Labs and companies

One of the main employers for Applied Scientist roles. It publishes its hiring process and Leadership Principles as pages, which makes them primary source material for behavioral interview prep.

A frontier lab, and the maker of Claude. For the career layer, what matters is that it wrote its policy on candidates' use of AI as a public page. The hiring page itself is the best prep material.

In the history layer it stands for the branch of "alignment as a core research direction". It split off from OpenAI in 2021. Constitutional AI: Harmlessness from AI Feedback replaces most human preference labeling with a set of written principles, and has the model critique and rewrite its own answers. Beyond the Claude series, it released the protocol in MCP：给每个系统一个统一的插口, which made "models connecting to tools" an industry standard. In Opinions and predictions, Dario Amodei's long essay is the most often cited judgment for the frontier layer.

In the 1990s, LeCun's convolutional networks and Vapnik's SVM competed in the same building. The two routes of the second wave (neural networks vs. statistical learning) met here.

Newell and Simon's Logic Theorist (1956) was the first AI program. XCON, speech recognition (Sphinx) and self-driving (NavLab) all came out of this line of work.

An AI editor company (legal entity: Anysphere). It is one of the fastest-hiring AI-native startups. Its hiring page is a good sample of "work-sample style" role requirements.

The main funder of AI in the 1960s and 70s. Its 1974 cuts were the US side of the first AI winter. In the 1980s its Strategic Computing Initiative gave expert systems another push. When you read AI history, its name stands for the funding curve.

A lab in Hangzhou funded by the quant fund High-Flyer. Its R1, released in January 2025 with open weights and a publicly stated low training cost, showed that reasoning models are not one company's secret.

A nonprofit research group that tracks AI trends. It publishes open datasets on training compute, model size, chip cost, and when we run out of data. When I talk about scaling and compute economics, I check the numbers here first.

Google Brain (2011) turned deep learning into infrastructure. The Transformer, BERT and TPU all came from there. In 2023 it merged with DeepMind to form Google DeepMind, which entered the race with Gemini.

Google's AI lab (DeepMind was founded in 2010 and merged with Google Brain in 2023). It has public official hiring pages for research roles and research engineering roles.

In terms of history, it is the standard-bearer of the second half of the second wave: Playing Atari with Deep Reinforcement Learning taught a network to play Atari from pixels. Mastering the game of Go with deep neural networks and tree search used self-play plus search to beat top human players. Highly accurate protein structure prediction with AlphaFold applied the same approach to protein structures and won a Nobel Prize. Training Compute-Optimal Large Language Models corrected the recipe for scaling: for the same compute, data has to grow together with parameters. After the 2023 merger with Google Brain, the Gemini series put it back on the front line of LLM competition.

The transformers library and the model hub standardized how pretrained models are distributed. It is the "GitHub" of the open-weights ecosystem.

Samuel's checkers program (1959), Deep Blue (1997), Watson (2011), and statistical machine translation (1990s) are all here. In each of the three waves, it stood at the peak of the mainstream method of its day.

A small lab in Lugano, Switzerland, and the birthplace of the LSTM. Around 2011 it won several vision competitions with GPU convolutional networks, before AlexNet.

One of the big-tech employers with the most AI roles. Besides FAIR, it has a superintelligence lab. Its careers page is a good sample of how a big company titles its AI roles.

FAIR, which LeCun founded in 2013. PyTorch and LLaMA both came out of it. It is the biggest driver of the open-weights route.

ResNet came out of Microsoft Research Asia. Since 2019 Microsoft has invested in OpenAI and been its exclusive compute provider, and Copilot made code generation a mass-market product for the first time.

Bengio's lab. In 2014 it produced both the attention mechanism and GANs in the same year. One of the three places (Toronto, Montreal, Edmonton) where Canada kept deep learning alive in academia.

Founded in Paris in 2023. Mistral 7B and Mixtral showed that a small model with good data can beat larger models. The main player in Europe's open-weights approach.

The lab Minsky and McCarthy founded in 1959. It was the stronghold of the symbolic school: ELIZA, SHRDLU and the Lisp machines all came out of it. It merged into CSAIL in 2003.

CUDA (2007) turned the GPU into a general-purpose compute device. After AlexNet, it became the physical foundation of deep learning. The compute economics of the third wave revolve around its supply cycle.

A frontier lab, the author of the GPT series. Research engineer roles are its main hiring line, and its official interview guide is public.

In the history layer, it is the lead of the third wave. Improving Language Understanding by Generative Pre-Training through Language Models are Few-Shot Learners showed that the road of "the same objective, scaled up ten times" could keep going. Scaling Laws for Neural Language Models wrote that down as a formula. Training language models to follow instructions with human feedback taught the continuation machine to follow instructions. ChatGPT then put it in everyone's hands. The later GPT-4 Technical Report no longer discloses the architecture, which marks research moving from papers to products. For the nodes, see The third wave: scaling and LLM and Pretraining and scaling.

Founded by McCarthy in 1963. Expert systems (DENDRAL, MYCIN) came out of it. Forty years later, ImageNet and DPO came from here too. It shows up in all the three waves.

The home of Hinton's lab, where deep learning was kept alive through the winter. Deep belief networks (2006) and AlexNet (2012) were both built here.