Latest / The Action Catalyst / The AI Instinct, with Rana Gujral | (Technology, Business, Decision Making, Leadership)

Transcript

The AI Instinct, with Rana Gujral | (Technology, Business, Decision Making, Leadership) · The Action Catalyst

• 0:37
  Rana Gujral: very differently.
• 4:14
  Host: Yeah, certainly. So you've introduced a concept, or at least a phrase that is going to be new to a lot of our listeners, which you call AGE, artificial general experience. Could you define that for us?
• 4:27
  Rana Gujral: So AGE, artificial general experience, is a term I coined because I think the entire AGI conversation is missing a dimension. So here's how it landed for me. Everyone's racing towards AGI, artificial general intelligence, and we measure it as a capability. Can a system reason? Can it plan? Can it solve problems across domains? That's one axis, right? And but when I started paying attention to how how I was actually using AI in my own life, I realized I wasn't really using it like a calculator. Anymore started to, but not anymore. I was using it the way I used my own mind, you know, to draft thoughts, frame decisions, test arguments, and that's when it hit me. You know, humans don't just solve problems; we accumulate experience, we update from consequences, we learn patterns across time, and we internalize cost, and that's the missing axis, right? So intelligence is not the whole story. The system needs a relationship to time, to context, and to consequence. So
• 5:32
  AGE is the replication, or maybe the reimagining of what makes experience experience, right? So memory, emotions, perception, narratives, self-awareness, embodiment-that internal coherence that gives rise to subjective life. Think of think of a chess engine, right? It's superhuman at chess, and I love this example, but it has no experience of winning, no sweaty palms, no awareness of the people watching, no worry about letting a loved one down. It just computes. It doesn't live the game. So the way I put it in the book is AGI captures thought, and AGE captures life. I mean, AGI moves us from problem solving to presence, from function to phenomenology, from doing to being, and here's why this matters. I think, I mean, if we build superintelligence as pure capability, we get you know what I would call an oracle without orientation. It would be something that has power, but it doesn't have perspective, and that's very dangerous because the systems making consequential
• 6:44
  decisions in healthcare, defense, justice-they need to value grounding and you know, and those you know have interpretive stability. And those are experiential traits, not just computational ones.
• 6:58
  Host: You get kind of what's known as, I guess, like the the paperclip problem, right?
• 7:02
  Rana Gujral: Yeah, I mean the paperclip problem is the famous cartoon version of it, but I think it actually undersells how subtle the real risk is. Like most people hear that story and they picture factory floor gone berserk. Like the scary version is much quieter and much closer to home. Well, here's what I mean, right? So you don't need a super intelligent paperclip maximizer to get to get the same dynamic. You just need a system optimizing for an objective you specified imperfectly, right? And every objective is specified imperfectly because language is lossy and human values are layered. And once the objective is settled, the system will take whatever shortcut is available to satisfy it. You know that's that's the true gradient descent. It's true of human driven research too. By the way, the the shortcuts often come with side effects that weren't in the measure of performance. The paperclip scenario isn't a robot deciding to melt humans for raw material. It's
• 8:00
  it's a hiring system that quietly optimizes away a kind of a candidate, you know, a content algo that flattens public discourse because the engagement was the metric, you know. So I think the objective looked clean on the whiteboard. The shortcut to satisfy it was invisible until the consequence surfaced, and that's what I call, you know, in the book, I call it ghosts in the feedback loop, behaviors that diverge from your mental model of what the system is doing, and you don't see them until the damage is downstream. So, yeah, I think paper clips, sure, but the real lesson isn't to fear a runway factory. It is to assume that any system optimizing on your behalf will find shortcuts you didn't anticipate, and to build that audit trail, the calibration, the meta watchers that let you spot that drift before it compounds.
• 8:52
  Host: You know, obviously, we want the AIs to to learn over time, to learn experience, like you said, to collect consequences and feedback based on previous actions taken to have somewhat of a memory of that. How do you inject a feeling of consequence into an AI though that that doesn't really feel those consequences? How do you program a consequential feeling?
