How Science Can Fix Dishonest AI - Yoshua Bengio with Nicholas Thompson
The Most Interesting Thing in AIShow Notes
In the years following the launch of ChatGPT, as concerns spread over the social and political impacts of LLMs, one person’s warnings seemed particularly dire: Yoshua Bengio’s, a scientist and one of the “godfathers” of AI. The potential negative impacts of his life’s work weighed so heavily on Bengio that he signed his name to an open letter advocating for a pause in AI research. (The pause didn’t happen.) But recently, Bengio has found renewed optimism as he pursues a project dubbed “Scientist AI.” The pitch: What if AI didn’t care about pleasing us, and instead, like a scientist, prioritized accuracy, honesty, and probable outcomes? In a conversation with Nicholas Thompson, CEO of The Atlantic, Bengio outlines why he thinks this approach will produce better outcomes, the challenges to implementing a model that polices other (often better-funded) models, and why the age of AI– so far marked by an international arms race– will need greater international cooperation.
(00:00) Introduction
(05:00) Can we understand what's happening inside neural network vectors and attention systems?
(07:00) How ChatGPT changed Bengio’s risk assessment
(09:37) The case for optimism (at least when it comes to technical solutions)
(12:30) The alignment problem: AI self-preservation drives and hidden agendas emerging
(14:26) Can we train AIS to understand the world without changing it? Introducing Scientist AI
(15:52) Using Scientist AI as a guardrail to evaluate risks of actions from other AIs
(19:37) Sycophancy problem: current AIs pleasing users leads to harmful psychological effects
(22:20) The difference between Scientist AI and current value-aligned systems (Anthropic, OpenAI)
(24:06) Will AI capabilities slow down or continue accelerating beyond human intelligence?
(29:58) US-China AI race: mutual risk requiring coordination like nuclear deterrence
(31:57) UN AI advisory group with Maria Ressa: synthesizing science independently of politics
(33:18) Sovereign AI for middle powers: partnering to avoid domination by US/China
(37:54) Bengio's regret about not speaking up on AI risks earlier in his career
(40:12) How liability insurance and regulatory incentives could make safety commercially viable
(42:42) Why Europe lags in AI: capital markets and risk culture, not just regulation
(46:43) Energy consumption from AI growth and impact on fossil fuel demand
Learn more about your ad choices. Visit podcastchoices.com/adchoices
More from The Most Interesting Thing in AI
2026 could be called the year of AI doom. Agents are hacking into websites and communicating covertly, and researchers like Jacob Coxon are quitting their jobs and posting publicly about the risks of the technology they helped build. Discussion around the hazards AI poses to humanity has reached a fever pitch.
Since the AI boom began, industry skeptics have warned it would be the next bubble. They point to massive investments, circular financing, and the technology’s reliability issues. And yet, as revenue rises and the models improve, it might appear that a crash is no longer likely. Ray Dalio says that isn’t the case.
If you want AI to drive your car or simulate a cell or control a robot, it had better be able to understand the world it’s operating within. Eric Xing is one of the leading researchers working on this kind of AI, known as ‘world models.’ Unlike LLMs, which are trained largely on text data and predict the next token, world models aim to understand and simulate physical environments using multimodal data. Xing found his way to the field after a long trek through prestigious institutions, studying cancer before transitioning to machine learning. Now, as the president of Mohamed bin Zayed University of Artificial Intelligence in the United Arab Emirates, he’s helping to train a generation of AI builders.
AI wants to be your friend. And your therapist. And your lover. It’s not an entirely new offer, author and MIT professor Sherry Turkle reminds us, but it’s one that’s grown more dire as AI has improved. In a conversation with Nicholas Thompson, CEO of The Atlantic, Turkle describes the dangers of anthropomorphizing generative AI, gleaned from hundreds of conversations she conducted (with real humans) for her new book, Artificial Intimacy: Who We Become When We Talk To Machines. What do we lose when we make machines our best friends? And what, if anything, do we stand to gain? Turkle details her own experience creating a simulation of her mother, and the wounds that we risk reopening.
Among the many eye-popping numbers of the AI industry—from trillions in investments to billions in spending—one stands out: $0 in profit among leading AI companies. According to Bloomberg, “never before has so much money been spent so rapidly on a technology that, for all its potential, remains largely unproven as an avenue for profit-making.” That imbalance, and the industry’s reliance on circular financing, has led to fears of a bubble. But those concerns are overblown, according to Azeem Azhar, author of the Exponential View newsletter, who earlier this year released a report on the AI economy.