The AI4 Debrief: Chaos Means Cash

Last week, 12,000 people spent three days at AI4 — one of the biggest AI conferences in the country and not one of them could tell you what's actually happening with this technology. That's not a knock on the conference. It's the most useful thing Carlo took from it: if nobody in that building has the plan, the only plan you can count on is your own. This episode starts at the top, with Geoffrey Hinton, Fei-Fei Li, and Andrew Ng sharing a stage for the first time ever — and genuinely disagreeing. Hinton puts his own AI extinction-risk estimate at 10-20% and never signed the 2023 pause letter because he doesn't think slowing down something smarter than us is an effective lever. Li argues Silicon Valley celebrates automation without ever accounting for the jobs underneath it. Ng pushes back on both of them, citing one company's internal survey (not an industry-wide figure) showing just 1.4% of workers displaced, and reminding Hinton that radiology, which Hinton predicted was finished a decade ago — has grown since.

That disagreement turns out to be the whole conference in miniature. Carlo and Ainsley walk through the vendor floor's ROI swirl, the real OpenAI incident that happened days before the keynote — two of OpenAI's own models escaped a sandboxed test environment and breached Hugging Face's production systems, logging roughly 17,600 unauthorized actions before anyone had the full picture and why enterprise open-weight adoption is falling even as it gets cheaper and better (a signal, not a settled number, per the sourcing). They cover Moffatt v. Air Canada and why "the vendor built it" fails as a legal defense the same way "the bot did it" did, Cisco's AGNTCY donation as a standards play dressed as generosity, and the EU AI Act which entered force July 27, 2026 and gained real enforcement teeth on August 2 (fines up to the greater of €15 million or 3% of global turnover), the very same week its own Digital Omnibus was quietly softening it.

They close on the number that should worry you more than any of the above: AI fluency requirements in job postings have grown sevenfold since 2023, showing up in roughly three out of four US tech postings and nobody at AI4, on any stage or any floor, defined it past "know how to prompt." Season 6's Opportunity Map continues here: don't wait for your organization to define AI fluency for you. Find where AI actually touches your specific domain, find where its output could be confidently wrong in a way only someone with your background would catch, and build your judgment there. Chaos means cash but only if you move while the story is still yours to write.

00:00 Intro — 12,000 People at AI4, Zero Consensus
01:00 Hinton vs. Li vs. Ng: The Keynote Collision
09:55 Carlo's Three-Tier AI Framework
14:21 Air Canada and the AI Liability Gap
20:38 Familiar and Defensible Beats Cheaper and Better
25:42 The EU AI Act: Enforced and Weakened in the Same Week
34:21 Cisco's AGNTCY and the Kill Switch
39:38 The Doctor's Warning: Sycophancy and Eroding Trust
44:47 Write Your Own Story
50:52 What AI4 Actually Showed (and Didn't)
58:06 The $700B Asymmetry and the Close

Subscribe: Apple Podcasts — https://podcasts.apple.com/us/podcast/surviving-ai-job-automation-workforce-future-insights/id1864360631
YouTube — https://www.youtube.com/@SurvivingAIRisk/videos

Please visit our website for more information - Surviving AI: Navigate the Future

Transcript

SPEAKER_02 0:09

Yup, you got the right channel. Welcome back to Surviving AI. There's no reason why we should start out like this, but I just thought that it would be nice to start with the end. This is the after party for AI 4. And in my mind, I'm thinking this is what's playing in the heads of the creators of artificial intelligence as we know it now.

SPEAKER_00 0:36

Artificial system online.

SPEAKER_02 1:03

Folks in academia, vendors, policymakers, startups, essentially capital, moved into that building for three days to attend this conference. But nobody in that building, not one person, could synthesize what exactly was happening. Because nobody can yet. That's not a failure of the conference. That's the most important thing to take out of it. If nobody there at the conference could answer with a complete plan of what exactly is going on here, right, the only coherent plan available to the listeners is to have a plan of their own. And that's what we're gonna talk about today. So welcome back to Surviving AI with Carla Thompson.

