"if you want to make decisions in the real world our perception is limited so we are always uncertain about the state of the real world and we need to make decisions under uncertainty... fundamentally to build an intelligent system you need a system that can represent uncertainty that can update its uncertainty and then can use that to make good decisions under uncertainty" - Zubin Ghahramani [00:01:40]
"and in fact that's the definition of information right so information a bit of information that we use in computer science is the reduction of your uncertainty by a factor of two that's what a bit is and so collecting information is the way we reduce our uncertainty" - Zubin Ghahramani []
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"that sort of adversarial example shows us that um it's not correctness that we care about alone it's actually correctness and confidence we don't want systems that can be overconfidently wrong" - Zubin Ghahramani [00:08:32]
"I'm pretty boring in that I was interested in AI as a teenager so when I was 14 or 15 I already wanted to work in AI... neural networks were the counterculture people were very um enamored of the old school of AI" - Zubin Ghahramani [00:09:58]
"probabilities are at the heart of everything we do in machine learning but what's missing is we're not really doing what I said which is the careful representation of probabilities we're actually sort of hoping that the models represent probabilities okay because we've trained them on enough data" - Zubin Ghahramani [00:15:13]
"I would actually like our AI systems to be you know more rational better at representing and manipulating probabilities than humans are" - Zubin Ghahramani [00:19:28]
"for all problems that matter I would rather have an AI system that knows when it doesn't know than an AI system that is arrogant and overconfident" - Zubin Ghahramani [00:43:43]
Speakers & Credentials
Hannah Fry (Host): Professor, mathematician, author, and radio/television presenter specializing in the patterns of human behavior and mathematics.
Zubin Ghahramani (Guest): Professor of Information Engineering at the University of Cambridge, Co-Lead of Frontier AI at Google DeepMind, and former Vice President of Research at Uber. Pioneer in Bayesian machine learning and probabilistic AI.
1. Executive Summary
Contemporary artificial intelligence models display unwavering authority even when producing factually incorrect outputs, highlighting a critical lack of machine self-doubt 00:00:05.
Intelligent decision-making requires systems that can represent, update, and operate under probabilistic uncertainty due to incomplete sensory information 00:01:40.
Machine learning must distinguish between aleatoric uncertainty (inherent randomness in the environment) and epistemic uncertainty (lack of knowledge reducible by collecting data) 00:04:18.
Adversarial vulnerability demonstrates that modern neural networks optimize for correctness without properly calibrating confidence, making them prone to high-confidence errors 00:08:08.
Large Language Models (LLMs) operate as next-token predictors using statistical approximations, frequently "faking" confidence rather than executing rigorous probabilistic inference 00:20:08.
Groundbreaking applications like Google DeepMind's GenCast (weather forecasting) and AlphaFold (protein structure prediction) succeed explicitly because they embed probabilistic uncertainty representations [00:26:53, 00:30:16].
Beyond pure compute and dataset scaling, reaching true Artificial General Intelligence (AGI) requires architectural breakthroughs in Bayesian continual learning, energy efficiency, and data efficiency 00:36:56.
The Imperative of Machine Uncertainty in Decision-Making
Perception across all biological and synthetic intelligent systems is inherently constrained by incomplete environmental telemetry, requiring decision-making under fundamental uncertainty 00:01:40.
Autonomous navigation systems illustrate the critical need for self-awareness of internal limits; when encountering rare edge cases ("long tail" scenarios) like a sudden hailstorm combined with horses leaping into traffic, a system aware of its own epistemic limits safely decelerates 00:02:41.
Machine learning divides uncertainty into aleatoric uncertainty (irreducible environmental randomness, such as a pedestrian's sudden turn direction) and epistemic uncertainty (reducible parameter or model ignorance) 00:04:18.
Epistemic uncertainty is systematically mitigated through information acquisition; in information theory and computer science, one bit of information is formally defined as the reduction of uncertainty by a factor of 2 00:04:44.
Cognitive Science, Human Perception, and Probabilities
Unconscious biological perception operates via probabilistic inference, synthesizing prior knowledge with noisy sensory input to maintain environmental models 00:05:36.
Evolution hardwires survival heuristics; for example, a hiker hearing rustling leaves in Costa Rica immediately factors in the prior probability of jaguars, whereas the same sound in London produces a completely different prior distribution 00:05:56.
While human perceptual and motor systems execute precise implicit probabilistic inference, human conscious cognition is notoriously deficient at explicit mathematical probability estimation, frequently falling victim to cognitive fallacies documented by Kahneman and Tversky [00:06:28, 00:06:49].
Developmental psychology and cognitive science reveal that infants and toddlers learn implicitly using identical probabilistic and Bayesian formalisms to build predictive world models without explicit mathematical formulation 00:07:08.
