Glossary
Who is Ilya Sutskever
In short
Ilya Sutskever is a deep learning researcher, a co-founder of OpenAI and the founder of the company Safe Superintelligence.
People · Ilya Sutskever
MMOOZZGG.com editorsChecked: 1 min read
For a child
Remember the game where one person says a word and the other has to say it in another language? Sutskever taught a computer a similar trick. The program first listens carefully to the whole phrase and remembers its meaning, and then builds a phrase in another language by itself, word by word. He also helped teach a computer to recognize a thousand different objects in photos.
In plain words
Sutskever co-authored two well-known papers on neural networks. In 2012, together with Alex Krizhevsky and Geoffrey Hinton, he trained a large neural network to recognize objects in photos. In 2014, he and his colleagues described the sequence-to-sequence scheme: one network reads the whole phrase, and a second one writes a translation or an answer. This idea is close to how chatbots work today. Sutskever was one of the founders of OpenAI and its chief scientist, worked on GPT-3, and in June 2024 founded the company Safe Superintelligence.
For business
Two things from Sutskever’s work are used in companies. The encoder-decoder scheme from the 2014 paper influenced machine translation, document summaries and answers from a knowledge base. The GPT family of language models he worked on is the basis of assistants that write letters and answer clients. For a business, what matters is not the author’s name but a test on your own tasks: models make mistakes, translate terms in different ways and can confidently make up facts. So before launch, collect a set of real examples and count how many answers need editing. How to teach staff to set tasks for such models is covered in the material on generative AI.
For a specialist
In the paper ImageNet Classification with Deep Convolutional Neural Networks (NIPS 2012), Krizhevsky, Sutskever and Hinton trained a convolutional network with about 60 million parameters and 500,000 neurons: five convolutional layers, two fully connected ones and a softmax over 1000 classes. In September 2014, Sutskever, Vinyals and Le published Sequence to Sequence Learning with Neural Networks: a multilayer LSTM encodes the input sequence into a fixed-length vector, and a second LSTM decodes it into the target (encoder-decoder). On English-to-French translation (WMT’14), the model scored BLEU 34.8 against 33.3 for a phrase-based statistical system, and reversing the word order in the input noticeably improved quality. Sutskever co-authored the GPT-3 paper Language Models are Few-Shot Learners (May 2020), a model with 175 billion parameters. In June 2024, together with Daniel Gross and Daniel Levy, he founded Safe Superintelligence with offices in Palo Alto and Tel Aviv.
Example
An online store translates product cards for its foreign website with an encoder-decoder model. An editor checks a sample of a hundred cards and builds a glossary so that the model does not translate names in different ways.
Related terms
Questions and answers
Sources
- Gigazine, Ilya Sutskever launches Safe Superintelligence, 20.06.2024gigazine.net
- Krizhevsky, Sutskever, Hinton, ImageNet Classification with Deep Convolutional Neural Networks, NIPS 2012papers.nips.cc
- Sutskever, Vinyals, Le, Sequence to Sequence Learning with Neural Networks, 2014arxiv.org
- Brown et al., Language Models are Few-Shot Learners, 2020arxiv.org
- Safe Superintelligence Inc., official website, 2024ssi.inc
Not sure whether your company needs this?
AI solutions map