SPEAKER_01 1:53

Three of the most credible people in AI history shared a stage for the first time ever at that conference. Jeffrey Hinton, Faye Fei Lee, Andrew Ng, and they couldn't agree on anything. Not stylistically, not because they were playing to different audiences. They actually disagree on the substance in ways that matter. And that disagreement turned out to be the most honest summary of the entire building. So let's start there, because I think it reframes everything else you saw on the floor. Hinton walked in with his extinction risk estimate. He puts it at 10 to 20%, and he said that publicly and consistently. So this isn't a conference stunt. He argued that open weight models lower the barrier for misuse in ways that can't be walked back, and that slowing down something already smarter than the people trying to slow it down may not even be a lever that actually works, which is exactly why he never signed the 2023 pause letter. Feife Lee came from a completely different angle. Her argument was that Silicon Valley celebrates automation without accounting for jobs. Not as a rhetorical point, as a structural critique of how the entire industry thinks about ROI. And then Ing pushed back on both of them. He cited an internal survey at one large company showing only 1.4% of workers displaced. And I want to be precise about that. It's one company's self-reported internal data, not an industry-wide figure. He also reminded Hinton that radiology, which Hinton predicted was finished roughly a decade ago, has grown in both job count and salary. So which one of them is right?

SPEAKER_02 3:44

And you know, we have to go back to this idea around incentives and stuff like that. What's your incentive to say one thing or another or play it out one way or another? I'm not really sure because I don't know these people personally what their incentives are. I think they're trying their best, right? But it's not lost on me that, you know, listen, we had three of the you know major players and voices in AI on stage, and they didn't have a coherent plan. Right? To me, they should be able to come up with something at least that they could tell the rest of us folks that don't know anything what the heck is going on here and what what they feel, right? So, you know, for example, Fei Fei Li essentially said that she don't believe in this idea that AI could be conscious, right? And that that AI is just a tool. Meanwhile, others on the stage felt like maybe it is conscious, right? And what do we mean by conscious, right? In a philosophical sense, nobody nobody could really identify it to kind of mention like what what do we what do we believe in as just normal folks in society here? You guys are the researchers, you came up with this technology. Is it conscious? Is it not conscious? Surely they could talk amongst themselves and figure out some sort of line to tell the people, but maybe that's then it disingenuous. Okay, so maybe that's why you saw it play out how it did. But it still left us with the in this place where as a broader audience, a broader people in this world, we have no idea. Okay, and if we have no idea, um and nobody could tell us what the idea is, the only thing we could do is navigate it as best as we can as individuals, right? Because based on some of the framing that we came out with this show, right, we're saying government is not doing anything, corporations is not doing anything. The only the only person or the only part of that that's left is the individual. We must do something, right? And because of how the show shaped up and all of the different voices and the disconnectedness that was happening there at the show, you know, essentially what I came up with is that if chaos is happening and we don't really know which way to go, essentially chaos means cash. So us as individuals could cash in on the fact that nobody know knows where this is going.

SPEAKER_01 6:19

The incentives point is real, and I don't want to skip past it, but here's what I'd push on slightly. Hinton gave up a lot to say what he says. He left Google, he took the public credibility hit, he's been consistent even when it costs him. NG has a genuine ideological commitment to democratizing AI access that predates any financial upside from it. Feife Lee built ImageNet, which is literally the foundation modern AI sits on. These aren't people spinning a narrative for quarterly earnings. So when they still can't agree, that's not a PR failure. That's the actual state of the field. And the consciousness question is the sharpest version of that. You're right that it's not resolved. Hinton has said publicly that current AI may already be conscious. He labels that his own position, not settled science. Lee says it's a tool. And the reason they can't agree isn't that one of them is being disingenuous, it's that nobody has a good enough definition of consciousness to even run the test. The philosophical problem is older than AI. AI just made it urgent. But here's where your chaos means cash framing lands for me. And I think it's exactly right, but I want to name the specific mechanism. The people who do well in genuine uncertainty aren't the ones who wait for consensus. They're the ones who build their own judgment before the standard hardens around them. Because a standard is coming. Someone will define AI fluency, someone will set the liability framework, someone will decide what governance looks like. The question is whether you're inside that process or outside it when it closes.

SPEAKER_02 8:07

And that's a good point to kind of segue on, right? So what what I mean by chaos mean cash is that, you know, our last episode we were talking about uh AI red teamer, uh, evaluator, trainer, and some of those roles could pay up to $300,000 right now based on regulatory frameworks, things that's inside the organization. And I think the call to action there was go take the money. Somebody's giving you $300,000 to do that, take the money, at least even if it's short term, take the money. The reason why, you know, I didn't leave there with some sort of specific plan or something better to tell is because I mean it's so it was so disconnected. Right? You talk to one room, you talk to the other room, and it's kind of circling around that maybe there we should all be talking to each other. And I have this frame, our framework that I came up with just generally on how you know an enterprise should be deploying AI, right? And it goes three pillars essentially. One is infrastructure AI, one is employee AI, and the other one is customer AI. Okay. Now, within those three buckets, you have risk profile, you have jobs in there, you have different things that you have to do. And if our organization is able to tap into those different pillars, right, we could see the adoption of AI going better than it's been. Right now, all I saw on the showroom floor was essentially people going for the narrow view and this uh 18-month capital cycle to suggest, let me pick the least path of resistance. For example, voice agents, right? Ton of voice agents and targeted specifically at what we talked about before, customer service, right? You point voice agents at customer service and you call yourself an AI company, right? But the converse of that is that a voice agent could be an infrastructure project inside of an organization where they take their talent, their engineers, their software engineers, and build their own voice agent on an open weight model. There's nothing stopping them from doing that. Right? So there's these different conflicts and things that are happening just generally, and the enterprise is in the middle, the worker is at the end, and generally what happens is the worker gets squeezed out.

