Nikhil KamathWTF is Artificial Intelligence Really? | Yann LeCun x Nikhil Kamath | People by WTF Ep #4
CHAPTERS
- 0:31 – 1:49
Yann LeCun’s early life near Paris and the long obsession with intelligence
Yann shares his upbringing in the suburbs of Paris, his engineer father’s influence, and how early curiosity in science and technology shaped his path. He frames his career as both understanding intelligence and building machines to uncover it.
- •Grew up near Paris; father was an engineer and major influence
- •Early interest in science/technology and seeing himself as an engineer
- •Core obsession: uncovering the mysteries of intelligence
- •Belief that building intelligent machines is a route to understanding intelligence
- •Potential societal consequences of intelligent machines
- 1:49 – 3:12
Engineer vs. scientist: understanding vs. creating (and why you need both)
Yann distinguishes the goals of science and engineering while arguing real progress often blends both. Technology enables new scientific discovery (telescopes, microscopes), and AI research sits at this intersection.
- •Scientists aim to understand the world; engineers aim to build new things
- •Scientific progress often depends on technological tools for measurement
- •AI requires both theory (what is intelligence?) and practice (build systems)
- •Building systems can test hypotheses about intelligence
- •Technology-development feedback loop drives discovery
- 3:12 – 4:06
From electrical engineering to math, research, and invention
Yann describes studying electrical engineering (not computer science) and gradually shifting toward fundamental questions in mathematics, physics, and AI. Independent projects with math professors pulled him into research and invention.
- •Studied electrical engineering; computers were present but not central initially
- •Increasing interest in fundamental math/physics/AI questions
- •Independent research projects with mathematics professors
- •“Favorite activity”: invent new things and understand them differently
- •Research mindset shaped by building + explaining mechanisms
- 4:06 – 5:22
“Godfather of AI,” credit in science, and being publicly visible
He rejects the “godfather” label, emphasizing science as collective progress where many contributors are forgotten. He also explains how public presence (talks, social media, opinions) amplifies prominence—sometimes positively, sometimes not.
- •Dislikes “godfather” framing; science isn’t an individual pursuit
- •Progress comes from collision of ideas across many people
- •Public memory over-attributes credit to a few names
- •His social media presence and policy opinions increase visibility
- •Prominence isn’t only about technical output—also public stance
- 5:22 – 7:40
Teaching, fame at NYU, and heroes in science
Yann reflects on teaching dynamics and the shift toward “celebrity” moments (students asking for selfies). He discusses how scientific heroes emerge, noting many major contributors remain unknown compared to figures like Einstein.
- •Interactive teaching matters more than passive lecturing
- •Growing recognition: selfies after class
- •Scientific “heroes” can exist, but many contributors remain invisible
- •Einstein as a public figure; Newton less so
- •Public engagement can shape who becomes widely known
- 7:40 – 11:14
Three big world problems through Yann’s lens: knowledge and intelligence deficits
Yann argues many global problems stem from inadequate understanding, poor decision-making, and failure to seek facts. He connects this to education and AI’s potential to amplify collective human intelligence to solve issues like climate change.
- •Uses causal models: identify effects and underlying causes
- •Many problems trace back to insufficient knowledge/mental models
- •Climate change as a key example with political + technological dimensions
- •Concern: people uninterested in facts or lacking access to information
- •AI and education as ways to make humanity “smarter” overall
- 11:14 – 14:35
What is AI? The blind-men-and-the-elephant view of intelligence
Yann reframes the question: to define AI, you first confront the multi-faceted nature of intelligence. Using the elephant analogy, he explains how AI historically over-focused on single slices—especially reasoning-as-search—while ignoring others.
- •Defining AI depends on defining intelligence—multi-aspect problem
- •Blind men/elephant analogy: partial perspectives dominate debates
- •Early AI (1950s) emphasized reasoning and logical search
- •Reasoning problems often reduce to optimization/search
- •Planning examples: traveling salesman, stacking objects, robot arm trajectories
- 14:35 – 20:42
Two historical branches: GOFAI (search/logic) vs learning/perception (neural nets)
He contrasts the dominant “reasoning as search” tradition (GOFAI) with the brain-inspired learning approach. The learning branch initially surged then stalled due to limitations, yet seeded pattern recognition and later deep learning.
- •GOFAI dominated until the 1990s; focused on logic, rules, search, planning
- •This tradition largely ignored perception/understanding of raw sensory input
- •Second branch: learning in networks of simple units (neurons)
- •Early neural nets had limited capabilities; field rebranded as pattern recognition/adaptive filters
- •Marvin Minsky’s pivot and Perceptrons critique shaped the field’s trajectory
- 20:42 – 28:34
Perceptrons and supervised learning: how early learning systems worked (and why finance is hard)
Yann explains the perceptron’s mechanism—weighted sums over pixels and iterative weight updates—and how this resembles linear regression. He then addresses a core challenge: systems can overfit retrospective financial data that may reflect randomness rather than stable structure.
