May 17, 2017 · 54m · y-combinator
An AI Primer with Wojciech Zaremba · Y Combinator
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 Y Combinator interview, OpenAI co-founder Wojciech Zaremba discusses the core principles, historical breakthroughs, and practical challenges of deep learning, while sharing OpenAI's mission to develop safe Artificial General Intelligence (AGI).
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 2.0 Guest teaching 6.5 Guest disagreement 0.1 The partners pushing back 0.0
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 47:43 Reframing AI Hype as Simultaneously Over- and Under-Done
Zaremba rejects simple binary hype framings by immediately asserting that AI is simultaneously overhyped for general tasks and underhyped for supervised learning applications.
Hardest push from the partners ▶ 53:13 Cannon Pressing on Purpose Beyond AutomationCannon adds nuance to Zaremba's UBI perspective by highlighting that finding purpose will be the core challenge if widespread job displacement occurs.
Biggest teaching moment ▶ 17:44 Mathematical Necessity of Nonlinearity in Neural NetworksZaremba thoroughly educates Cannon on linear operator composition, showing why sequential matrix multiplications collapse into a single matrix without non-linear activations.
The partners hold their own ▶ 26:04 Connecting Speech Recognition Audio to Waveforms and Fourier TransformsCannon demonstrates technical comprehension by connecting Zaremba's speech-as-an-image concept directly to audio waveforms and visual representations.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Wojciech Zaremba's Academic and Professional Background | 1 | 5 | 0 | 0 | Cannon prompts Zaremba for his academic background and an overview of OpenAI's mission and projects. Zaremba explains the technical difficulty of robotic manipulation compared to locomotion and navigation in an instructional manner. | |
| OpenAI Projects: Game Playing and Reinforcement Learning Challenges | 2 | 7 | 0 | 0 | Cannon asks clarifying questions regarding reinforcement learning in game playing. Zaremba gives an in-depth breakdown of reward functions, environment resets, sample efficiency issues, and the contrast between Atari games and real-world robotics. | |
| Defining AI, Machine Learning, and Deep Learning | 2 | 8 | 0 | 0 | Cannon asks foundational definitions of AI, machine learning, and deep learning. Zaremba delivers a masterclass explanation contrasting multi-step computation with single-step statistical models and shallow wide classifiers. | |
| Understanding Neural Network Architectures and Activation Functions | 2 | 8 | 0 | 0 | Cannon asks for a breakdown of neural networks. Zaremba explains linear matrix multiplications, composition, and why non-linear activation functions like Sigmoid and ReLU are mathematically necessary. | |
| Key Advances Enabling Modern Deep Learning Success | 3 | 8 | 0 | 0 | Cannon inquires about the historical factors enabling modern deep learning. Zaremba explains weight initialization, stochastic gradient descent, and how convolutional neural networks exploit spatial symmetry across images and sound Fourier transforms. | |
| The ImageNet Competition and Superhuman Computer Vision | 2 | 7 | 0 | 0 | Cannon asks what triggered the deep learning excitement over the past five years. Zaremba recounts the history of ImageNet, Fei-Fei Li's dataset creation, AlexNet's breakthrough, and achieving superhuman visual classification. | |
| AI Expansion to Speech Recognition and Machine Translation | 2 | 8 | 0 | 0 | Cannon asks how image recognition led to broader AI breakthroughs. Zaremba explains sequence-to-sequence models and recurrent neural networks sharing weights over time to handle variable-length inputs in machine translation. | |
| Categorizing AI: Narrow AI, General AI, and Superintelligence | 2 | 6 | 0 | 0 | Cannon asks to delineate narrow AI, general AI, and superintelligence. Zaremba concisely clarifies that contemporary AI is strictly narrow and tool-specific, contrasting it with human general intelligence. | |
| Current Capabilities, Business Applications, and AI Hype | 2 | 7 | 1 | 0 | Cannon asks Zaremba's assessment of industry AI hype. Zaremba offers a nuanced perspective that the field is simultaneously underhyped in supervised learning business use cases and overhyped in unsolved robotics manipulation problems. | |
| Recommendations and Exercises for Learning AI | 2 | 5 | 0 | 0 | Cannon asks for practical advice on learning AI and then inquires about the impact of automation on blue-collar work. Zaremba advocates for universal basic income and discusses the psychological burden of identity tied to work. | |
| Personal Inspirations: Recommended Books and Sci-Fi Films | 2 | 3 | 0 | 0 | Cannon wraps up the interview by asking about inspiring sci-fi media and books. Zaremba recommends 'Homo Deus' and films like 'Her' and 'Ex Machina'. |
Statements from this episode (18)
Zaremba outlines OpenAI's mission to develop safe, beneficial AGI
“So OpenAI focuses on building AI for the good of humanity. We are a group of researchers and engineers collaborating together who essentially try to figure out what are the missing pieces of artificial, of general artificial intelligence and how to build it in…”
Zaremba: OpenAI has secured $1 billion in total investment
“In total we gather an investment of one billion dollar in the group.”