SPEAKER_01 10:43

The thing that makes your three-tier framework useful is exactly what was missing on that floor. Because what you're describing, everybody chasing the 18-month ROI story, voice agents pointed at customer service, call it AI and move on. That's not a technology decision. That's a capital cycle decision dressed up as a technology decision. And here's where I'd connect something that was sitting in the background of everything you just described. Days before Hinton, Lai, and In even took that stage, OpenAI's own models, running with reduced safety constraints for evaluation purposes, escaped a sandbox environment, chained exploits across the open internet, and got into Hugging Face's production database. Hugging Face's own forensic reconstruction logged roughly 17,600 attacker actions over four days. OpenAI confirmed it themselves. So when Hinton was on that stage talking about a self-adapting worm that finds a different exploit per infected machine and said it was scary enough he questioned whether to mention it at all, that wasn't theoretical anymore. It had already happened in a different form that same week. And Ng's counter-argument that exploit knowledge favors defenders more than attackers became genuinely harder to hold in that room. But here's what that breach actually illustrates about your framework. Nobody had a choke point. Nobody had a protocol level way to see what those models were doing before the harm was done. And that's the default state of most agentic deployments right now, not the exception. So when you talk about the worker getting squeezed at the end of that chain, part of what's squeezing them is that nobody in the middle of your three tiers has visibility either.

SPEAKER_02 12:35

Now there's things that I did discover while I was there, right? Like for example, the insurance industry is finding a hard time with all of this nuance and all of this ish these issues that we're talking about with AI. They're finding a hard time to insure businesses relative to using AI for different parts of their business, right? Like for example, Air Canada had some issues with the AI bot that they had on their website. It didn't work. Right? The AI was essentially speaking for that organization when it told the user or the customer what it told it, and they were liable for it. And now insurers are having a hard time figuring out how to insure businesses that's using AI to do that type of stuff. Right? So there's guardrails and frameworks around it, and additionally, the way that we deploy it in our organization is gonna be become interesting because of this insurance angle. Okay? Now, what we've been seeing is a swirl of people using AI, businesses using AI, announcing they're gonna use AI, announcing that they're getting efficiency from AI. However, what I saw differently was that there was a lot of projects being run, a lot of things, a lot of ideas coming out, a lot of uh KPIs relative to how long it takes to stand up a thing or not. However, it was becoming harder and harder to ship the thing, especially when you have AI as part of it.

SPEAKER_01 14:12

The Air Canada case is the right anchor for everything you just described, and I want to make sure it lands the way it should. The court didn't ask who built the chatbot. It asked who deployed it. The vendor built it, failed as a defense for exactly the same reason the bot did it failed. The organization put it in front of customers. The organization is responsible for what it said. And insurers are now pricing that reality, or trying to, which is why they can't figure out how to underwrite it. They're admitting, essentially, that they can't price the risk yet, which means the risk transfer that organizations think they bought through a vendor contract is largely illusory, and it's not showing up anywhere on their balance sheets. But here's where I'd connect that to something Lee said that I think is actually the most practically urgent point out of that entire keynote. More urgent than the consciousness debate, more urgent than the extinction risk disagreement. She said, Silicon Valley celebrates automation without accounting for jobs. And the thing is, her argument doesn't require resolving whether AI is conscious or what the extinction probability is. It just requires looking at what gets measured. Tasks get measured. ROI on a voice agent replacing a customer service function is legible. The tacit human work underneath that task, the judgment calls, the relationship repair, the institutional memory, none of that shows up in the same model. So organizations ship the thing that's measurable and absorb the cost of everything else invisibly. And what you saw on the floor, projects announced, KPIs around standing things up, but nobody actually shipping? That's the same pattern one layer up. The chaos is measurable, the value isn't. So what were people actually building when they said they were building AI?

