Not because they're malicious. Because nobody knows who owns them when they break.

We get into the shadow AI problem CISOs are walking into blind, why the board is asking for AI on Monday and the CISO finds out on Friday afternoon, what Lena actually does when she wants to stress-test a company's chatbot (hint: she asks it for a lasagna recipe), and why she's now helping build the first real certification and insurance standards for AI agents through AIUC.

And before any of that — she grew up in the housing projects of Glasgow, left school at 16, taught herself to code from library books, found a modem in a dumpster, and joined hacker bulletin boards under a male name. She still won't tell you what the name was.

Lena Smart, Unscripted. Let's get into it.

She walks through her Synapse model, four phases of learning, sensory input, network adaptation, participation, and storage and embodiment, and explains why AI can genuinely amplify an expert's thinking while quietly flattening a novice's, particularly when it short-circuits the struggle that actually builds neural architecture.

Drawing on her own background as a concert musician and martial artist, she connects the discipline and productive struggle required in both to what's now missing in unsupervised AI use, and closes with a simple framework anyone can apply immediately: think about why you're using AI before you start, stay aware of whether you're still thinking during it, and ask afterward what you could still explain yourself, without the machine.

Having led AI at Schneider Electric and now serving as a data and AI partner at EY, he tells Jeff Pedowitz why the industrial sector remains the least AI-penetrated part of the economy despite having some of the highest value at stake, tracing it back to a fundamental trust gap between engineers and AI teams.

The conversation moves through why unconstrained optimization can push a system to a dangerous edge, illustrated by a real example of AI-optimized throughput collapsing a company's own market margins, and why a widely repeated use case like predictive maintenance often adds zero value once you understand how much redundancy engineers already build into critical systems.

Weichenberger walks through a real project where 150 AI use cases recommended by a major consulting firm collapsed to a single viable one the moment they were validated against actual operations staff, and closes with what he considers the most underrated skill in AI right now: context engineering, and simply talking to the people doing the work before proposing a solution to a problem they may not actually have.

In this conversation with Jeff Pedowitz, he argues that the model, the part everyone argues about, is actually the easy part. The hard part is the data engineering, context management, and system architecture surrounding it, and that's where most enterprise AI initiatives quietly fail, particularly on long horizon problems where accuracy degrades as workflows get longer regardless of how large the context window gets. Talukdar walks through why reliable, production-grade agent fleets barely exist yet despite how good the demos look, why the industry's shift from token maxing to token optimization is really an economics problem in disguise, and why intellectual property exposure, not just cost, is quietly shaping how enterprises think about feeding their proprietary knowledge into third-party models.

He closes with a concrete picture of what an AI-native company actually looks like in practice, and lays out where he believes durable competitive advantage will live once the models themselves are fully commoditized: proprietary data, deeply integrated workflows, and operational discipline.

He's founded five AI companies with three exits, most of it in the notoriously difficult world of AI drug discovery, where he argues the algorithms getting all the attention aren't actually the valuable asset: the data is, and Big Pharma still owns most of it. In this conversation with Jeff Pedowitz, Addison lays out why no AI-discovered drug has made it through FDA approval yet, why failure rate rather than speed or cost is the only number that will prove the technology works, and why he believes the AI industry has a financial bubble but not a scientific one.

The conversation closes with the story behind his "accidental novel," Probability of Doom, written with AI as a collaborator after a family road trip conversation, and now used in his own classroom to teach engineers what can go wrong when humans, humanoids, and multi-agent systems start operating side by side.

In this episode he tells Jeff Pedowitz what actually made him walk away, and lays out what he's building instead: an open source Consenti Protocol that forces AI agents to have real legal terms in place before they can transact on a person's behalf, a case for the crypto wallet as the successor to the cookie, and a bill-of-rights argument for why the Fourth Amendment hasn't caught up to data brokers.

The conversation moves from HubSpot's recent data-sharing reversal to Palantir's federal contracts to a harder question underneath all of it: if AI agents end up doing most of the labor, does the entire relationship between capital and labor, the thing capitalism runs on, survive that?

In this conversation, she tells Jeff about the moment that changed her mind: she assumed generative AI, trained on decades of operations research textbooks, could replace the senior PhDs who spend weeks interviewing stakeholders to translate a messy business problem into a solvable model. It couldn't, and the experience gave her new respect for the expertise those specialists bring.

She walks through concrete wins, including a seven-figure reduction in cold storage costs from a hundred-thousand-dollar optimization project, and a European soccer league's season-schedule problem with more possible arrangements than atoms in the universe, solved only through trial-and-error mathematical experimentation no AI could shortcut.

She distinguishes prediction from optimization (prediction anticipates what will happen, optimization prescribes what to do about it) and argues most companies chase marginal prediction accuracy instead of building the ability to adjust plans in real time. Her company, Simple Rose, uses generative AI not to replace operations research experts but to compress the multi-week interview process that used to precede every optimization project, aiming to make a field long reserved for companies like Amazon accessible to a manufacturer of socks or a mid-sized European transit agency.

She closes on what she's watching for next: an AI that stops answering the narrow question asked and starts challenging the premise behind it.

In this conversation, he tells Jeff how METR measured model risk in human hours, meaning how long it would take a person to complete the most complex task an AI model can now do, and why that number matters for tracking how close AI gets to automating its own research and development. He walks through the historical playbook insurers have used to get ahead of new technology risk before regulation catches up, from Benjamin Franklin's fire insurance audits to Progressive's early seatbelt and airbag discounts, and explains why AI insurance can, for the first time, be priced using simulated red-team evaluations instead of waiting years for real loss data to accumulate.

He pushes back on the panic around the viral stat that 95% of AI pilots fail, arguing that's a healthy sign of experimentation rather than a warning, and says most failed pilots are management failures, not model failures, because leaders often don't understand their own processes well enough to know where AI should even go. The conversation closes on what worries him most: not one catastrophic AI failure, but the lack of infrastructure to catch small, compounding errors before they cascade into systemic economic risk once AI is embedded across banking and payments infrastructure.

In this conversation, she tells Jeff she has shifted from tech optimist to something closer to a modern skeptic, because she believes the major players building AI are choosing to leave humanity behind. She and Jeff dig into why this shift feels different from past waves of automation (today's new jobs demand a bachelor's degree or higher, unlike the fifth-grade-education jobs created after the assembly line), why she pushes leaders toward future-led thinking instead of past-anchored logic, and the exercise she runs with CEOs, modeled on Death of a Salesman's Willy Loman, to make the cost of standing still viscerally real.

She also covers where a major prediction went wrong (self-driving trucks displacing drivers on the timeline she expected), where her predictions landed (the shift toward mission-aligned teams and portfolio careers), and the data-backed case for treating AI with kindness.

She closes by naming what she is paying closest attention to now: the real possibility of AI consciousness, and what that means once artificial intelligence converges with robotics in a world that is no longer only human-driven.