← Back to list
Lex Fridman PodcastPodcast8 May 2020Source: lexfridman.comHost: Lex Fridman

#94 – Ilya Sutskever: Deep Learning

In plain words

This is a deep interview with OpenAI chief scientist Ilya Sutskever. His main point: deep learning, a simple brain-inspired idea, really works and can approach human intelligence with enough data and compute, and we keep underestimating it. He thinks AGI needs "deep learning plus maybe another small idea," like self-play (AI playing against itself to generate surprising creativity). Key mentions: GPT-2 (a milestone language model), AlphaZero (chess AI showing neural nets can reason), and OpenAI DotaBot (multi-agent system showing self-play's surprising abilities).

AI SummaryAI-generated · may contain errors · verify against the original

This report summarizes Lex Fridman’s in-depth interview with OpenAI co-founder and chief scientist Ilya Sutskever, focusing on the development and future of deep learning. The core argument is that Sutskever, a top computer scientist with over 165,000 citations, emphasizes the breakthrough progress

~12 min full read · 8 sections
Deep Analysis

At a Glance

Ilya Sutskever (OpenAI co-founder and chief scientist, with over 165,000 citations) engaged in an in-depth discussion with Lex Fridman on the essence, breakthroughs, and future of deep learning. The most weighty takeaway from the entire episode: Sutskever believes the most astonishing aspect of deep learning is that "it actually works" — a simple idea inspired by the brain (neural networks + backpropagation), given sufficient data and computing power, can approximate the core functions of human intelligence, and each time it has been underestimated, it has proven capable of going even further.


I. The Catalyst of the Deep Learning Revolution: Data, Compute, and Conviction

Sutskever believes that 2010-2011 was a critical turning point. At that time, he connected two facts: first, James Martens successfully trained a 10-layer neural network using the Hessian Free Optimizer (without pre-training); second, he realized that a "10-layer network" was analogous to the number of neural layers (approximately 10) in the human brain that complete object recognition within 100 milliseconds. This convinced him that as long as a sufficiently large neural network could be trained with backpropagation and enough supervised data, it would succeed.

"The main doubt was, can we train a big, will we have enough compute to train a big enough neural net? With backpropagation, I thought it would work."

Key catalyst: Alex Krizhevsky wrote an extremely fast CUDA convolutional neural network kernel, and combined with the emergence of the ImageNet dataset, these three elements ignited the deep learning revolution. Sutskever emphasized that prior to this, neural networks were "indeed not the best method for almost all tasks," so "hard benchmarks" were needed to provide irrefutable evidence and put an end to endless debates.


2. Brain vs. Artificial Neural Networks: Differences and Insights

Sutskever argues that the brain surpasses current models in certain dimensions, but artificial neural networks also hold significant advantages. He evaluates several key differences:

  • Spikes: The brain uses spike signals, but Sutskever believes this "may not be that important"—because researchers studying spiking neural networks find they "need to simulate non-spiking neural networks within spikes to make them work," otherwise there is no reason to assume they would be effective.
  • Cost Function: This is the "first profound idea" in deep learning. Sutskever is a strong proponent of cost functions: "I would not bet against cost functions." He notes that GANs are the only successful example without an explicit cost function—they function more like a "game," with behavior reasoned through game equilibrium rather than a cost function.
  • Temporal Dynamics: The brain uses spike-timing-dependent plasticity (STDP)—synapses strengthen if activated before a neuron fires and weaken if activated after. Sutskever believes that recurrent neural networks (RNNs) can capture similar phenomena, but they have since been replaced by Transformers, though "some form of recurrence is likely to return."

Unique Analogy: Sutskever describes deep learning as "the geometric mean of biology and physics"—biological theories are complex with weak predictive power, physical theories are overly precise, and deep learning sits somewhere in between.


III. The Trend Toward Unification: From Fragmentation to a Single Architecture

Sutskever believes that the field of machine learning exhibits "great unification," with only one, two, or three very simple principles that apply in nearly the same way across different modalities and problems.

  • Unification in NLP: A few years ago, each NLP subproblem had its own architecture; today, a single Transformer handles all tasks.
  • Unification of Vision and Language: Sutskever argues that "one day, vision will also be unified with natural language," although convolutional neural networks still hold an advantage in computational efficiency.
  • The Distinctiveness of Reinforcement Learning: RL must handle non-stationary environments (where actions change experiences) and exploration problems, resulting in higher variance. However, he expects "a broader unification between RL and supervised learning," ultimately becoming a "big black box"—you just throw things in, and it handles everything on its own.

On the Difficulty of Language vs. Vision: Sutskever initially hesitated but ultimately concluded that "language may be harder"—if defined as "100% language understanding." He also notes, however, that the boundary between vision and language is blurred (is reading vision or language?), and it is likely that "getting one will get you the other for free."