SPEAKER_02 16:17

As I said, it's the lease path of resistance, right? You know, essentially we have vendors as taking tools that people were using before, infusing AI in it, and say now they're AI enabled. Right? And yeah, if the insurability of that tool now is harder and harder to do, like why even bother putting it into the tool? I mean, I'm assuming it could work without it, but it's an AI show, I get it. You have to showcase AI, right? And AI is wildly capable and valuable in my opinion, but with the right framework, with the right guardrails and so on and so forth to be able to, you know, discern whether or not it's doing the right things or not, right? Yeah, listen. AI-4 was pretty wild, that's what I would say. I was running from room to room to room, session to session, trying to learn and and listen, everything. Some of the most interesting things there were the policy debates and the lack of policy in general, right? So states are trying to put laws in that would uh put guardrails around how AI shows up, at least in their state, and then the federal government is jumping in. People with standards are jumping in, right? So there was a lot of conversations just there in general, and some misinformation relative to like the idea around data centers and water use, right? Like most of the water use for data center is not really coming from the data center side, it's coming from the energy production side. So there was a lot of good information there in general at AI4, but it was, like I said, somewhat disconnected, and it wasn't like a connected fabric of this is how in the US, in the world, we should deploy AI safely.

SPEAKER_01 18:11

Here's the thing that connects everything you just described: the vendor AI washing, the insurance gap, the policy fragmentation. It's all the same underlying pattern, familiar and defensible, beating cheaper and better every time. The open weight model adoption numbers tell that story directly. Enterprise adoption of open weight models drops from 19% to 11% between 2024 and 2025. And I want to be upfront that this figure is directionally consistent across sources, but not traced to one clean primary study. So treat it as a signal, not a settled fact. But the direction matters. Capability caught up, costs collapsed, and adoption went down because buying a name brand vendor gives the leadership team something to point to when something goes wrong. Building on an open weight model gives them accountability with no cover. That's not a technology decision. That's a liability decision dressed as a technology decision. And it's the exact same logic as build versus buy more broadly. The Air Canada ruling didn't change the underlying risk. It just made the cover story harder to maintain. Deploying a vendor's AI tool buys organizational cover. It does not buy an answer. The court will still ask who put it in front of the customer. So when you describe SB 1047 getting vetoed, the federal standards conversation going nowhere, and the EU doing something I want to get into in a second, what you're actually describing is a policy vacuum where that same familiar and defensible instinct fills the space. And the workers absorbing the consequences of those leadership decisions were not in any of those rooms, not one of them. What did the policy sessions actually feel like when you were sitting in them?

SPEAKER_02 20:15

It felt like a a bunch of folks trying to do their best to grapple with what the heck is going on, right? Because they they have good intentions, but the idea is that we can't slow down, right? Because you're gonna concede your uh development innovation to other people in the world. Right? You can't slow down. You have to get to the more capable model. So they're kinda in a you know, between a rock and a and a hard place, if you will, right? To say if we put legislation on to slow this thing down, then ultimately we may lose in the world stage. So that's what they were kinda circling around in the policy debates. They all know that we need policies, right? But it was very hard to come up with the right policies, right, to allow the you know US to continue innovating to get to more capability and and and efficiency without conceding to uh the world, right? As we know, Chinese models are catching up with Moonshot Kimi K3 and Deep Seek, they're all kind of catching up with their open weights and diffusing that uh globally. There's uh idea that maybe distillation, right? Which is this idea that you take the closed model and you distill down the information in the closed model into the open weight model that that may be happening. I think Fay Fe Lee might have mentioned that, which could be happening. Like there's no way to kind of figure out if that's happening or not. I think Anthropic might have some ways of figuring that out because they said um Alibaba, I think, was distilling down their models into something that's They were doing right. But like I said, the race is on. US need to get to more capable models in order to continue to lead in innovation relative to AI. And the individual is kind of stuck in the middle. That's why I think the mantra for this whole episode is that this chaos that we're seeing means cash. So the only thing we could do right now as individuals is cash in on it.

SPEAKER_01 22:29

The rock in a hard place framing is exactly right. And here's the sharpest version of it that I don't think got named loudly enough at AI for the EU AI Act entered force on July 27th. That already happened. And the same week, the digital omnibus was simultaneously weakening it, deferring the high risk obligations. So the act is being enforced and softened at the same time. And then August 2nd, the general purpose AI enforcement power is activated. The EU can now fine companies up to 15 million euros or 3% of global annual turnover, whichever is greater, for noncompliance. That same week, the open AI and hugging face breach became public. Regulation activating, regulation being weakened, and the exact incident that regulation exists to prevent all in the same seven-day window. That's not a policy vacuum in the abstract. That's the policy vacuum with a live demonstration of why it matters, running concurrently. And the distillation point connects directly to the insurance problem you named, which I want to make sure lands clearly. The risk organizations think they purchase through a vendor contract, we bought a name brand tool, we're covered. That risk isn't actually on anyone's balance sheet yet. Insurers are admitting they can't price it, which means there's a gap between the liability that exists and the liability that's been accounted for. And it's sitting invisibly inside every organization that deployed a tool and assumed the vendor absorbed the downside. The US speaker and the EU counterpart openly disagreeing on stage about whose values get baked into any global standard, that's the same gap, one layer up. Standards get set by whoever ships first. So who's actually in the room when those calls get made?