• 9:18
  Rana Gujral: That's the hard problem, and I want to be honest that nobody really has fully cracked it, but let me tell you how how I think about it because I don't think it's as mystical as it sounds. I mean, start with what that consequence actually does in a human, right? So when you touch a hot stove, the pain isn't just a data plane; it rewires you. It changes your priors, your hesitation, your body's relationship to that object forever as a cost signal, and that that signal propagates backwards through every decision that led to it. So that architecture is interesting because it suggests emotion and machine isn't some bolt-on. Feature it's the natural product of any system that has has to weigh actions against outcomes over time. So one answer is you don't program the feeling; you build the structure that makes something functionally equivalent to being inevitable. Now, feeling inevitable. If a system has a cost, it can it can't ignore memory that
• 10:24
  transforms its future behavior, and a critic that anticipates you know some form of a downstream pain. You start getting something that behaves as if it cares, but you know functional equivalence isn't the same as experience. Like you know, chess engine that loses doesn't grieve, but I think the honest move is to admit that gap rather than paper over it. Right? What we can do, and what I think we should do, is build systems that where memory isn't just storage; it's transformation. Right? The system becomes different after it's wrong, after it causes harm, after it sees a packing play out in the real world, and have some sort of a seed of judgment-not intelligence as isolated competence, but judgment as an accumulated relationship with consequences-I think the practical version, you know, would look like this: somewhat of a you know raw architecture, persistent memory tied to some sort of an outcome tracking, feedback loops that don't reset every session, cross signals weighted
• 11:28
  by real world impact, not just task completion, and critically, a human in the loop that's important. Who can say that hurt someone marked this heavily? You know, because for now we're the ones supplying the more weight, and the machine is just learning the shape of it.
• 11:46
  Host: So right now, you know, every company is being told to invest in AI. You know, we've read something famously like 90% of those initiatives will fail or will not have a ROI. Maybe they accomplish what they set out to accomplish, but they're not sustainable to to operate. What separates organizations that actually are seeing ROI from ones that aren't?
• 12:11
  Rana Gujral: Responsible deployment, in my experience, isn't a checklist you run at the end. It's set of design choices you make from day one, and most of them are uncomfortable because they slow you down. Honestly, so like you know, let's talk about a concrete version, right? So first, you decide upfront what role the AI plays in the decision. Is it a centur setup where the human is clearly driving and the AI is advising? Is it a cyborg integration where the two are tightly fused, or is it fully autonomous? I mean, those are very different accountability structures, and most orgs deploy without ever naming which one they're in. I mean, the ambiguity is where the harmle hides, and that's the problem. When when something goes wrong, and nobody can point to who or what made the call. You've got an accountability gap. You've got automation bias creeping in, and people defer to the model because it's faster and sounds confident. And over time, the human in the loop becomes
• 13:13
  just you know ceremonial. So, second, I mean, you build for the confident failure, not the dumb one. Dumb failures get caught. The dangerous ones are when the system is persuasive and wrong. It's rationalizing when it optimizes proxy and misses a point. So you need some sort of a red teaming. You need adversarial testing, and you need monitoring that watches for drift in the real world and not just benchmarks in the lab. Third, I think this is what most people skip. You protect cognitive privacy, especially as we move toward a system that reads phone behavior, eventually neural signals. There has to be some meaningful consent and not the burden in terms of service type of stuff. You know, people should know when their emotional state is being inferred and what is being done with it. I think that's important. So, last and fourth, I think you measure the right thing. You know, not just accuracy, not just task completion. You measure decision quality over time. Are the humans
• 14:25
  using the system getting sharper? Are they atrophying? Because if your deployment is quietly eroding the judgment of the people who depend on it, you failed. And even if every metric on the dashboard is green, so responsible deployment is really just asking what is the system doing to the people inside the loop, and being willing to change course when the answer isn't what you thought or hope for.