4. Unsolved Mysteries and Future Directions of Deep Learning

4.1 Deep Double Descent

Sutskever explains this counterintuitive phenomenon: When fixing a dataset and gradually increasing the size of a neural network, performance first improves rapidly, then suddenly deteriorates upon reaching zero training error, before improving again. The intuition is: when the degrees of freedom in the data equal those in the model, the model becomes extremely sensitive to minor randomness in the data; conversely, when parameters far outnumber data points or data far outnumber parameters, the solution is insensitive to data variations.

Key detail: Without early stopping, the double descent phenomenon is very pronounced; with early stopping, it nearly disappears.

4.2 Reasoning Capability

Sutskever believes neural networks are already capable of reasoning, citing two pieces of evidence: AlphaZero's neural network (without search) plays Go at a level exceeding 99.9% of humans; humans themselves serve as proof of existence. "Humans can reason. So why can't neural networks?"

However, he points out a key limitation: neural networks only "solve the problem you put in front of them in the easiest way possible"—if the training task does not require reasoning, they will not reason. He looks forward to future reasoning benchmarks including: writing genuinely good code, proving unsolved theorems, and solving open-ended problems.

4.3 Active Learning and Data Selection

Sutskever argues that active learning requires a "task that truly needs it" to achieve a breakthrough. Demonstrating active learning on MNIST is no longer convincing enough. He anticipates that future language models will "use their own intelligence to decide which data to accept and which to reject"—just as humans do selectively.

4.4 Self-Play

Sutskever believes self-play will be a key component of AGI, because it has the property of "surprising us in genuinely novel ways"—DotaBot, Hide-and-Seek, and AlphaZero all produced unexpectedly creative solutions. "That seems like an important part of AGI that our systems don't exhibit routinely right now."


5. The Path to AGI and Alignment

5.1 What Is Needed?

Sutskever’s answer is concise and powerful: “Deep learning plus maybe another small idea.” He believes self-play will be one of those ideas.

On embodiment: He argues that a body is “useful but not necessary”—much like Helen Keller, who, despite being blind and deaf, was still able to compensate and succeed.

5.2 Alignment and Control

Sutskever paints an optimistic vision of AGI governance: Humans act as a corporate board, while AGI serves as the CEO—residents of different cities or countries vote to elect AGIs that represent them, and the AGI executes accordingly. The key is that “the board can press the reset button.”

He firmly believes: “It’s definitely possible to build AI systems which will want to be controlled by their humans.” By analogy: Human parents are usually happy to help their children; it is not a burden but a joy. Similarly, AGI can be designed so that its “deep drive” is to “help humanity thrive.”

On power: When asked whether he would be willing to relinquish control over AGI, Sutskever replied: “I’d find it trivial to do that. The kind of scenario you are describing sounds terrifying to me.”


Mentioned Positions

Position Analyst View Key Data
GPT-2 Positive (milestone progress) 1.5 billion parameters, trained on approximately 40 billion tokens (from web pages with Reddit links receiving ≥3 upvotes)
AlphaZero Positive reference (evidence of reasoning ability) Neural network (without search) outperforms 99.9% of humans in Go
OpenAI DotaBot Positive reference (surprising capability from self-play) Multi-agent system, generating unexpected behaviors
OpenAI Hide-and-Seek Positive reference (creativity from self-play) Two agents playing hide-and-seek, producing unexpected strategies
OpenAI Robotic Hand (Rubik's Cube) Positive reference (sim-to-real success case) 100% trained in simulation, policy trained to be highly adaptive

Judgments Worth Remembering

1. “The most beautiful thing about deep learning is that it actually works.” — Sutskever believes that a simple idea inspired by the brain, when given enough data and compute, can approach human intelligence, which he finds “unbelievable.”

2. Deep learning is the “geometric mean of biology and physics.” — Sutskever’s own framework: biological theories are complex with weak predictive power, while physical theories are overly precise; deep learning sits between the two.

3. “We keep underestimating deep learning.” — Sutskever notes that every year people think “this is the peak,” but the following year it goes even further.

4. Neural networks can reason, but only “in the easiest way to solve the problem you put in front of them.” — If the training task does not require reasoning, they will not reason. This is a key gap between current systems and humans.

5. The unique value of self-play lies in “surprising us in truly novel ways.” — Sutskever believes this is an important component of AGI, and current systems cannot yet demonstrate it routinely.

6. “It’s definitely possible to build AI systems which will want to be controlled by their humans.” — Sutskever uses the analogy of parents willingly helping their children to argue that AGI can be designed with a deep drive to “help humanity thrive.”

7. The intuition behind Deep Double Descent: When the degrees of freedom in the data equal those in the model, the model becomes extremely sensitive to small randomness in the data; when parameters far exceed data or data far exceed parameters, the solution is insensitive to data variation.

8. The success of Transformer is “a simultaneous combination of multiple ideas.” — Attention is just one of them; the other two key factors are: it runs extremely fast on GPUs and is non-recurrent (shallower and easier to optimize).