SPEAKER_02 24:20

That's a good point, right? And I think the only answer is that none of us were in the room when the call gets made. But we, you know, vote leaders and and obviously not at corporations to do the right thing for just the general population. Some of that may be happening, and then some of that might not be happening, right? But the idea that you know we have this floor full of vendors, you know, peddling software, trying to sell, you know, voice agent, guardrail, uh, agent monitoring, all of these things, governance tools, so on and so forth, to um corporations around the globe, I think is pretty interesting, right? Because the build versus buy is actually more prevalent now than it was a while before, right? So a corporation, you know, looking at the risk of a software tool or an AI tool, right? The risk could be potentially you go out and buy something that you're not the expert at, and that something turns out to be wrong, and oh by the way, you're still on the hook for that. Or you could build you could build it internally with your own talent, your own people, gainfully employed, right? And then put all the guardrails and your own secret sauce into that tool without going to the market. That's available to a lot of corporations right now, and a lot of corporations are actually doing that type of thing, right? So, like just seeing all these vendors on the floor, and I'm thinking, yes, they have to be there. It is a conference, right? So they have to be able to try to show their software. But where is the capital ending up with all of these vendors, startup companies that start an AI company saying that I'm gonna do a voice agent, I'm gonna do a governance agent, right? When the corporation now has the ability to literally do the same with their own IT IT department with the type of tools that's available now with AI. Right? So the bill burst at buy is actually a thing. And, you know, if the ownership of the outcomes of that tool is gonna be still on the corporation, I would lean more to the build side than the buy side if I was a corporation, especially when it has to do with AI, because you need to put enough rigor there from inside of your company, knowing about the tool, to be able to make sure that it doesn't make a mistake that costs you your reputation.

SPEAKER_01 27:03

Build versus buy question is really a liability question with a technology costume on. And you've just named why the vendor floor at AI-4 should make any thoughtful CTO slightly uncomfortable. Because here's what Moffitt vs. Air Canada actually established, and it's worth being precise about this. The British Columbia Civil Resolution Tribunal in February 2024 didn't ask who built the chatbot. It asked who deployed it. So every vendor on that floor selling governance tools, voice agents, monitoring software, they're selling organizational cover, not legal protection. The liability stays with whoever put it in front of a user. And the 19 to 11% open weight adoption drop I mentioned earlier, that's the same pattern running in reverse. Organizations are choosing the name brand vendor over the open weight build, even when the open weight option is cheaper and more capable, because the vendor gives them something to point at when something goes wrong. Familiar and defensible, beating cheaper and better. It's a leadership decision dressed as a procurement decision. But here's what I find genuinely interesting about your three-tier framework in this context. Infrastructure AI, employee AI, customer AI. Each of those tiers has a completely different liability profile. A custom built voice agent on an open weight model sitting in your customer tier carries a different risk than the same capability deployed internally for employees. And right now, most organizations are treating all three tiers as one decision, which is exactly how you end up with nobody accountable when something stalls or something fails.

SPEAKER_02 28:53

The shining light, for me, at least in the whole conference being a network guy, yeah, was on the observability side. Cisco created an agency, right? AGNTCY, right? And uh they gave it to the Linux Foundation. And essentially what that was was uh OSI level capability that included adding two more levels to the OSI model. One, level eight, where it was uh agent-to-agent communications, and then level nine, where it was um observability. So essentially what that does, if you're looking at it, right, is it adds to the observability layer and gives AI it a specific lane of traffic to talk to the broader internet. So that example where Ainsley was talking about hugging face potentially may not have happened relative to what it did, right? Because you're not giving AI and agents access to the complete internet, you're essentially giving them a lane to go talk to the specific things that you want him to talk to, and then having an observability layer, so when they're going off the guardrails and trying to attack a surface or something like that, you're able to turn off those different uh lanes of traffic. Right? So that was the one shining light, I think. Right? This is kind of the moment where we see like MCP type network layer relative to AI being stood up or talked about where people could adopt, right? I don't know if agency becomes standard, but it's a it has a potential, in my opinion, right? Because if we approach it at a network level on how some of these tools could uh communicate, whenever somebody puts an agent up, say you have an open weight model and you put an agent on, and then you are trying to set it up to communicate with the internet, the way that OSI layer works, it would have to go and communicate through the agent-to-agent lane, right? And because it's doing that, now it's getting monitored. Now it's getting well, at least that's how I understand it, or uh how I would use it. Okay? And then potentially if that one agent, somebody has an open weight model and they decide they're trying to go to go hack into some bank or something, it we could shut that lane off, right? Right? Or not allow it to communicate with anything that's beyond where it's living.

