People

He matters for "can machines think": he turned that philosophical question into a game you can actually run. In 1950, when computers could only calculate ballistic tables, he predicted machines that learn.

First author of GPT-1, GPT-2, CLIP and Whisper. OpenAI's "generative pre-training" line took shape in his experiments.

First author of AlexNet. He wrote the code that trained a convolutional network on two GPU cards himself. The breakout moment of deep learning carries his name.

What matters about him at the history layer is not his research but his teaching. From the 1989 reproduction of LeCun's network to "one hour on LLMs", he explained the technical thread of the three waves to more people than anyone else.

The teacher who has explained machine learning to the most people: his ML course on Coursera was the entry point for a generation of engineers, and he runs The Batch, which squeezes the frontier into one page every week. On this map he matters for the "teaching" edge, not for research.

Co-author of the Chinchilla paper. He founded Mistral in 2023. Mistral 7B beat larger models with fewer parameters, and it is the main example of the European open-weights approach.

The 1959 checkers program improved its evaluation function through self-play and ended up playing better than its author. He is also the one who coined the term "machine learning".

First author of the Transformer paper. Took all the recurrence out of sequence models and kept only attention. It became the backbone of every large model since.

CEO of Hugging Face. The transformers library and the model hub standardized how pretrained models are distributed. The open-weights ecosystem has a "GitHub" thanks in part to him.

Former OpenAI governance researcher and lead author of the AI 2027 scenario. He wrote "AGI within a few years" as a dated timeline. So of all the opinions, this one is the easiest to falsify and the most worth reviewing when the dates come due.

While VP of Research at OpenAI, he led GPT-2, GPT-3 and scaling laws. In 2021 he left with a team and founded Anthropic. He is the author of the "scaling + safety" narrative.

The reinforcement learning lead behind DQN, AlphaGo and AlphaZero. The "self-play + search" line, from TD-Gammon to AlphaZero, came together in his hands.

Founder of DeepMind. AlphaGo changed the image of deep learning from "recognizing pictures" to "making decisions". AlphaFold took it into scientific discovery, and that work won the Nobel Prize.

The founder of Cyc. Starting in 1984, he spent almost forty years hand-coding common sense. It was the longest and most thorough experiment of the symbolic school, and how it ended is the best case study for why knowledge can't be written by hand.

Host of the long-form interview podcast: he questions one person for three hours and gets people inside the labs to say the judgments they don't usually put in blog posts. Most of what we hear from key people at the frontier comes through him.

Founder of SemiAnalysis. He works out the cost of AI from chips, power and data centers. Most of the numbers in the compute economics section trace back to him.

The father of expert systems. DENDRAL showed that domain knowledge matters more than general reasoning. He summarized this as knowledge engineering, and he personally drove the expert-system industry of the 1980s.

The person who started ImageNet. She bet on data scale at a time when algorithms got more attention than data. The 2012 AlexNet breakthrough happened on her dataset.

The inventor of the perceptron. He was the first to have a machine learn weights from examples and build it in hardware. His approach was suppressed for a generation and only proved right after 1986.

The person who held out longest through the neural network winter: he popularized backpropagation in 1986, reignited deep learning in 2006, and built AlexNet with his students in 2012. After leaving Google in 2023, he turned to warning about the risks.

Inventor of the GAN (said to have come up with it in one night), the starting point of the deep generative model line; also the first author of the textbook Deep Learning.

From AlexNet to seq2seq to GPT, he was there at the turning point of each of the three waves. Former chief scientist at OpenAI and the earliest evangelist of the scaling faith. He left in 2024 and founded SSI.

Co-founder of Anthropic. Before that, he worked on policy at OpenAI. He has run the Import AI weekly newsletter for nearly ten years. It is the most reliable feed I know for "which papers are worth noticing this week."

A physicist by training and the first author of the scaling laws paper. He turned "how much loss does more compute buy" into a power law you can extrapolate. Co-founder of Anthropic.

