Azalia Mirhoseini

Mirhoseini was born in Iran.⁠[1]

She earned a BSc in electrical engineering at Sharif University of Technology and a PhD in electrical and computer engineering at Rice University.⁠[2]

She conducted research at Google Brain, AnthropicAnthropicAnthropic is an artificial intelligence research and product company founded in 2021 that develops the Claude models, assistant, and tools such as Claude Code, and operates as a public benefit corporation. Dario Amodei and Daniela Amodei are its co-founders, and its research covers model safety, interpretability, and steerability.Open the full entry, and Google DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry.⁠[2]

In 2017, she co-authored the sparsely gated mixture-of-experts layer. A gating network selects a small set of experts for each input, allowing model capacity to grow without a proportional rise in computation.⁠[3]

The team tested expert layers with up to 137 billion parameters in language modeling and machine translation.⁠[3]

In 2019, MIT Technology Review named her an Innovator Under 35, highlighting her research on reinforcement learning for chip layouts.⁠[1]

The team introduced its reinforcement-learning chip-layout method in 2020, with the related Nature article appearing in June 2021. A September 2024 addendum named it AlphaChip; Mirhoseini and Anna GoldieAnna GoldieAnna Goldie is an AI researcher and co-founder and CEO of Ricursive Intelligence. She previously worked at Anthropic and Google DeepMind and co-led AlphaChip.Open the full entry were equal first authors of the addendum.⁠[4]

In a subsequent response, the authors reported deployment in three TPU generations and emphasized learning from existing chip blocks before transferring to new layouts.⁠[5]

She co-authored Constitutional AI, which uses human-chosen principles for self-critique and revision, followed by training with AI preference feedback.⁠[6]

She created and taught CS229S, Systems for Machine Learning, with autumn offerings in 2023 and 2024. The course addresses efficiency and scalability in training, fine-tuning, and inference.⁠[7]

Her Scaling Intelligence Lab studies scalable intelligence, self-improvement, verification, and the mutual optimization of AI and computing systems.⁠[8]

In autumn 2025, she co-taught CS329A, Self-Improving AI Agents, with Aakanksha Chowdhery, covering tool use, memory, reasoning, planning, and evaluation.⁠[9]

The Okawa Foundation listed her as a 2025 overseas research-grant recipient for “Towards Self-Improving AI Systems.”⁠[10]

In 2023, Nature investigated performance claims in the study. The withdrawn item was a separate News & Views commentary, not the research paper.⁠[11]

The authors’ 2024 response reported that Nature had closed its investigation after review. They also argued that the critical evaluation used different pretraining, compute, and convergence conditions.⁠[5]

Cheng and colleagues’ updated 2026 assessment used new public 7-nanometer benchmarks and updated Circuit Training, identifying scalability and pretraining methodology as questions still needing confirmation.⁠[12]

In Large Language Monkeys, she studied repeated sampling: more attempts can increase the fraction of problems with at least one correct answer. DeepSeek-V2-Coder-Instruct coverage on SWE-bench Lite rose from 15.9% with one sample to 56% with 250.⁠[13]

Without an automatic verifier, selecting the correct answer remains a limitation; coverage is not final selection accuracy.⁠[13]

She co-authored KernelBench, which evaluates language-model GPU kernels on 250 PyTorch workloads. Its fast_p metric requires both functional correctness and a specified speedup.⁠[14]

The study found that execution and profiling feedback improved iterative results, while models still struggled to generate efficient kernels directly.⁠[14]

At Simons Institute, she discussed parallel sampling, serial revision, and inference-step combinations, stressing limits from verification quality and computational cost.⁠[15]

She described continual learning as a research direction, without establishing unlimited improvement: fine-tuning could reduce answer diversity, and experiments needed more compute.⁠[15]

In 2025, she and Anna Goldie founded Ricursive Intelligence.⁠[16]

In January 2026, Lightspeed announced that it was leading the Series A, emphasizing co-development of AI models and specialized chips.⁠[17]

As of October 2026, the company website lists $335 million in backing, with investors including Sequoia.⁠[18]

Her co-authored Intelligence per Watt was submitted in 2025 and updated in 2026. It combines accuracy and power consumption, testing over 20 local models, eight accelerators, and around one million single-turn queries.⁠[19]

The latest version reports accurate local responses to 88.7% of tested queries, varying by domain and not generalizable to every task.⁠[19]

She contributed to OpenJarvis, a local-first personal AI framework organized around models, inference engines, agents, tools and memory, and learning, with energy, latency, and cost included in evaluation.⁠[20]

The project studies improvement from local execution traces and permits cloud calls when necessary; local-first does not mean every configuration is entirely offline.⁠[20]

The LLM-as-a-Verifier project, announced in April 2026, lists Mirhoseini as an equal adviser. It improves assessment of agent trajectories through scoring granularity, repeated verification, and decomposition of evaluation criteria.⁠[21]

Azalia Mirhoseini Personal Website: Official website: https://www.azaliamirhoseini.com/ (opens in a new window)⁠[22]

Stanford University: Official website: https://www.stanford.edu/ (opens in a new window)⁠[23]

Rice University: Official website: https://www.rice.edu/ (opens in a new window)⁠[24]

Ricursive Intelligence: Official website: https://www.ricursive.com/ (opens in a new window)⁠[18]

Scaling Intelligence Lab: Official website: https://scalingintelligence.stanford.edu/ (opens in a new window)⁠[8]

Google DeepMind: Official website: https://deepmind.google/ (opens in a new window)⁠[25]

X Public Account: Official account: https://x.com/Azaliamirh (opens in a new window)⁠[26]

LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry Public Account: Official account: https://www.linkedin.com/in/azalia-mirhoseini-a5308522/ (opens in a new window)⁠[27]