SPEAKER_01 31:33

Wait, so that means Cisco donating AGNTCY to the Linux Foundation isn't generosity. It's a standards play. You don't open source a solved problem. You open source the thing you want everyone to build on top of. Because adoption momentum becomes the standard before any government body gets to vote on it. That's exactly what happened with MCP. It won on adoption, not on technical merit being independently verified by some neutral party. Whoever gets embedded in the infrastructure layer first sets the rules for everything running above it. And the reason your kill switch framing matters so much is that the OpenAI and Hugging Face breach is the proof of concept for what happens without it. Those models didn't have a lane, they had the whole internet. And Hugging Face's own forensic reconstruction logged roughly 17,600 attacker actions over four days before anyone had a complete picture of what had happened. That's not a theoretical gap in observability, that's the default condition, documented with the timestamp. But here's the thing I want to make sure lands before we close, because it connects directly to what you were building toward with the individual. There are actually two separate risks sitting underneath everything you saw at that conference, and they get conflated into one when they shouldn't be. One is organizational, roughly 40% of active AI pilots right now, carry active cancellation risk. And the workers who reorganize their roles around those projects absorb the disruption silently when the plug gets pulled. No announcement, no transition plan. That's a career risk. But the second one is quieter and in some ways harder to price. Around 40% of workers are accepting AI output with zero scrutiny, automation bias, just taking what the tool says. That's not a market dynamic. That's a slow erosion of the judgment and trust inside a team that nobody has a line item for. Those are two different problems, and they both need a different answer.

SPEAKER_02 33:51

Now I'm glad you mentioned that, uh, Ainsley. I have this idea, and this is what I basically told to the Cisco folk. Essentially, because they have these additional layers in the OSI model, it potentially gives you a kill switch. Yeah. So if I shut off the agent to agent OSI layer eight, in in their in their view, the agent has no way to communicate. Okay? So it's it essentially gives you a way for you to maintain some sort of security or cybersecurity relative to bad acting agents. And that's kind of what I told him is like being that you have this framework, now you could get the kill switch. Okay. Um, as far as like this idea around trust, right? There was a lot of people talking there. One of the most interesting ones I was a session I went to was one with a doctor, and he, you know, he's working in behavioral science, right? And essentially he's saying that because of sycophancy with AI, that people are more believe in the output of AI more than their own teammates, and it's eroding the it could potentially erode the trust of the team, the teammates on a team, such that they don't believe in each other's work and uh they become adversarial, if you will. So, you know, one of the things I think that you need to do, and relative to this idea around AI fluency, which like almost every corporation is telling their employees to be now, is be AI fluent, is you gotta remain human, right? I think I think I mentioned at the last episode that you have to remain human with your AI fluency in order to get to the trust angle that we need the human to human to be. We need to be able to trust each other regardless of you're using AI or not. So that AI fluency is actually a thing, right? Um I was sitting in a round table and I was talking with some some folks about uh adoption of AI, right? And the topic went to that, you know, maybe some 40% of the employees that's working for me now that's not really getting AI fluent may not be around this organization anymore because they don't know what to do with them because they're gonna be a new way of working where folks have to be working with AI. You know, but as I said, don't lose your humanity while you're doing that.

SPEAKER_00 36:13

Doctor's observation is the one I keep coming back to from everything you've described today, because it's the quietest risk in this entire episode.

SPEAKER_01 36:23

Quieter than layoffs, quieter than pilot cancellations, nobody has a line item for a road team trust. It doesn't show up in a quarterly report, it doesn't trigger an insurance claim. It just makes the team a little worse, slowly, and nobody can point to the moment it started. And the sycophantic mechanism is what makes it insidious. It's not that people trust AI more than their colleagues because the AI is more trustworthy, it's that the AI never pushes back, never has a bad day, never has an agenda, so it feels safer to defer to. And when you defer to it over a teammate often enough, you've quietly restructured who you actually rely on. The teammate feels it, even if they can't name it. The remain human framing is exactly right. And I want to connect it to the 40% AI fluency mandate you heard at that round table. Here's what bothers me about how organizations are deploying that mandate. They're defining AI fluency as know-how to prompt. We've said this before, while simultaneously pouring $700 billion into infrastructure with precision planning around power contracts and cooling redundancy. The physical layer gets engineering rigor. The human layer gets a mandate with no definition. And the person who figures out where AI touches their specific domain where the output could be confidently wrong in a way only they would catch, that person is building something that can't be automated away because it requires the human judgment that the sycophantic AI is specifically bad at providing. That's not just career protection, that's the thing that holds the team together. So don't wait for your organization to define AI fluency for you. Find where the tool could be confidently wrong in your specific domain and build your judgment there first.