He founded Google Brain and turned deep learning from a research project into Google's infrastructure (DistBelief, TensorFlow, TPU). He is now Google's Chief Scientist.

Founder of NVIDIA. CUDA turned the GPU into a general-purpose computing device. Since AlexNet, the GPU has been the physical foundation of deep learning, and this whole wave of scaling pays its bill to him.

He gave the field its name (1956 Dartmouth proposal), invented Lisp, and held to the line that intelligence is formal reasoning. In the 1973 Lighthill debate, he was the one defending AI.

The author of ELIZA. He was surprised that users got emotionally attached to a pattern-matching program, and he spent the rest of his life criticizing AI's overblown promises. The difference between "seems to understand" and "really understands" starts with him.

Bayesian networks made reasoning under uncertainty computable and brought probability back into an AI field ruled by logic. He represents the statistical side of the second wave, and he later turned to causal inference.

Co-author of LSTM and head of IDSIA. Around 2011, his lab won several competitions with GPU convolutional networks. He has long argued that many "new" ideas were already done by his group.

First author of ResNet. Residual connections made networks with a hundred layers trainable. Since then they have been standard in every deep network, including the Transformer.

The founder of DeepSeek. He used a quant fund's compute and a small team to build open-weight reasoning models, and showed that frontier capability is not the property of a few American labs.

The author of Lil'Log: every few months she writes a long post that walks one area (agent, RLHF, diffusion models) from start to finish. It is the first survey I open when I need to look up when needed.

One of the founders of the MIT AI Lab. His book Perceptrons left neural networks dormant for more than ten years, yet he started out in neural networks himself (SNARC). The standard-bearer of the symbolic school.

Author of Interconnects and a post-training researcher at Ai2. He also wrote the public book on RLHF. On post-training and open models, his judgment is the most first-hand.

Co-author of the Transformer. Also proposed sparsely-gated mixture of experts (MoE) and multi-query attention. Left Google to found Character.AI, then was bought back in 2024.

The original authors of RLHF: in 2017 they trained a reward model on pairwise human comparisons, and in 2022 the same method produced InstructGPT. After that, they focused on alignment research.

Author of the standard reinforcement learning textbook and originator of temporal-difference learning. His 2019 essay The Bitter Lesson is the most cited single page of the scaling era. 2024 Turing Award.

CEO of OpenAI. He turned a nonprofit research lab into a product company built around ChatGPT. He is also a main driver of the compute arms race.

Author of Build a Large Language Model (From Scratch): a book and companion repo that write a GPT line by line, the reference answer for the L2 reproduction project. Ahead of AI tracks new papers.

His 1991 master's thesis already pointed out the vanishing gradient problem, and LSTM was his answer. It was the default backbone of sequence modeling before the Transformer.

Co-creator of Django. For the past few years he has logged, every day, how LLM tools work in practice. The term prompt injection and the "lethal trifecta" framing both come from his blog. He is the most diligent front-line observer of the application engineering layer.

He runs Latent Space. He pushed the "AI Engineer" job title and its annual conference into use. Half of the vocabulary in the role family section comes from here.

SHRDLU is the peak demo of symbolic natural language understanding. He later admitted himself that this path could not get out of the blocks world, and he turned to human-computer interaction.

Author of word2vec. He cut the language model down to a single prediction task, which let him train word vectors on a billion words. Pretrained representations became routine starting with his work.

Author of statistical learning theory (VC dimension) and the support vector machine. In the 2000s, the classifiers with theoretical guarantees were his. Deep learning only overtook them in 2012.

With Walter Pitts, he wrote the neuron as a logic unit. This paper is the shared starting point of both the neural network and symbolic computation lines.

Founder of convolutional networks. In 1989 he used backpropagation to recognize zip codes. He later founded Meta's FAIR and argued for open research, and he pushed the open route for LLaMA.

The 2003 neural language model is a common source of both word vectors and the LLM. In 2014, his lab produced the attention mechanism and GANs. He stayed in academia and founded Mila.