Eric Steinberger
Last reviewed
Jun 8, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,457 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,457 words
Add missing citations, update stale details, or suggest a clearer explanation.
Eric Steinberger is an Austrian computer scientist and technology entrepreneur who is the co-founder and chief executive officer of Magic.dev, a San Francisco company building what it calls an "AI software engineer" that aims to automate large parts of software development. Before founding Magic in 2022, Steinberger worked on game-theoretic reinforcement learning at Meta AI's Fundamental AI Research lab, where he collaborated with Noam Brown on poker-playing systems, and he co-founded the climate-education nonprofit ClimateScience. Magic became one of the most heavily funded AI coding startups, raising more than 400 million dollars from backers including former Google chief executive Eric Schmidt, and it drew wide attention for claiming a context window of 100 million tokens, which the company said was far larger than that of any other reported model.[1][2]
Steinberger was born and raised in Vienna, Austria, and has said that an early love of mathematics turned into an obsession with artificial intelligence by the age of 14.[7] He attended HTBLVA Spengergasse, a technical secondary school in Vienna, in a track for gifted students; a classmate there, Sebastian De Ro, would later become his co-founder at Magic.[7] In 2017 he worked as a student researcher in robotics at the Vienna University of Technology (TU Wien), and he later held a brief research role at the Massachusetts Institute of Technology.[7]
Around 2019 Steinberger enrolled to study computer science at the University of Cambridge, but he left after roughly a year to pursue research and entrepreneurship full time; published accounts differ on whether he completed any degree.[1][7] During the same period he was a researcher in deep reinforcement learning at Facebook AI Research, later part of Meta AI, from about 2019 to 2021.[7][8] His work centered on game theory and counterfactual regret minimization, the family of algorithms behind superhuman poker programs. He released open-source frameworks such as PokerRL and Deep-CFR, wrote the 2019 paper "Single Deep Counterfactual Regret Minimization," and co-authored the 2020 paper "DREAM: Deep Regret Minimization with Advantage Baselines and Model-Free Learning" with Noam Brown and Adam Lerer.[8][11] He stayed engaged with safety research after founding Magic, appearing among the many co-authors of the widely cited 2025 position paper "Chain of Thought Monitorability."[8]
In 2019, while at Cambridge, Steinberger co-founded ClimateScience with fellow student Isabel Key. Registered as a charity in the United Kingdom, the organization produces free climate-education materials and online courses, and Steinberger served as its founder and chair of the board before shifting his focus to Magic.[7]
Steinberger founded Magic, incorporated as Magic AI, Inc., in 2022 with Sebastian De Ro, the former chief technology officer of FireStart, an Austrian business-process-automation company based in Linz.[1][7] Steinberger has said he started the company after concluding that artificial general intelligence was closer than he had previously believed, and that automating software engineering was the most direct path toward it.[6] The company's stated ambition is unusually direct: in 2024 Steinberger wrote that Magic wanted to "build an AI model that can design, code and secure the next version of itself," a goal he acknowledged would take years.[4]
Magic positions its product as an "AI software engineer" rather than an autocomplete tool, competing in the same broad market as AI code generation assistants such as GitHub Copilot and Cursor. Its central bet is that a coding model becomes far more capable when it can hold an entire codebase, its documentation, and its dependencies in memory at once, rather than retrieving fragments. As of 2026 the company remained research-focused and comparatively small, with roughly 109 employees reported in April 2026, and it had not launched a broadly available consumer product, instead continuing to train larger models and build software-engineering tools on top of them.[10]
Magic's models are branded LTM, short for Long-Term Memory. In June 2023 the company unveiled LTM-1, a prototype it described as the first model built on a new "Long-term Memory Network" architecture, with a context window of about 5 million tokens, enough to take in roughly 500,000 lines of code. Magic cautioned that LTM-1 had fewer parameters than the frontier large language models of the time and was therefore less capable in other respects.[3]
In August 2024 Magic announced LTM-2-mini, which it said was its first model trained for a 100 million token context window, equivalent to about 10 million lines of code or 750 novels.[2] The company said the 100 million token figure was by far the largest context window of any model it was aware of, comparing it with the 2 million token windows of Google's Gemini models.[2][9] Magic argued that conventional attention was impractical at that scale: it estimated that running Meta's Llama 3.1 405B with a 100 million token context would require on the order of 638 NVIDIA H100 GPUs per user just to store the key-value cache, whereas its own sequence-dimension algorithm needed only a small fraction of a single H100's memory per user and was roughly 1,000 times cheaper per decoded token.[2] Independent reporting treated these figures as company claims rather than independently verified benchmarks.[9]
Alongside the model, Magic released HashHop, an open-source evaluation for long-context systems. HashHop asks a model to memorize pairs of random, incompressible hash strings and then complete multi-step chains across them, with the order shuffled so the model cannot exploit recency or semantic shortcuts. Magic presented it as a more rigorous successor to the popular "Needle in a Haystack" test, which it argued was too easy because hidden facts stood out from the surrounding text.[2]
Magic raised money rapidly across several rounds. After an early seed investment, it announced a 23 million dollar Series A in 2023 backed by Alphabet's growth fund CapitalG, former GitHub chief executive Nat Friedman, and the investor Elad Gil. In February 2024 it raised a 117 million dollar round led by the investment partnership of Nat Friedman and Daniel Gross, at a reported valuation of about 500 million dollars.[5][10]
The defining round came in August 2024, when Magic announced 320 million dollars in new funding from investors including Eric Schmidt, Sequoia Capital, Atlassian, Jane Street, CapitalG, Elad Gil, and the Friedman and Gross partnership.[1] The round brought the company's total funding to roughly 465 million dollars, although Magic and some outlets put the cumulative figure at about 515 million dollars.[1][2] Reuters had reported the previous month that Magic was seeking a valuation of about 1.5 billion dollars; the company did not officially confirm a valuation for the round, and the 1.5 billion dollar figure is best understood as reported rather than confirmed.[1][10]
The August 2024 announcement also disclosed a partnership with Google Cloud to build two supercomputers for training. Magic-G4 would use NVIDIA H100 GPUs, and Magic-G5 would use NVIDIA's newer GB200 NVL72 (Blackwell) systems, scalable to tens of thousands of Blackwell GPUs over time; some reports put the combined capacity on the order of 160 exaflops.[1][2]
| Round | Date | Amount | Selected investors | Reported valuation |
|---|---|---|---|---|
| Series A | 2023 | $23 million | CapitalG, Nat Friedman, Elad Gil | Not disclosed |
| Series B | February 2024 | $117 million | Nat Friedman and Daniel Gross, CapitalG, Elad Gil | about $500 million |
| Growth round | August 2024 | $320 million | Eric Schmidt, Sequoia, Atlassian, Jane Street, CapitalG | about $1.5 billion (reported) |
As of June 2026, Steinberger remained co-founder and chief executive of Magic, which continued to develop larger LTM-2 models and to position itself among the most ambitious attempts to automate software engineering. His combination of early game-theoretic AI research, a climate-education nonprofit, and a heavily funded long-context startup has made him one of the more closely watched young founders in the field.[6][10]