Correctness, Confidence, and Adversarial Vulnerabilities
High predictive accuracy is insufficient for deployment; artificial intelligence models must simultaneously provide well-calibrated confidence metrics 00:08:08.
Imperceptible pixel perturbations applied to an image of a school bus can trick a deep neural network into classifying it as a cheetah with 99% confidence, demonstrating dangerous miscalibration 00:08:08.
Current Large Language Models exhibit high confidence coupled with structural fragility; when challenged by a human user, they routinely flip-flop or offer obsequious apologies rather than holding grounded probabilistic positions [00:09:08, 00:20:32].
Rather than maintaining explicit probability distributions over internal states, large language models mimic reasoning trajectories seen during pre-training on trillions of tokens [00:15:13, 00:20:20].
Historical Evolution of AI and the Bayesian Paradigm
During the late 1980s, classical AI was dominated by brittle, rule-based expert systems, while neural networks represented an underfunded counterculture movement [00:10:52, 00:12:26].
In 1986, the publication of the two-volume Parallel Distributed Processing and the foundational backpropagation paper catalyzed an early neural network wave 00:10:21.
Early artificial neural networks in 1989 were trained on dedicated hardware platforms like the Thinking Machines Connection Machine—a parallel supercomputer housing 65,000 processors featuring blinking red lights inside an opaque cube, which possessed far less computational power than a modern Google Pixel smartphone [00:12:04, 00:41:43].
Bayesian machine learning applies Bayes' Rule, taking a prior probability distribution, multiplying it by the likelihood of observed evidence, and renormalizing the product to yield an updated posterior distribution 00:16:12.
Sequential Bayesian updating operates as a formal model for both continuous perceptual inference and parameter learning without information destruction [00:17:41, 00:18:06].
Scaled Probabilistic Engineering: Weather, Biology, and LLMs
Google DeepMind's GenCast weather forecasting system generates 15-day weather predictions in 8 minutes (versus hours on traditional supercomputers) by combining diffusion models with ensemble probability trajectories to track severe events like Hurricane Melissa 00:26:53.
AlphaFold models protein folding uncertainties directly by outputting color-coded structural confidence maps (pLDDT scores), differentiating rigid molecular structures from flexible, wiggling physical domains 00:30:16.
In LLM architecture, internal model uncertainty can be estimated by calculating the semantic entropy or output distribution variance over next-token prediction vectors [00:23:17, 00:23:44].
Probability calibration dictates that if a weather model predicts a 70% chance of rain across a set of historical days, it must have rained on exactly 70% of those days 00:31:28.
The AGI Debate: Architectural Innovation vs. Pure Scale
The modern AI landscape is split between scaling proponents (who argue scale and data additions suffice) and architectural innovators who insist AGI demands fundamental structural changes 00:36:22.
Modern pre-training workflows suffer from catastrophic forgetting; human-like continual learning requires Bayesian updating where posterior distributions seamlessly become new priors for incoming data streams [00:37:06, 00:38:01].
The human brain operates on an energy footprint of approximately 20 watts (equivalent to an energy-efficient light bulb), standing in contrast to the massive megawatt data centers required to train LLMs 00:38:38.
Bayesian methods offer a principled mathematical path toward drastic improvements in data efficiency, enabling systems to learn effectively from sparse inputs 00:39:50.
The central bottleneck of exact Bayesian inference is computational intractability (NP-hard/complete problems), forcing modern researchers to pioneer hardware-software co-evaluations and scalable approximations [00:40:48, 00:41:10].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Bit Definition (Information Theory)
2x Uncertainty Reduction
Reduction of system uncertainty by a factor of 2 per bit of information gained.
Bayesian Updating & Rational Belief Revision [00:16:12]
Bayesian updating establishes that rational belief under uncertainty requires combining prior probability distributions with incoming evidence likelihoods to yield a posterior distribution. In contemporary machine learning, modern LLMs disregard explicit Bayesian mechanics, attempting to fake reasoning by sampling from pre-trained static token distributions. Ghahramani asserts that true AI safety, continual learning, and rational decision-making require hard-coding probabilistic inference directly into system architectures to prevent catastrophic forgetting and flip-flopping under pushback.
Aleatoric vs. Epistemic Uncertainty Division [00:04:18]
This framework divides system doubt into aleatoric uncertainty (stochasticity inherent in the physical universe, like a coin toss) and epistemic uncertainty (a system's parameter ignorance due to missing data). When applied to mission-critical AI like self-driving vehicles or clinical diagnosis, recognizing epistemic uncertainty directs the model to pause, slow down, or query additional sensors rather than making dangerous, overconfident predictions.
Semantic Entropy & Internal Probability Calibration [00:23:17]
Semantic entropy measures the spread and flatness of an LLM's next-token output distribution prior to sampling. A low-entropy, sharp distribution signifies high internal model confidence, whereas high-entropy distributions flag epistemic confusion. Calibrating this metric allows developers to expose genuine model doubt to end-users rather than relying on natural-language outputs that fake certainty.