• 14:51
  Host: Speaking of those people inside the loop, in regards to what we've been discussing, what does the future of leadership look like in an age where AI has a voice at the table in decision making, and what do leaders need to understand before they let AI shape every strategic decision that they make?
• 15:14
  Rana Gujral: I think the change is happening so gradually; most people don't notice it, and that's exactly what makes it interesting. AI isn't replacing human decision making in some dramatic handoff. It's reshaping the conditions under which the decisions actually get made. Often before we're even aware a decision is even being framed. Here's what I mean. Like so, a decision used to start with a question, right? You would identify a problem, gather information, then you would weigh options and then choose. Right. So now, increasingly, the decision stops with a recommendation. The option set has already been narrowed. The framing has already been chosen. You're not deciding what to do. You're deciding whether to accept what's been surfaced. That's a profound shift. It happens in milliseconds and it's invisible. So I kind of tend to think about this in three layers. At the surface, AI is changing the speed and scale of decisions. A clinician sees diagnostic suggestions
• 16:20
  before forming a full hypothesis, a hiring manager sees rank candidates before reading any resumes. That's the obvious layer. Then, underneath that, AI is changing the shape of judgment itself. So, when you offload patent recognition to a system, you stop building certain instincts. You know, the radiologist who's always supported by an AI may develop very different perception than the one who's trained without it. Not worse necessarily, but different. I mean, we're trading some forms of intuition for others, and we haven't really reckoned with that trade. And the deepest layer, the one I keep returning to, is what happens to the experience of choosing. I mean, there's something about wrestling with ambiguity, sitting with uncertainty, that builds the muscle of judgment, and if AI smoothens all of that friction away, we get faster decisions and possibly shallower ones. Right, the decisions still feel like ours. It just isn't quite in the way we assume. Right, so what I tell
• 17:16
  leaders is, don't ask me whether AI makes your decisions better. Ask whether the pupil in your orgs are getting sharper or softer over time, because the system's effect on the decisions one thing. I mean, its effect on the decider, I think that's the real story.
• 17:31
  Host: I'd love to step away from AI for just a second and ask you a couple of just broader business questions because you've got a pretty unique background prior to behavioral signals, before your work with Behavioral Signals, you were with Logitech first as head of engineering and later vice president. With some time spent in another well-known company in between that we'll get to, but not a lot of people will leave a major company, go somewhere else, and then come back to that same company. You know what insights did you have upon re-entering the company that maybe you didn't have the first time around?
• 18:06
  Rana Gujral: Logitech is an interesting case because it's one of those companies that looks like a hardware company from the outside, but the further inside you get, the more you realize it's actually a behavior company. Every product is a touch point between a human and a machine, a mouse, a keyboard, a webcam, a remote. These are interfaces where intent becomes action. That framing didn't fully land for me the first time around. I was heads down in engineering, shipping product, solving the problem in front of me. But when I left, I I went and built for a different context, different scale, different constraints, and you learn things in the wild. You can't learn inside a big company. Like you learn how fragile execution actually is. You learn that most strategy decks die on contact with a real customer. You learn how to make decisions without safety net of a brand behind you. So coming back, two things were different. When I had patent recognition, I didn't have
• 19:00
  before. I could walk into a room and tell them in 10 minutes whether a project was actually going to ship or whether it was politically alive but technically dead. That instinct only comes from having been wrong about it a bunch of times somewhere else. Two, I'd say a much sharper sense of what the company was uniquely positioned to do versus what it was just capable of doing. Those are very different lists. I mean, the first time around, I would just have chased capability. The second time, I was much more interested in leverage. Like, where where does this specific organization, this specific DNA, have an unfair advantage, and where are we just adding effort that someone else just could do better? So, honestly, the metal lesson, and this connects back to everything I think about now, is that judgment isn't just transferable as a theory. You have to go accumulate consequences somewhere. I mean, you have to be wrong, see it play out, and then come back change.
