Sutskever, Ilya

co-founder and chief scientist of OpenAI; co-author of the AlexNet paper; one of the most cited computer scientists in history

Ilya Sutskever: Deep Learning

Lex Fridman Podcast · May 8, 2020

• Ilya Sutskever
• OpenAI
• deep learning
• GPT-2
• transformers
• AGI
• backpropagation
• self-play
• double descent
• AI alignment

Core ideas

Sutskever traces his conviction that deep learning would work back to around 2010–2011, when James Martens's Hessian-free optimizer successfully trained a 10-layer neural network end-to-end without pretraining — a result he connected to a rough analogy with the brain (10 layers being loosely comparable to how many times neurons could "fire" within the ~100 milliseconds humans take to recognize an object) and to an early, later-refined theoretical intuition that having more data than parameters should prevent overfitting. He credits Alex Krizhevsky's extremely fast CUDA kernels for convolutional networks as the missing piece that let this conviction be tested at ImageNet scale, and frames the entire deep learning breakthrough not as new ideas but as old ideas (large neural networks trained end-to-end via backpropagation) combined with sufficient data, sufficient compute, and — critically — the "conviction" to combine them, given that neural networks had been consistently unconvincing on real benchmarks for decades and scientific communities reasonably distrusted them until hard, undeniable benchmark results (like ImageNet) arrived to settle the debate empirically rather than rhetorically.

On the brain-versus-artificial-neural-network comparison, Sutskever is skeptical that spiking (a major biological mechanism absent from most ANNs) is essential, arguing spiking neural network researchers effectively have to simulate non-spiking computation using spikes to get results — while flagging spike-timing-dependent plasticity as a biological learning rule worth studying in simulation. He identifies the cost function as deep learning's foundational, almost invisible-in-retrospect idea, contrasting it with GANs (where behavior is better understood via game-theoretic equilibrium than any single cost function, similar to how biological evolution or economies aren't well described by one) while still expressing strong confidence in cost-function-based approaches generally. He argues machine learning has surprising "unity" — the same core principles (gradient-based optimization) now unify what used to be fragmented sub-fields, with computer vision and NLP already converging toward shared architectures (transformers displacing bespoke NLP pipelines) and reinforcement learning, while requiring distinct techniques for exploration and variance reduction because of its fundamentally non-stationary setting (actions change future observations), likely to eventually integrate more tightly with supervised learning.

Sutskever discusses his paper on "deep double descent" (bigger models and more data can temporarily hurt performance): plotting test performance against model size while holding data fixed, performance improves, then gets worse right at the point where training error reaches exactly zero, then improves again as the model grows past that point — an effect that also appears in linear models and whose intuition he ties to how sensitive a model is to "spurious," random details in the training set: when model capacity roughly equals data capacity, there's a near one-to-one correspondence between the two, making small dataset changes shift the learned model substantially, whereas heavily over- or under-parameterized regimes are more robust to this noise; the effect can also be substantially mitigated with early stopping. Responding to Geoffrey Hinton's suggestion that backpropagation might need to be discarded, Sutskever clarifies Hinton's actual claim was narrower (worth investigating brain-inspired alternatives if backprop truly isn't found in biological learning) and states his own strong belief that backpropagation solves a sufficiently fundamental problem (finding a neural circuit under constraints) that it's unlikely to be displaced soon.

On reasoning, Sutskever argues AlphaZero already constitutes an "existence proof" that a single neural network (without search) can perform something functionally like reasoning at superhuman level in a constrained domain (Go), and that neural networks trained on tasks that don't require reasoning simply won't reason, since networks "solve exactly the problem you pose... in the easiest way possible" — meaning the path to more general reasoning likely requires the right training tasks (like theorem-proving or competitive programming) rather than fundamentally new architectures. He revisits and revises his earlier framing of neural networks as "the search for small circuits" (analogous to but computationally tractable relative to Solomonoff-style search for the shortest program generating the data, which is uncomputable): he now prefers describing it as finding large circuits whose weights nonetheless encode a small amount of information, which he argues explains generalization — the amount of information "transmitted" from data into parameters during training ends up surprisingly compact. He emphasizes training remains the fundamental, immovable constraint shaping all deep learning architecture choices ("we've got to stay trainable... this is an invariant we cannot violate"), which is why practical progress in areas like program synthesis will likely come from training large neural networks to produce programs rather than from directly searching program space.

On language models, Sutskever explains GPT-2 (a 1.5-billion-parameter transformer trained on ~40 billion tokens from Reddit-linked web pages) and the significance of the transformer architecture, arguing its success comes from the combination of several factors — heavy use of attention, GPU-efficient design, and non-recurrence (making it shallower and easier to optimize) — rather than attention alone. He connects language model scale to emergent semantic understanding, citing earlier work on a "sentiment neuron" where scaling an LSTM from 500 to 4,000 cells caused a single neuron to spontaneously represent review sentiment (a semantic rather than syntactic property) — his explanation being that once a model exhausts the syntactic patterns available at smaller scale, it's forced to start capturing semantics to keep improving prediction. He was genuinely surprised by how well GPT-2 worked despite theoretically expecting language models to improve with scale, comparing the leap to a sudden jump in GAN image quality, and expresses interest in future work enabling models to actively/selectively choose what data to learn from (analogous to human selectivity) rather than indiscriminately absorbing everything, though he notes such capabilities need a genuine problem requiring them to drive real research progress rather than artificial benchmark demonstrations. He also discusses OpenAI's staged release of GPT-2 as reflecting AI's transition from a "childhood" to a "maturity" phase where potential negative impacts (like cheapened disinformation) warrant caution even under uncertainty, expressing hope that AI developers can gradually build enough mutual trust to discuss such risks across organizational and competitive boundaries.

On AGI, Sutskever expects deep learning plus "maybe another small idea" to suffice, naming self-play as one likely component precisely because of its demonstrated capacity to produce genuinely surprising, creative, useful solutions (citing OpenAI's Dota bot, multi-agent hide-and-seek, and AlphaZero), and describes viewing simulation not as a limitation but a tool with particular strengths — citing OpenAI's Rubik's-cube-solving robot hand (trained entirely in simulation, yet its resulting policy proved robust to physical perturbations like a glove or stuffed giraffe never seen in training) as evidence that simulation-to-reality transfer is a solvable, already-demonstrated problem whose feasibility will grow as transfer learning generally improves. He considers embodiment useful but not necessary for intelligence (citing Helen Keller as evidence that missing sensory/interactive modalities can be compensated for), tentatively affirms that sufficiently brain-like artificial neural networks should in principle be capable of consciousness (leaning on an "existence proof" argument from human consciousness), and offers a compelling articulation of AI alignment: he envisions AGI systems that can be designed to want to be helpful and controlled by humans in the same intrinsically motivated way that parents want to help their children — not as an externally imposed constraint but as a genuine internalized drive — describing an ideal governance picture where humanity functions like a board of directors able to "reset" an AGI system that represents them, analogous to democratic oversight of a powerful CEO. Asked whether he personally could relinquish power over a system that powerful, he says he'd find it "trivial" and finds the scenario of concentrated personal control "terrifying," expressing hope (not certainty) that people generally behave better than expected "when it really counts." He closes on meaning and happiness, rejecting the premise that life has an externally-defined objective function (distinguishing this from evolution's implicit objective of survival/reproduction, which he says doesn't answer the meaning-of-life question) and offering a view that happiness comes substantially from perspective and how one chooses to interpret experience, rather than from specific accomplishments.

Excerpts