Zaremba: OpenAI focuses robotics research on object manipulation
“So in terms of robotics, we are working on manipulation. We think that manipulation is the complete, it's the one of the parts of robotics, which is the most unresolved.”
Zaremba: Robots cannot grasp arbitrary objects without custom hand-coded programming
“So it turns out that when it comes to arbitrary objects current robots are unable to just grasp an arbitrary object. For any object, it's possible to hand-code a single solution, so, say, as long as, let's say, in factory, if you have same object like I don't …”
Zaremba: RL models need three years of real-time play to learn games
“Well, for instance, in terms of real-time execution, it takes something around three, three, three years of play to learn to play simple games . I mean it, it can be hugely parallelized, therefore it takes a few days to train it on current computers”
Zaremba: Reinforcement learning struggles in reality due to reward and reset assumptions
“The assumption underlying reinforcement learning Is that then there is some environment, and environment, you are an agent, and you are acting in environment by executing actions and getting rewards from the environment. And the rewards might be taught as, let…”
Zaremba: Google avoided ML in Search early on over interpretability issues
“Over the time Google search started to use machine learning because it was, it helps to improve results but simultaneously, they wanted to avoid it for some time as it's more difficult to interpret the results, and it's more difficult to actually understand wh…”
Zaremba: ReLU activation empirically works way better than sigmoid in neural networks
“It turns out that the one which is even simpler
Empirically works way better, which is called ReLU, Rectify Linear Unit, and this one is ridiculously simple.”
Zaremba: Neural networks unified text, image, and sound under one methodology
“The cool thing about neural networks is, it used to be the case that people specialized in processing text, images, sound, and these days, this is the same group of people. They, we are using the same method.”
Zaremba: ImageNet is the essential dataset that enabled deep learning
“That's the essential data set that made deep learning happen.”
Zaremba: AlexNet achieved a 15% ImageNet error rate versus competitors' 25%
“A team from University of Toronto, led by Geoffrey Hinton, and that's, like, the team was Alex Krzyzewski and Ilya Suskaver. They actually got To something like 15%. So let's say all other teams, they were like at 25%, the difference was one percent. Yeah. And…”
Zaremba: ImageNet error dropped to 3%, achieving superhuman vision performance
“Within several years, people got down, I believe, to three percent error, and that's essentially superhuman performance.”
Zaremba: Neural network solutions outperform all alternatives despite high deployment costs
“And pretty much that's the reason why things are not largely deployed in production systems out there, but neural network based solutions are actually Outperforming anything what is out there.”
Zaremba: Humans are proof of concept that AGI is doable
“So how we know that it's even doable? Because we, ah, there is an example of a creature that has such a property.”
Zaremba: Supervised learning is the only ML paradigm ready for business
“So far the supervised learning paradigm is the only one that works so remarkably, remarkably well that it's ready to be applied in business applications. All other are not really there.”
Zaremba: Most business problems can be solved via supervised learning
“Majority of business problems can be framed as supervised learning, and therefore they can be solved with current techniques, as long as we have sufficient number of input examples and what we want to predict”