Anna Goldie

Goldie earned bachelor’s degrees in computer science and linguistics at MIT, followed by a master’s degree.⁠[1]

Her 2011 master’s thesis, CHATTER, studied dialogue systems for language learning.⁠[2]

In 2017, she co-authored a study of neural machine translation architectures with Denny Britz, Minh-Thang Luong, and Quoc LeQuoc LeQuoc Le (Vietnamese name Le Viet Quoc) is an artificial intelligence researcher and early member of Google Brain who led the 2012 experiment in which a neural network running on 16,000 processor cores taught itself to recognize cats from YouTube videos, and who helped develop deep learning methods including sequence-to-sequence learning (seq2seq), paragraph vectors, and neural architecture search (AutoML). In August 2026, he left Google with Jeff Dean, Sanjay Ghemawat, and Oriol Vinyals to co-found the AI company Discovery Loop.Open the full entry. The team ran hundreds of English–German translation experiments totaling over 250,000 GPU hours and released an open-source framework.⁠[3]

At Google Brain, she co-founded and led the Machine Learning for Systems team, applying reinforcement learning to computing systems and chip design.⁠[1]

The chip-layout research she co-led was published in Nature in 2021, with Goldie and Azalia Mirhoseini as co-first authors.⁠[2]

In a Nature interview that year, she explained that the system places components sequentially and receives a reward based on layout quality. She highlighted constraints involving area, power, heat, and application requirements, and said some layouts were used in Google TPUs manufactured in January 2021.⁠[4]

In 2021, she was selected for MIT Technology Review’s global Innovators Under 35 list. Her profile highlighted reinforcement learning for faster chip layout and the prospect of AI and hardware design advancing one another.⁠[5]

She recalled that unfamiliar machine-generated layouts initially worried chip designers about manufacturing viability; the designs ultimately became physical chips.⁠[5]

She and Mirhoseini subsequently moved from Google Brain to 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, returned to Google, and then founded their company together.⁠[6]

She worked on Claude at Anthropic and later on Gemini as a senior staff research scientist at Google DeepMind.⁠[2]

Nature flagged concerns about the paper’s performance claims in 2023. In September 2024, it announced that post-publication review had resolved the issues, closed its investigation, and published an addendum.⁠[7]

In September 2024, Goldie and Mirhoseini reported AlphaChip’s use in three successive generations of Google TPUs and in Axion data-center processors. MediaTek had also extended the method for its own chip designs.⁠[8]

The method pretrains on existing chip blocks before tackling new layouts. Its graph neural network represents relationships between components to transfer experience across chips.⁠[8]

In 2025, she earned a PhD in computer science at Stanford, supervised by Christopher D. Manning. Her dissertation examined sequential decision-making in chip design and language modeling, addressing problem formulation, representation learning, reward attribution, and multi-step trajectory data.⁠[9]

In April 2025, she and collaborators released the SWiRL preprint. The method iteratively generates reasoning and tool-use data, decomposes multi-step trajectories into sub-trajectories, and applies data filtering and reinforcement learning. The paper reports experiments on question answering and mathematics.⁠[10]

In 2025, she and Mirhoseini founded Ricursive Intelligence, which publicly launched in December. Goldie became CEO and Mirhoseini CTO.⁠[11]

After a $35 million seed round led by Sequoia, the company announced a $300 million Series A led by Lightspeed in January 2026 at a $4 billion valuation. Its product direction is helping chip companies design chips, rather than manufacturing chips itself.⁠[6]

The company pursues a cycle in which AI designs hardware that supports further AI development. Shortening the full chip-development cycle and achieving recursive self-improvement remain goals, rather than established completed outcomes.⁠[12]

On August 27, 2026, Goldie and Mirhoseini were jointly featured in TIME100 AI, which highlighted their effort to co-optimize chip design and AI models.⁠[11]

Personal website and publications: Official website: https://www.annagoldie.com/ (opens in a new window)⁠[2]

Ricursive Intelligence: Official website: https://www.ricursive.com/ (opens in a new window)⁠[12]

Google DeepMind: Official website: https://deepmind.google/ (opens in a new window)⁠[8]