- •Perceptron (1957): classify simple shapes via weighted sums + threshold
- •Training: adjust weights based on errors (increase/decrease)
- •Early practical use limited to simple pattern recognition
- •Supervised learning general principle: correct output guides parameter updates
- •Finance modeling risk: retrospective fitting vs inherently noisy/random processes
- 28:34 – 37:37
The AI “tree”: AI → machine learning → deep learning; types of learning (supervised, RL, self-supervised)
Yann lays out a clean taxonomy: AI as a broad problem space; GOFAI as rule/search methods; and machine learning as training from data, with deep learning as today’s dominant approach. He differentiates supervised learning, reinforcement learning, and self-supervised learning—highlighting why self-supervision fueled modern chatbots.
- •AI is a field/problem; techniques include GOFAI and machine learning
- •GOFAI: rule-based inference, logical deduction, graph search
- •Traditional ML: regression, trees, SVMs, boosting, Bayesian methods
- •Reinforcement learning: feedback is only good/bad; efficient for games via self-play but sample-inefficient
- •Self-supervised learning: predict missing/corrupted parts (text/image) without human labeling; foundation for LLM training
- 37:37 – 50:11
Neural nets, backprop, convnets, transformers: core building blocks explained simply
Yann gives intuitive explanations of multilayer neural networks and backpropagation, then moves to architecture choices: convolutional nets for natural signals and transformers for token sets with attention-like comparisons. He clarifies what “neurons” mean in ML and why architectural bias improves data efficiency.
- •Deep learning breakthrough: stacking nonlinear layers; backprop adjusts weights end-to-end
- •ConvNets: local receptive fields + shared weights; translation equivariance; inspired by visual cortex
- •Why architecture matters: leverages natural structure (neighboring pixels/audio similarity) to learn faster
- •Transformers: token-to-token comparisons; permutation equivariance; used heavily in language
- •Neuron as abstraction: weighted sum + nonlinearity; analogy to airplane wing vs bird wing
- 50:11 – 59:55
Language models from Shannon to modern LLMs: generation, n-grams, and scaling effects
He traces language modeling from Shannon’s next-symbol prediction and n-gram tables to neural language models and today’s transformer-based autoregressive LLMs. Yann highlights what scaling buys—massive parametric memory and fluent language manipulation—while underscoring that much of it is retrieval rather than deep reasoning.
- •Claude Shannon: predict next letter; conditional probabilities enable generation
- •n-grams: context length increases table size exponentially (sparsity + memory limits)
- •Neural language models (Bengio): predict next word via neural net distribution over vocabulary
- •Modern LLMs: transformers trained on large internet-scale corpora; autoregressive generation loop
- •Emergent capabilities: impressive syntax/grammar; limitations: minimal real reasoning and brittle mistakes
- 59:55 – 1:04:45
Why LLMs hit a wall for physical reality: continuous worlds, video prediction, and memory limits
Yann argues LLMs excel in discrete domains (text) but struggle with continuous, high-dimensional reality (images/video). He explains LLM memory as parameters + context window and outlines the need for persistent, hippocampus-like memory and world models for planning and robotics.
- •LLMs work best for discrete tokens; physical world is continuous/high-dimensional
- •Video prediction requires modeling distributions over enormous pixel spaces—mathematically intractable in naive form
- •LLMs lack true understanding of physics; can pass exams but fail common-sense world tasks (cat comparison)
- •Two LLM “memories”: learned parameters and limited prompt context/working memory
- •Need persistent memory and world models to enable planning and system-2 reasoning
- 1:04:45 – 1:12:52
JEPA and objective-driven AI: predicting in representation space (world models + planning)
He introduces JEPA (Joint Embedding Predictive Architecture) as an alternative to pixel-level generation: encode past and future into abstract embeddings and predict in that space. This supports hierarchical, longer-term prediction and planning, linking learned world models to system-two reasoning and search over action sequences.
- •JEPA replaces pixel prediction with prediction of abstract encoded representations
- •Encoders compress away unpredictable details; prediction happens in representation space
- •Longer-horizon prediction requires more abstract representations
- •World model framing: state + action → next state enables planning sequences toward goals
- •System 1 vs System 2: LLMs as reactive system-1; objective-driven AI aims at deliberative system-2
- 1:12:52 – 1:36:05
Future outlook, data/compute infrastructure, and practical advice for India: builders, investors, and society
Yann forecasts a mostly utopian trajectory where new architectures plus better memory and world modeling could reach human-level intelligence within ~5–10 years (optimistically). He stresses data quality, multilingual/cultural coverage, distributed training, the rise of open-source platforms, and the importance of local compute (data centers/inference). He closes with concrete guidance for entrepreneurs and investors and a redefinition of human intelligence in an AI-rich world.
- •Timeline: human-level intelligence possibly within a decade, but not by scaling LLMs alone
- •LLM pipeline evolves: ongoing need for high-quality data filtering; need broader multilingual/cultural datasets
- •Open source platforms likely to dominate (Linux analogy); distributed training across regions
- •India context: local compute is crucial for training sovereignty and ultra-low-cost inference at scale
- •Career/business advice: build on open-source foundation models (e.g., Llama), fine-tune for verticals (legal, accounting, fintech, internal enterprise search, health, education)
- •Human intelligence shifts upward: deciding what to do; humans “manage” AI assistants; jobs evolve not disappear
- •Final definition: intelligence = collection of skills + fast learning + zero-shot problem solving