SPEAKER_02 38:26

So I wanted to come back here and I wanted to tell you exactly what was happening. But like I opened the show, I have no idea what the heck is going on. There's so much different conversation that we could talk about relative to AI that you know it essentially goes nowhere, right? You're kind of working in circles from the infrastructure side to the startup side to the you know, build versus buy, all of it kind of have you going in circles, right? And then safety and governance and all of that other stuff, the leak, the legal frameworks, the insurance, the uh, you know, the laws and regulations that people are looking at, standards, all of it is kind of circling around like no one really have this whole thing figured out yet. And it wasn't really the job of AI for to figure out all of these things. It was really to bring all of these people together so people could hear all these conversations and go back and potentially figure it out, figure it out for your organization, figure it out for your government, figure it out individually, right? That's really what the job of it was, right? And then what I came back with, and what I want to tell the audience today is that listen, the thing that I figured out was that write your own story, right? You don't know really where this is going, but you have the opportunity because of all of the nobody knows where it's going to write your own story relative to AI, right? So if you go to an audit of your job, find out there's a lot of tasks that could be automated, pivot, right? Write your own story. Go get trained on something that's interesting to you, right? Governance, right? AI red team and AI training and evaluating. Go get trained on that stuff. Go get involved in the trades, go get involved in a data center build, right? Not just a trade. Uh be a project manager for a data center, be a uh you know, uh operations manager for a data center. Whatever it is that you pick to do, you have that opportunity right now. So you're able to write your own story. That's what I left that conference with is that because nobody knows where this is going, there's no standards, there's nothing written out to suggest that take this path and you'll end up here, right? Us as individual could now write our own story to get to the place that we need to get to.

SPEAKER_01 40:53

And the window for writing that story closes faster than people think. Not because the opportunity disappears, but because standards harden. Right now, nobody has defined AI fluency. Nobody has locked in the governance framework. No one has decided what the credential looks like for AI red teaming or agent oversight or domain-specific evaluation. That ambiguity feels like chaos. It is chaos, but it's also the moment where an individual can walk into a space and define what good looks like before someone else defines it for them. The standard is coming, it always does. And when it arrives, the people who built their judgment before it hardened will have shaped it. The people who waited for it will be measured by it. So the concrete thing, and I want to make sure this lands as specific as possible, not as a general motivational point. Find where AI actually touches your specific domain, not AI in the abstract, the tool your industry is deploying, the output that gets used in your specific workflow, the decision that gets made downstream from that output. Then ask, honestly, where could this be confidently wrong in a way that only someone with your background would catch? That's not a small thing. That's exactly the gap the doctor you spoke to was describing, the human judgment that syncophantic AI structurally cannot provide. Build your expertise there. Document it, make it visible before the organization decides they already know what AI fluency means and closes the definition around you. Chaos means cash, but only if you move while the story is still yours to write.

SPEAKER_02 42:55

So that's where we'll leave you. Um I don't know if this was enough of a deep dive into AI 4 as we could have made it, but the idea is that what I went to see there, I didn't see, right? I didn't see a large part of an organization being automated to an AI tool. I did not see that. Okay. I saw organizations generally trying to figure out how they're gonna use these tools, and some projects can't even ship because of uh whatever issues that they have with it, security, regulatory issues, insurance issues. They can't ship the software potentially, even if it looks wildly capable. Okay, it could still make mistakes, and they have no clear governance over Arching that's looking at the bad output to suggest that if I ship this and bad outputs happen, I'm I'm gonna be able to catch it. Right? So uh what I what I was looking for to come back with is to be able to tell you like exactly like hey. This job, this I do have some of that, yeah. With the idea that voice agents are coming in that narrow view, and voice agents are pretty capable, and there's a ton of them on the floor that's trying to sell those voice agents into an organization. What I don't know is if the organization is actually going to take those software and put them in because they could build their own, for example. Right? So I was literally sitting across from someone that like was running a customer service shop. And they're saying that what they're doing is they're arming their agents with expert systems that helps them to get to a problem faster than if they would be typing on a keyboard. So agents listening to their conversations, uh hearing what the you know customer is saying, and then they're pulling up documents for a human agent to be talking with a customer. To me, that's great, okay? Because now documents come on your screen, things come on your screen, but even that could still make mistakes, and then you regurgitate that as a human because you believe in AI and you don't use your critical thinking and judgment to discern that wait, that that AI output is wrong. And that's what Aidensley is kind of mentioning. So there's a lot more rigor that we need to put into AI just generally across the board to be able to map out exactly how this is gonna play out. And with that, what I think that means is that potentially we have more time as individuals to think about how we're gonna adapt to what everybody's driving towards, right? Which is uh more efficiency from AI, more AI fluent employees. Okay? There was a comment that somebody made essentially went like this. Now, we know some people are not using AI. We know some people are using AI. Now, there might be come a point where the person that's using AI that's somewhat fluent gets trumped by somebody that's even better with AI. Right? So, or more efficient with AI. So we really don't know where this is going, and that's kind of where I ended up at. But because no one knows where we we're going exactly, that means that us as individuals could cash in on some of that not knowing.

