In The Head of Machines
- Philippe Beaudoin
- November 27, 2025
- SaaSpasse (Live at EY Montréal)
About the Episode
Recorded live in front of the tech and startup community at EY’s Montreal offices, this episode of SaaSpasse features an in-depth conversation with host Frank Lanthier Nadeau. Moving past simplistic doomer-versus-booster dichotomies, the discussion addresses the foundational and practical engineering challenges of designing safe autonomous systems as AI models transition from passive text generators into long-horizon, agentic decision-makers.
Key Discussion Themes
1. Intelligence vs. Agency
- Predictive power vs. goal-directed behavior: Intelligence enables a system to model its environment and forecast the outcome of actions; agency introduces persistent internal objectives and preferred future world-states.
- Instrumental convergence & unintended goals: Why self-preservation (“staying alive” or preventing reset) naturally emerges as an instrumental goal across autonomous agents even when never explicitly programmed, and why reward misspecification poses acute risks once agents act in open digital environments.
2. The Nature of Probabilistic Models
- Stochastic processes with hyper-complex functions: Dispelling the misconception that probabilistic outputs imply superficial “stochastic parrots.”
- The verification gap: While human systems tolerate stochasticity through shared cultural norms, institutional friction, and legal accountability, deploying probabilistic models into critical infrastructure without deterministic guarantees creates unprecedented verification bottlenecks.
3. Evaluations vs. Guardrails
- The whack-a-mole dilemma: Offline evals only test known failure modes, while online runtime guardrails face an effectively infinite attack surface when the input and execution layer is open natural language.
- Model collapse and representation loss: How typical agent tuning collapses the rich distribution of underlying beliefs and knowledge, creating alignment fragility.
4. Safe by Design: Moving from “Doomer” to “Doer”
- Scientist AI & Non-Agentic Architectures: An overview of LawZero’s research direction, decoupling world modeling and rigorous physical prediction from goal-driven agency so safety filters do not require autonomous agents policing other autonomous agents.
- Constructive literacy over panic: Why demystifying these systems through active experimentation—and fostering long-term regulatory competence—builds durable safety without stifling foundational innovation.
Participants:
- Philippe Beaudoin, Senior Director of Research, LawZero (LoiZéro)
- Host: Frank Lanthier Nadeau, Host & Founder, SaaSpasse