Benjamin "Ben" Mann
Last reviewed
Jun 7, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,436 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,436 words
Add missing citations, update stale details, or suggest a clearer explanation.
Benjamin Mann, known professionally as Ben Mann, is a software engineer and artificial intelligence researcher who is one of the co-founders of Anthropic, the AI safety company behind the Claude family of large language models. Before helping to start Anthropic in early 2021, Mann was a member of the technical staff at OpenAI, where he was one of the three co-lead authors of the 2020 paper that introduced GPT-3. As of 2026 he co-leads Anthropic's product incubator, Labs, alongside Instagram co-founder Mike Krieger.[1][2]
Mann studied at Columbia University, earning a Bachelor of Science in computer science from its Fu Foundation School of Engineering and Applied Science. He attended from 2007 to 2011 and graduated magna cum laude.[7][8] While a student he was active in the Theta Tau engineering fraternity and the university's culinary society. Public records do not disclose his date or place of birth.
Mann has said that the trajectory of his career changed after he read the philosopher Nick Bostrom's book on superintelligence, which convinced him that ensuring advanced AI is safe and beneficial should be the central problem of his working life. That conviction drew him away from conventional software engineering and toward AI research and safety.[2][7]
After college, Mann worked as a software engineer, including a period as a senior software engineer at Google. There he contributed to consumer products, most notably the Waze Carpool ride-sharing service, and spent time in Google's Area 120 in-house incubator.[7][8] He is also credited as a co-inventor on a US patent covering a method for identifying and ranking attributes of entities. Earlier in his career he worked at startups focused on automation and at the Machine Intelligence Research Institute (MIRI), a nonprofit dedicated to the long-term safety of advanced AI, before moving into frontier model research.[1][7]
Mann joined OpenAI as a member of the technical staff, where he worked on infrastructure, efficiency, and safety. He contributed to the GPT-2 and GPT-3 projects and was one of the three co-first authors, listed second after Tom B. Brown, of the landmark 2020 paper "Language Models are Few-Shot Learners," which introduced GPT-3, an autoregressive language model with 175 billion parameters that demonstrated strong few-shot performance across many natural-language tasks.[3][9] The paper has accumulated tens of thousands of citations and is widely regarded as a turning point in the development of large language models.
Beyond the research itself, Mann has said his work spanned both the research and product sides of GPT-3, including helping with the technical transfer of the model to Microsoft Azure infrastructure as OpenAI built out its commercial partnership with Microsoft.[2]
In late 2020 Mann left OpenAI as part of a group of researchers, led by siblings Dario Amodei and Daniela Amodei, who believed AI safety should be made the explicit top priority of a frontier lab. The group incorporated Anthropic in January 2021. Alongside the Amodeis, the co-founders included Jared Kaplan, Sam McCandlish, Tom Brown, Jack Clark, and Chris Olah. Dario Amodei became chief executive officer and Daniela Amodei president.[1][6]
At Anthropic, Mann has described himself as the company's "engineer of last resort." He says he has held roughly fifteen different roles over the company's history, including managing security and operations, building out networking and infrastructure, and starting the product team from scratch.[2] His product work helped shape Claude and the surrounding developer ecosystem as Anthropic grew from a research-focused startup into one of the most heavily funded AI companies in the world.
Mann is a co-author on several influential machine-learning papers, spanning capabilities research, safety alignment, and mechanistic interpretability. His best-known contributions are summarized below.
| Year | Publication | Mann's role |
|---|---|---|
| 2020 | Language Models are Few-Shot Learners (GPT-3) | Co-first author |
| 2020 | Scaling Laws for Autoregressive Generative Modeling | Co-author |
| 2021 | A Mathematical Framework for Transformer Circuits | Co-author |
| 2022 | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback | Co-author |
| 2022 | Constitutional AI: Harmlessness from AI Feedback | Co-author |
After GPT-3, much of Mann's published work focused on making models more helpful and less harmful. He was a co-author on Anthropic's 2022 paper on training a "helpful and harmless" assistant using reinforcement learning from human feedback (RLHF), and on the influential "Constitutional AI" paper, which introduced a method for aligning models against a written set of principles, or "constitution," using AI-generated feedback rather than relying solely on human labels.[4][5] The technique, sometimes described as reinforcement learning from AI feedback (RLAIF), became a defining part of how Anthropic trains Claude. Mann has been a vocal advocate for this kind of scalable oversight, arguing that human supervision alone will not keep pace as models become more capable.[2]
Beginning around 2024, Mann started an internal effort at Anthropic to incubate experimental products built on the frontier of Claude's capabilities. The small team operated by quickly building prototypes, testing unpolished tools with a handful of early users, and scaling the ideas that worked into reliable products.[1][2] According to Anthropic, this approach produced Claude Code, which grew from a research preview into a business generating roughly a billion dollars in revenue within six months, and the Model Context Protocol (MCP), an open standard for connecting AI models to external tools and data that the company says reached around 100 million monthly downloads.[1]
On January 13, 2026, Anthropic announced that it was formally expanding this incubator, named Labs, and that Mike Krieger, the Instagram co-founder who had served as Anthropic's chief product officer, would step into Labs to co-lead it with Mann. The reorganization placed Labs under the office of president Daniela Amodei, with Ami Vora taking over as head of product and Rahul Patil continuing as chief technology officer. Anthropic said Labs would expand its headcount as it pursued new product bets, structuring its work around early-stage exploration, testing with select users, and scaling successful ideas.[1] As of 2026, co-leading Labs is Mann's primary role at the company.
Mann is an outspoken commentator on the trajectory of artificial intelligence and the case for prioritizing safety. In a widely circulated 2025 interview, he argued that focusing on safety is what gave Claude its distinctive character, and he reiterated that a desire to put safety first was the reason he and his colleagues left OpenAI to found Anthropic.[2]
He has offered concrete and sometimes stark predictions. Mann has proposed an "economic Turing test," in which a system would be judged to have reached artificial general intelligence if it could perform the economically valuable work of a human, and he has suggested that this threshold could plausibly be met around 2027 or 2028. He has publicly entertained the possibility that superintelligence could emerge as early as 2028, and has said he believes significant labor-market disruption, on the order of roughly 20 percent unemployment, may be an unavoidable consequence of rapid AI progress. Mann frames these forecasts as reasons to invest more heavily in alignment now, warning that once systems reach superintelligence it may be too late to make them safe.[2]
Mann's most cited work, the GPT-3 paper, is among the most referenced publications in modern machine learning, and his co-authored papers on RLHF, Constitutional AI, and transformer interpretability are frequently cited within the AI safety and alignment literature.[9] As a co-founder of Anthropic, he is regularly invited to speak at industry events, including the AI Conference and the Open Data Science Conference, where he has presented on building Claude and on aligning AI systems with human values.[8] He maintains a personal website and is active on the social platform X under the handle @8enmann.