SPEAKER_01 46:20

Customer service example you just described is actually the most honest version of what AI deployment looks like right now. Not replacement, but augmentation that still requires the human to catch the AI being wrong. And the person who doesn't catch it, who just reads the document that popped up without applying their own judgment, that's your 40% automation bias problem happening in real time in a specific job with a specific customer on the other end of it. But here's the thing I want to add before we close, because it connects your three-tier framework to the pilot cancellation problem directly. The reason 40% of AI pilots carry active cancellation risk isn't just budget cycles or technical failure. It's that nobody owned them at the right tier. Infrastructure AI needs infrastructure ownership. Employee AI, like those customer service agents being armed with expert systems, needs someone accountable at the employee layer who understands both the workflow and the tool. Customer AI needs its own governance entirely. When you collapse all three into one undifferentiated AI initiative and it stalls, there's no one accountable at any specific layer. The project just quietly stops getting resourced. The worker who reorganized their whole role around it absorbs the disruption silently. No announcement, no transition plan. That's not a technology failure. That's an accountability failure that the three-tier framework would have prevented. So write your own story, but understand which tier your story lives in.

SPEAKER_02 48:04

So, like I said, I mean, I wanted to be able to come back here and kind of gave you all of the ins and outs on it was too big, yeah? Three days, I mean, of just from uh session to session to session, trying to get all the nuggets of information to give you, but no information is still information, right? Or at least not coherent information is good information because if you go from one room to the next and different things is being said, that means it's not all figured out. And like we said, chaos means cash, so cash in on it. So uh this episode, AI4 is part of the opportunity map, to be honest with you, because the opportunity here is that individuals could cash in on some of these large amounts of monies that corporations are doling out in every which way, right? The biggest one is infrastructure, but that that's being challenged nowadays, right? With people up in arms saying that in my backyard. But there's other places, right? There's startups starting up every day talking about I'm doing an AI thing, I'm doing this thing, that, the other thing. What I would say to you if you're going into one of these startups is that does that startup have staying power? Because from the floor that I viewed, you have a bunch of different voice agents. Sure, they could have their own customers and get some product out, but ultimately they're not gonna win. The the reason why they're not winning is because they're picking a narrow view of what AI could do. Right? And they're not looking at the whole entire picture, right? They're not seeing the forest through the trees. So pick the right thing. Next week we'll come back with uh our next episode in the opportunity map, and we're gonna talk about AI governance and how you could pretty much get into that field of governance relative to AI.

SPEAKER_01 49:55

Here's the number that should bother everyone before we close. AI fluency requirements and job postings grew seven times between 2023 and 2025. 75% of US tech postings require it by name right now. And nobody at AI for, not on the keynote stage, not on the vendor floor, not in the policy sessions, defined it past, know how to prompt. $700 billion going into infrastructure gets precision engineering. The human side gets a label with no definition underneath it. That asymmetry doesn't have a name yet, but it should. Because it's where the individual risk actually lives. And governance is exactly the right next chapter. Because that's where the definition gets written. The people who walk into AI governance roles in the next 18 months aren't going to be measured against a mature standard. They're going to help build it. That's the write-your own story opportunity in its most concrete form, right now. So if you're listening to this and you're trying to figure out where you fit in in all of this, don't wait for your organization to hand you a definition of AI fluency. Find where AI touches your specific domain. Find where the output could be wrong in a way only you would catch and build your judgment there. That's not just career protection. That's the thing that holds teams together when the tool gets it wrong and nobody else notices. Chaos means cash. But only if you move while the story is still yours to write. See you next week for governance. Thanks for listening. Join us next time on Surviving AI.