#58 - Juan Pablo Carbajal: Tadpoles, tolerance and why AI alignment is really a human problem

2026-03-02

Download

We train AI systems to optimize for metrics — but what if the real alignment problem isn't in the machine? What if it starts with us?Juan Pablo Carbajal is a physicist, interdisciplinary researcher, and educator at the Eastern Switzerland University of Applied Sciences (OST). Originally from Argentina, his career spans physics, robotics, biomechanics, agronomy, machine learning, and water research — driven not by any single discipline but by a deep need to understand the systems underlying everything he enc...

We train AI systems to optimize for metrics — but what if the real alignment problem isn't in the machine? What if it starts with us?

Juan Pablo Carbajal is a physicist, interdisciplinary researcher, and educator at the Eastern Switzerland University of Applied Sciences (OST). Originally from Argentina, his career spans physics, robotics, biomechanics, agronomy, machine learning, and water research — driven not by any single discipline but by a deep need to understand the systems underlying everything he encounters.

In Episode 58 of The Grand Challengers Podcast, Juan Pablo joins host Peter Marcus Bach to reframe the AI alignment debate as fundamentally a human one. He argues that the same narrow, single-metric optimization we apply to AI — rewarding output over understanding — is already distorting science, education, and society. From the "publish or perish" crisis in academia to how we evaluate students with one-dimensional scores, the pattern is the same: pressure systems to hit a number, and unintended behaviours emerge.

The conversation traces Juan Pablo's journey from collecting tadpoles in an Argentine river (with a memorable food-chain disaster) to studying embodied AI at the University of Zurich's AI Lab under Rolf Pfeifer, where intelligence is inseparable from its physical body. Along the way, he shares lessons from plant biology — how nitrogen-fixing bacteria and their "cheater" free-riders coexist through tolerance rather than punishment — and explains why physics-informed AI may be the antidote to purely data-driven approaches that mistake local patterns for universal truths.

Topics include: the limits of data-driven machine learning, Gestalt psychology and machine vision, why the word "robot" means slave, complex systems and emergence, the role of culture as humanity's alignment mechanism, and why understanding history is our best defence against repeating its mistakes.

Chapters:
0:00 Intro
1:59 Guest Intro & Yerba Mate
6:47 JP's Secret Mate Recipe
11:13 Summarizing JP's passions into a keyword
13:08 The Tadpole Story
17:24 JP's dilemma of what to study and his lifelong mentor
22:57 Versatility of a physicist and the Sokal affair
25:58 Critique of the current education systems
31:23 From physics to magnetic flux leakage detection
36:51 A life lesson on tolerance from bacteria
46:12 Embodied AI, Robotics and Studying Intelligence
54:07 Physics-informed AI
59:21 The challenge with prior knowledge - example from Gestalt Psychology
1:05:08 The Real Alignment Problem
1:22:49 JP's current exciting next steps
1:26:12 Q&A Start
1:26:43 What does innovation mean to you?
1:28:38 Key moment, book, person
1:31:10 Time Management
1:33:32 Favourite childhood memory
1:33:56 Biggest challenge to date
1:35:58 Advice for young professionals
1:36:35 What would you most like

Detailed shownotes over at: petermbach.com/podcast for more in-depth information about each episode.
Join the community over at: linkedin.com/company/tgcpodcast and let's connect across the world!

Subscribe and listen to the podcast (and do please leave a review/rating) on Spotify, Apple Music, YouTube or wherever you get your podcasts from.
Your subscription/follow will greatly help the show grow and reach a wider audience.

Hosted by Peter Marcus Bach (www.petermbach.com), follow me on: