do, I try to tell you both sides of the story. I want to be as honest I can be so that you understand where the context lies.
[00:40:18] You cannot underwrite what you cannot predict.
[00:40:23] An underwriter needs to know the odds.
[00:40:27] A signature under a probabilistic system, a system that may do things different tomorrow than it did today and can explain either.
[00:40:42] It's not a signature, it's hope.
[00:40:47] Well, technically, it's notarized hope.
[00:40:52] Yeah.
[00:40:54] This is the part, really of the entire show.
[00:40:58] And I'm going to tell you what I think. Right. Like we always do. The last three weeks on this program have been. I've told it in three separate ways. The constraint moved beneath the model, really, into power and permission. The authority moved beneath the structure. Agents acting within, you know, a boundary, but with no chain of command.
[00:41:23] And now that risk is moved into the foundation, systemic, according to three regulators in five days.
[00:41:37] Every one of those stories ends at the same door.
[00:41:43] So somebody has to be answerable.
[00:41:47] Not the model.
[00:41:49] Model can't sign.
[00:41:51] Not the vendor marketing department, because guess what? Enthusiasm isn't collateral.
[00:41:59] A named human or a named institution with something real to lose standing under the system with a pen.
[00:42:07] The engineering answer exists, and you've heard me describe it and its shape many times before.
[00:42:16] Systems built to be deterministic matters.
[00:42:22] Systems that give the answer to the same question, same answer.
[00:42:30] They can show their reasoning, and that keeps a human accountable.
[00:42:36] You can't underwrite that.
[00:42:40] You can underwrite a deterministic system.
[00:42:44] You can put your name under a machine that can explain itself.
[00:42:48] The industry knows how to build it this way, but most of it, at least so far, it's, you know, it's a. It's a harder choice because probabilistic is faster to guess what. The bill hasn't arrived yet this week. The bill arrives and it's in three envelopes and it's in three separate capitals.
[00:43:15] But here's what. Here's what I want you to carry out of this entire conversation that we've had.
[00:43:23] Regulators did not Decide this week that AI is dangerous. Okay? They decided something quieter and much bigger.
[00:43:33] AI is permanent.
[00:43:36] You don't stress test a fad.
[00:43:40] Do not put a passing trend under the bank of England systemic.
[00:43:48] It's not an insult, it's really an admission.
[00:43:51] The machine is now part of this structure.
[00:43:55] The way the grid is and the way that all banks are.
[00:44:00] We watch that exact moment take place, you know, and it's happened over and over again in, in the history of the world. And we just oftentimes don't pay attention to it.
[00:44:14] So aviation flew for decades on really nerves and, and goggles, right before anyone demanded a licensed pilot fly it. And you inspected the airframe, electricity burned down.
[00:44:29] Oh geez. City blocks before encode ever arrived. Right. Every technology that became load bearing passed through the same narrow gate. We just weren't paying attention to it. The year the world stopped asking about whether it was impressive and started asking whether it was answerable becomes pretty clear.
[00:44:51] This week was AI's time in at the gate.
[00:44:57] Now, permanent things, of course, right? They don't get startup rules.
[00:45:02] Permanent things get inspectors, capital requirements, deadlines that fall on Halloween.
[00:45:09] Permanent things get asked forever after that.
[00:45:13] The only question that ever really mattered in the first place, and that's if this fails, who signed for it.
[00:45:26] The banks are being stress tested, the clouds are being supervised. The intelligence, at least for now, remains untested.
[00:45:36] And that's the only layer of the, the stack that doesn't have a name yet.
[00:45:43] But that's not going to last very long.
[00:45:46] It never does. Right? The signature always comes for things the system stands on.
[00:45:54] And if yours is, you know, probabilistic, I'd be a little worried about that because it's coming.
[00:46:06] The machine though, it's really become too, too big to unplug now. It just has to become worth signing for.
[00:46:18] And companies are making those adjustments and those decisions today. We see it all the time, more so now even than a few months ago.
[00:46:28] Companies are cutting back on using large language models because the costs associated with those or the hallucinations or those kind of things. And they're trying to figure out a path forward, forward.
[00:46:43] I think, you know, they're starting to see that the future of AI and using AI isn't by plugging more or having more frontier models or having them do more.
[00:46:57] It's really about using AI for what it was designed for, which means using the right AI to solve the right problem, to meet their specific requirements, because that's how you're going to be able to sign for these things.
[00:47:17] So thank you for being here. I hope you enjoyed the show. We're going to have a great show next week. Again, I'm Dr. Alan Badot. This has been AI today. And I'll see you next week.
[00:47:31] Sa.
