Dec 18, 2014 · 48m · mad
Yann Lecun, Facebook // Artificial Intelligence // Data Driven #32 (Hosted by FirstMark Capital)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Data Driven NYC, host Matt Turck interviews AI pioneer Yann LeCun about his seminal work on convolutional neural networks, Facebook's AI strategy, the technical foundations of deep learning, and practical advice for AI startups.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 12.7% of the talking time here. How this is scored →
Matt as informed peer 2.6 Guest teaching 5.3 Guest disagreement 2.3 Matt pushing back 0.7
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 43:50 Yann dismisses biological neural simulation projects as 'nuts'
Yann delivers his strongest criticism of the interview, calling efforts to simulate full biological brain details and build spiking neural hardware 'completely nuts' despite massive funding.
Hardest push from Matt ▶ 35:42 Matt confronts Yann with Elon Musk's warnings on AI riskMatt directly challenges the optimistic narrative around AI by reciting Elon Musk's quotes about AI being an existential threat and 'summoning the demon.'
Biggest teaching moment ▶ 42:50 Yann uses the airplane versus bird analogy for AI designYann reframes the relationship between neuroscience and artificial intelligence, explaining that copying biological features like feathers or spiking neurons ignores the broader underlying principles of aerodynamics and intelligence.
Matt holds his own ▶ 22:17 Matt demonstrates technical understanding of backpropagationMatt shows strong domain familiarity by specifically framing backpropagation as a system that enables neural network nodes to learn from individual errors rather than global averages.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Yann LeCun's Career Journey and Early Neural Networks | 1 | 5 | 1 | 0 | Matt largely steps back as Yann recounts his career from electrical engineering in France to Bell Labs, AT&T, Google, NYU, and establishing the CIFAR deep learning 'conspiracy'. The host mostly offers brief conversational prompts. Yann educates the audience on the history of SVMs versus convolutional neural networks. | |
| AI Strategy and Content Understanding at Facebook | 1 | 4 | 0 | 0 | Matt asks a direct query about Facebook's AI strategy, allowing Yann to give an extended overview of content filtering and user interest matching. Yann explains the necessity of selecting 150 relevant posts out of 2,000 using automated understanding. The host accepts the explanation without pushing back. | |
| The Resurgence of AI and Technical Foundations of Deep Learning | 3 | 5 | 1 | 1 | Matt shows background knowledge by citing key industry moves such as Andrew Ng at Baidu and Google's DeepMind acquisition. Yann explains how deep learning automates feature engineering across hierarchical layers. The discussion remains open and collaborative. | |
| Computational Scale, Backpropagation, and Natural Language Processing | 4 | 6 | 2 | 1 | Matt asks informed technical questions about backpropagation error calculation for individual nodes and GPU computational scaling. Yann reframes backpropagation as an application of the chain rule and gradient descent before detailing natural language word embeddings like Word2Vec. | |
| Open Research Culture and AI Startup Opportunities | 4 | 4 | 1 | 1 | Matt demonstrates industry knowledge by quoting Jeff Hinton and referencing specific AI startups like X.AI and Vicarious. Yann details why the opportunity window for pure AI algorithm startups is closing as talent matures and large tech companies absorb experts. | |
| Addressing AI Safety, Ethics, and Existential Risk | 3 | 6 | 5 | 2 | Matt presses Yann on existential threats by reading long quotes from Elon Musk regarding AI regulation and 'summoning the demon'. Yann firmly rejects Musk's alarmist framing, distinguishing intelligence from autonomy and comparing current AI development to driving carefully through fog. | |
| Audience Q&A: Compute Bottlenecks, Neuroscience Inspiration, and Learning Resources | 2 | 7 | 6 | 0 | During audience Q&A, Yann forcefully attacks approaches that attempt to simulate biological brain details or build spiking neural network chips, calling them 'nuts' and misguided. He uses a detailed bird-versus-airplane analogy to explain why AI engineering should focus on underlying principles rather than copying biology. |
Statements from this episode (17)
LeCun: Deep learning's rapid industry adoption surprised early AI pioneers
“I mean, the whole success of deep learning, actually has taken some of us by surprise. We were all sort of Convinced from a long time ago that deep learning was going to take off at some point, but the speed at which has been picked up by, ah, industry and res…”
LeCun: AT&T's neural net read 20% of US checks in late 1990s
“Which at some point in the nineties were, in the late nineties, was reading something like 20% of all the checks in the U.S.”
LeCun turned down Larry Page's offer to direct Google's research lab
“And they offered me the job, and I turned it down.”
LeCun predicted in 2014 that AI would mediate human social interactions
“And started thinking about the next 10 years. What are the next 10 years going to be for social interactions? And it's pretty obvious to a lot of people that a lot of our interactions you know, with our friends and a lot of interactions with the digital world …”
LeCun: Apple, Android, and Microsoft mobile speech recognition use deep learning
“So probably many of you know that most You know, mobile speech recognition application search and stuff like that uses deep learning for the last two years. This is both for, you know, Apple, Android, Microsoft. All of those applications use deep learning for …”
LeCun: Deep learning automates feature engineering by integrating representation with classification
“And deep learning basically opens the possibility that you, to automate the process of feature engineering. So, it essentially learns to, It integrates the process of producing good representations with the process of learning a classifier in one fell swoop.”
LeCun: ConvNet training times dropped from two weeks to two days
“The first large convolutional nets that are appropriate for object recognition came up about two and a half years ago, and it would take about two weeks to train. The latest versions with the latest software implementations, we can train in about two days.”
LeCun predicted in 2014 deep learning would conquer natural language processing
“There's a kind of a sense that in the community that the next set of techniques to kind of fall to deep learning, if you want, will be natural language processing.”
LeCun: Facebook stole Word2Vec creator Tomasz Mikolov from Google
“I should mention that Tomasz Mikolov now works at Facebook AI Research. I stole him from Google.”
LeCun: Industry adopted deep learning faster than academia
“Industry picked up on deep learning faster than academia.”
LeCun: The window for deep learning startups will close quickly
“The window is going to close really quickly on, on deep learning you know, sort of opportunity for startups in deep learning.”
LeCun: Building superintelligent AI will take a long time
“Building machines that are super intelligent, you know, more intelligent than humans is going to take a long time.”
LeCun: The public thinks AI is more dangerous than it is
“The public thinks it's more dangerous than it actually is, so there is a whole segment of the public that will be extremely scared about the potential of AI, and that need, you know, the real dangers and the false dangers need to be explained.”
LeCun: Backprop designs vision systems better than human engineers
“In fact, those dumb learning algorithms, Backprop, namely, is better at designing vision systems than very experienced engineers and scientists are at designing vision systems.”
LeCun: Deep learning designs NLP systems better than linguists
“You know, those deep learning algorithms are better at designing NLP systems than linguists are, or natural language specialists.”
LeCun: Simulating biological brain neurons to achieve AI is 'completely nuts'
“There are people out there in Europe that, that have argued for the fact to kind of study neurons and neurons and neural networks natural neural networks in all the details, and then build giant supercomputers to simulate this very accurately, and somehow they…”
LeCun: Building chips for spiking neural networks is misguided
“The problem is, from the engineering point of view, nobody has actually demonstrated that spiking neurons actually work. Like there is no image recognition systems that are based on spiking neural networks. Why build chips for something we don't know works? So…”