Ilya Sutskever

Co-authored AlexNet with Hinton and Krizhevsky — the paper that launched the deep learning era. The thesis: larger networks, more data, and more compute consistently yield better performance. This 'scaling hypothesis' became Sutskever's organizing principle and drove the GPT series.

Contributed to the foundational OpenAI scaling laws paper showing that model performance improves smoothly and predictably with compute, data, and parameters. This wasn't obvious — many researchers expected diminishing returns. The scaling laws justified the massive compute investments in GPT-3, GPT-4, and beyond.

Led the team that developed RLHF to align GPT models with human preferences. The key insight: a model trained purely on text prediction is not necessarily helpful or safe; RLHF fine-tunes it to match what humans actually want. InstructGPT and ChatGPT both depend on this work.

The November 2023 OpenAI board crisis (Sutskever voted to fire Sam Altman, then signed the letter calling for his return) revealed deep internal tensions about safety vs. speed. Sutskever subsequently founded SSI (Safe Superintelligence) with the singular focus of building safe superintelligence before building a product.

Stated publicly (2024) that the era of internet-scale pre-training may be ending as web data is exhausted, and the next frontier is synthetic data and 'test-time compute' — models that reason longer and more carefully at inference time rather than simply being trained on more data.