Behnam Neyshabur
Behnam Neyshabur is Mirendil’s cofounder and CEO, working to accelerate AI research and development.
Contents15 sections
Key facts
2009–2017Education
He earned computer engineering bachelor’s and master’s degrees at Sharif University of Technology in 2009 and 2011.[1]
He completed his doctorate at the Toyota Technological Institute at Chicago in 2017, advised by Nathan Srebro, with the dissertation Implicit Regularization in Deep Learning.[2]
2015Capacity and Optimization
With Ryota Tomioka and Nathan Srebro, he studied norm-constrained feed-forward networks, examining capacity, convexity, and their characterization.[3]
That year, he, Ruslan Salakhutdinov, and Nathan Srebro proposed Path-SGD, an optimization method using path regularization and a geometry invariant to weight rescalings that preserve network outputs.[4]
2017Implicit Regularization
His dissertation investigated how optimization algorithms implicitly constrain model complexity and how this relates to generalization, comparing whether complexity measures explain observed deep-learning behavior.[5]
2017–2019Postdoctoral Research
From September 2017 to July 2018, he was a member of the Institute for Advanced Study’s School of Mathematics, researching optimization and generalization in deep learning.[6]
He subsequently conducted postdoctoral research in Yann LeCun’s group at New York University and joined the Simons Institute’s Foundations of Deep Learning program as a research fellow in summer 2019.[7]
2019Joining Google
He joined Google in 2019, later co-leading its Blueshift reasoning team.[8]
2019Studying Generalization Measures
A study he co-authored compared over 40 complexity measures across more than 10,000 convolutional networks trained with varied hyperparameters, examining their relationship with generalization.[9]
2020SAM
With Pierre Foret, Ariel Kleiner, and Hossein Mobahi, he introduced SAM, which considers both loss and sharpness to seek parameters whose neighborhoods have consistently low loss.[10]
The paper reports improved generalization on several image tasks and tests robustness to noisy labels.[10]
June 2022Gated State Spaces
He co-authored Gated State Space (GSS), studying long-range autoregressive modeling on books, code, and mathematics articles; adding local self-attention improved the reported results.[11]
June 2022Minerva
He co-authored Minerva, which adds technical training to general language pretraining for mathematical and scientific reasoning.[12]
Google notes that Minerva can make calculation or reasoning errors; even a correct answer may follow faulty steps that cannot be formally verified automatically.[13]
November 2022REPAIR
He co-authored REPAIR, investigating poor performance when interpolating neuron-aligned networks. It rescales preactivations to mitigate variance collapse in the interpolated network.[14]
2024Joining Anthropic
In late 2024 he joined 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 co-led Discovery, developing an AI scientist.[8]
2026Mirendil
Mirendil formally announced itself on June 24, 2026, reporting a founding team of 20 researchers and engineers.[15]
The company said Andreessen Horowitz and Kleiner Perkins led a $200 million seed round, followed by investment from NVIDIA and others.[15]
The company aims to broaden access to frontier AI R&D, training models and redesigning research workflows to accelerate science and technology.[16]
August 2026Compute Infrastructure
Google Cloud identified him as Mirendil’s cofounder and CEO and said the company was adopting its TPU and NVIDIA infrastructure for model training.[17]
Related Organizations, Websites, and Public Accounts
Behnam Neyshabur Personal Website: Official website: https://www.neyshabur.net/ (opens in a new window)[8]
Toyota Technological Institute at Chicago: Official website: https://www.ttic.edu/ (opens in a new window)[2]
Institute for Advanced Study: Official website: https://www.ias.edu/ (opens in a new window)[6]
New York University: Official website: https://www.nyu.edu/ (opens in a new window)[7]
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: Official website: https://deepmind.google/ (opens in a new window)[8]
Anthropic: Official website: https://www.anthropic.com/ (opens in a new window)[2]
Mirendil: Official website: https://mirendil.com/ (opens in a new window)[16]
X Public Account: Official account: https://x.com/bneyshabur (opens in a new window)[8]
Sources
- Curriculum Vitae (opens in a new window)
- TTIC Student Alumni (opens in a new window)
- Norm-Based Capacity Control in Neural Networks (opens in a new window)
- Path-SGD: Path-Normalized Optimization in Deep Neural Networks (opens in a new window)
- Implicit Regularization in Deep Learning (opens in a new window)
- Behnam Neyshabur (opens in a new window)
- Behnam Neyshabur (opens in a new window)
- Behnam Neyshabur (opens in a new window)
- Fantastic Generalization Measures and Where to Find Them (opens in a new window)
- Sharpness-Aware Minimization for Efficiently Improving Generalization (opens in a new window)
- Long Range Language Modeling via Gated State Spaces (opens in a new window)
- Solving Quantitative Reasoning Problems with Language Models (opens in a new window)
- Minerva: Solving Quantitative Reasoning Problems with Language Models (opens in a new window)
- REPAIR: REnormalizing Permuted Activations for Interpolation Repair (opens in a new window)
- Announcing Mirendil (opens in a new window)
- Mirendil (opens in a new window)
- Mirendil taps AI Hypercomputer TPUs and GPUs for pre- and post-training applications (opens in a new window)