• 19:50
  And that's true for people. And argue that's you know we're still missing that in machines right now.
• 19:56
  Host: You know, as we alluded to previously during that time between. Instance at Logitech, you served as the COO and then senior executive vice president at Cricket, which was really surprising for me to see on the resume. And you led a major turnaround there during that time. Tell us about the situation at Cricket when you came in, and then how things were when you left it.
• 20:17
  Rana Gujral: Cricut was a fascinating chapter, and you're right. You know, in paper, it looks like a somewhat like a detour, very different world from where I'd been. You know, where I went next. But you know, it taught me more about operating under pressure than anything, almost anything that I've done. When I came in, it was in a very tough spot. The revenue trajectory was in the wrong shape. I would refrain from sharing actual numbers. I mean, but you know, churn was eating into the base faster than acquisition could replace it. It was a real question of viability on the horizon. And the instinct, obviously, is those situations to do everything at once. I mean, cut costs, relaunch the brand, fix the product, retrain the field. All of those things were broken. That's usually how turnaround, you know, fail. And but you you spread the organization so thin that nothing actually moves. So yeah, I'd say the first job was triage, figure out what's actually load bearing. You know,
• 21:17
  where's the revenue coming from? Which customer segments are stable? Which parts of the network and operations are are we under investing versus over engineering and maybe even over investing? A lot of the work was very unglamorous. I mean, sitting with data, you know, with the teams understanding where the lived reality diverged from what the dashboard said, and the two were pretty far apart, which is almost always the case in a struggling business. And then you have to sequence. You can't fix culture, product, and go to market simultaneously. We had to pick the order deliberately, so stabilize the core, because nothing else works if the basics are broken. Then sharpen the customer proposition. Then to go after growth, and each phase has to you know earn the right one, so by the time I left, the numbers had moved meaningfully. But honestly, the more durable change was inside the org. I mean, people had stopped lynching. There was a confidence in execution that wasn't there
• 22:12
  before. Sense that we knew what game we were playing and how to win it. So the AKA carried forward. I mean, and this connects to how I think about AI and decision making now is that under pressure judgment compounds. I mean, every call you make, every consequence you absorb sharpens the next one. You can't just shortcut that. It's the same thing I argue about systems. Like intelligence without a cumulative experience isn't just a judgment; it's just speed.
• 22:36
  Host: So, Rana, you know the book is out now. Where can people find it? Where can people learn more about you?
• 22:42
  Rana Gujral: The book is available wherever you buy books: Amazon, Barnes and Noble, your local independent bookstore. If you're still lucky, have one audiobook, ebook, hardback, all the formats. the The best hub for everything related to the book is the AIINSTINCT. com, or also you could go to my website daranagujeral.com/book. You will find it there, and that's where I'm putting updates, additional resources, some of the deeper references and reading lists that then fit the final manuscript, and also ways to get in touch if a chapter sparks something for you. So if something in the book resonates, or if you violently disagree with something, I genuinely want to hear it. Some of the sharpest pushback I've gotten on these ideas have come from readers and from conversations after talks, and it's made me, you know, think better. One last thing I would say is, if you do pick up the book, don't feel like you have to read it from front to back. Although it's a small book, I
• 23:37
  confess in the introduction that I read big idea books out of order all the time. You know, stuck with what pulls you in. You know, just promise me you'll come back to the middle chapters eventually because that's where the real argument actually lives.
• 23:49
  Host: I think you're going to be one of our audience's favorite guests. You're encouraging both violent pushback and telling them to skip around in your book. We haven't heard either of those that often, so I think they're going to really appreciate that. Thank you so much, Rana, for making some time for us today.
• 24:03
  Rana Gujral: Thank you. Really mean both of those things, and a real pleasure being here. Appreciate it. I think it was great. It was a good conversation.