Continuous Learning via Posterior Progression [00:37:06]
Standard deep learning models require batch training phases, making them susceptible to catastrophic forgetting when updated naively. Posterior progression uses the current posterior distribution as the immediate prior for incoming real-time telemetry. This mathematical framework enables AI to continuously adapt to streaming real-world environments without corrupting past foundational knowledge.
6. Anecdotes
The Connection Machine Supercomputer (1989) [00:12:04]
Ghahramani recalls coding early natural language parsing neural networks on the Connection Machine—a legendary parallel supercomputer containing 65,000 processors inside a dark, opaque cube with thousands of blinking red lights. He shares this memory to highlight the dramatic leap in compute scaling, noting that this multi-million-dollar supercomputer was significantly slower than a modern smartphone.
Adversarial School Bus Manipulation [00:08:08]
Researchers altered a few imperceptible pixels on a digital photograph of a yellow school bus. While any human viewer still clearly saw a school bus, the deep neural network classified the modified image as a cheetah with 99% statistical confidence. Ghahramani cites this anecdote to illustrate how deep learning models lack properly calibrated confidence metrics.
Jaguars in Costa Rica vs. London [00:05:56]
Ghahramani presents a hypothetical scenario of walking through a jungle in Costa Rica versus a park in London and hearing rustling leaves. A human's perceptual system automatically applies prior geographical knowledge: in Costa Rica, the prior probability of a jaguar is non-zero, triggering immediate self-preservation responses. In London, the prior probability drops to near zero. He uses this example to show how unconscious human perception relies heavily on Bayesian priors.
The Self-Driving Car Long-Tail Encounter [00:02:41]
Ghahramani describes an autonomous vehicle navigating a rare edge case: a sudden, blinding hailstorm where a horse leaps in front of the vehicle. Standard neural networks trained exclusively on typical driving telemetry struggle to handle such rare combinations. An AI model equipped with calibrated epistemic uncertainty recognizes that it is in an unfamiliar domain and safely slows down.
7. References & Recommendations
Academic Institutions & Organizations
Google DeepMind [00:00:00] - Premier industrial AI research institute co-leading frontier model development.
University of Cambridge [00:00:34] - Academic institution where Zubin Ghahramani serves as Professor of Information Engineering.
University of Pennsylvania [00:10:08] - University where Ghahramani completed his undergraduate studies in computer and cognitive science.
Books & Publications
Parallel Distributed Processing (1986) [00:10:21] - Foundational two-volume book series by Rumelhart and McClelland that launched modern neural network research.
Backpropagation Paper (1986) [00:10:28] - Landmark paper co-authored by Geoffrey Hinton popularizing backpropagation in artificial neural networks.
Nature Review Paper on Probabilistic Machine Learning (2015) [00:14:28] - Seminal perspective paper written by Ghahramani outlining the central role of uncertainty in artificial intelligence.
People
Arvin Joshi [00:10:08] - Renowned computational linguist at Penn who hired young Ghahramani to study Parallel Distributed Processing.
Geoffrey Hinton [00:13:48] - Nobel Laureate and deep learning pioneer with whom Ghahramani worked closely on probabilistic deep learning models.
Daniel Kahneman & Amos Tversky [00:06:49] - Nobel Prize-winning cognitive scientists who documented human cognitive fallacies and poor explicit probability estimations.
David Spiegelhalter [00:34:10] - Cambridge statistician recognized for pioneer research on risk and uncertainty communication.
Media & Pop Culture
Data (Star Trek: The Next Generation) [00:19:09] - Referenced as an ideal representation of a hyper-rational, non-flip-flopping synthetic intelligence.
Technologies, Hardware & Software Systems
Connection Machine [00:12:04] - Highly parallel supercomputer from the late 1980s equipped with 65,000 microprocessors.
GenCast [00:26:53] - Google DeepMind's state-of-the-art ensemble diffusion model for medium-range probabilistic weather forecasting.
AlphaFold [00:30:16] - AI system developed by Google DeepMind that predicts 3D protein structures alongside output confidence scores.
Google Pixel [00:41:52] - Modern consumer smartphone cited to illustrate compute progress relative to 1980s supercomputers.
Sep 3, 2026
EP. 20 | 10,000 Pitches, 55,000 Deals, and AI: How Pranav Pai Finds Winning Startups | 2 Sept 2026 | Clearing The BLUR
"My only regret since '91 is we have not become antisocialist." Pranav Pai 00:04 http://www.youtube.com/watch?v=lPkjuRztHcg&t=0m4s "The assumptions the government makes—every Indian businessman is a crook—is a worst assumption you can make…
Connection Machine Parallel Processors
65,000
Number of individual microprocessors inside the Connection Machine cube used in 1989.