Christian Szegedy

Christian Szegedy is a Hungarian mathematician and machine-learning researcher, founder of Math Inc, first author of the GoogLeNet Inception paper and the original adversarial-examples paper, and a former founding team member of xAI from 2023 to 2025.
Nextomoro
8 min read
Christian Szegedy

Christian Szegedy

Christian Szegedy is a Hungarian mathematician and machine-learning researcher known for foundational work on deep convolutional architectures, neural-network normalization, adversarial examples, and the use of deep learning in formal mathematics. He is first author of "Going Deeper with Convolutions" (the GoogLeNet Inception paper, CVPR 2015) and "Intriguing Properties of Neural Networks" (2013, which introduced adversarial examples), and co-author with Sergey Ioffe of "Batch Normalization" (ICML 2015), winner of the ICML 2025 Test-of-Time Award. As of May 2026, he is the founder of Math Inc, focused on verified superintelligence via autoformalization, after a founding-team role at xAI from March 2023 to February 2025 and Chief Scientist at Morph Labs in mid-2025.

At a glance

Origins

Public biographical material on Szegedy is comparatively thin. He has no English-language Wikipedia entry as of May 2026, and the available record runs through his LinkedIn page, his X account, his Google Scholar profile, his Wikidata entry, the Mathematics Genealogy Project entry on his PhD, the published papers, and press coverage of his 2025 transitions. The Hungarian form of his given name appears in academic listings as "Krisztián Szegedy", including on programs at the Alfréd Rényi Institute of Mathematics in Budapest.

Szegedy studied mathematics at Eötvös Loránd University in Budapest from 1990 to 1992, then moved to Germany for doctoral work at the University of Bonn, joining the Research Institute for Discrete Mathematics under Bernhard Korte. The dissertation "Some Applications of the Weighted Combinatorial Laplacian" was completed in 2005.

Career

After the PhD, Szegedy joined Cadence Design Systems as a research scientist from 2005 to 2010, working on electronic-design-automation problems and producing patents in approximate placement methods, including US8572540B2. In 2010 he joined Google as a software engineer, becoming a Staff Research Scientist by 2015. The twelve-year Google period produced the body of work for which he is most widely known.

The first canonical contribution is "Intriguing Properties of Neural Networks" (Szegedy, Zaremba, Sutskever, Bruna, Erhan, Goodfellow, Fergus, 2013). The paper showed that imperceptible perturbations of an input image could cause a trained image classifier to misclassify with high confidence, and that the same perturbations frequently transfer across networks trained on different data subsets. The result introduced "adversarial examples" as a term and a research line that has continued for more than a decade.

The second is the GoogLeNet Inception architecture in "Going Deeper with Convolutions" (Szegedy et al., CVPR 2015). The paper introduced the Inception module, which combines convolutions of multiple receptive-field sizes within a single block, and used twenty-two such blocks to win the ImageNet 2014 classification track. "Rethinking the Inception Architecture" (Szegedy, Vanhoucke, Ioffe, Shlens, Wojna, CVPR 2016) introduced Inception v2 and v3.

The third is "Batch Normalization" (Ioffe and Szegedy, ICML 2015), which computes per-batch normalization statistics for each activation, allowing higher learning rates and less careful initialization. The paper received the ICML 2025 Test-of-Time Award. Szegedy is also a contributing author on "SSD: Single Shot MultiBox Detector" (Liu et al., ECCV 2016).

The later Google years pivoted to formal mathematics. Szegedy led the N2Formal team on neural theorem proving, including "HOList" (Bansal, Loos, Rabe, Szegedy, Wilcox, ICML 2019). His position paper "A Promising Path Towards Autoformalization and General Artificial Intelligence" (CICM 2020) argued that autoformalization, the AI-driven translation of natural-language mathematics into machine-checkable formal proofs, is a promising route to general reasoning capability. The paper has framed his subsequent commercial work.

In March 2023 Szegedy left Google to co-found xAI with Elon Musk, as one of the eleven publicly named founding team members. The team was assembled from senior researchers at Google DeepMind, OpenAI, and Google Brain, and the launch was publicly announced on July 12, 2023. Press coverage of his subsequent departure characterized him as the Grok 3 reasoning lead, working alongside Yuhuai (Tony) Wu, his former Google colleague.

Szegedy left xAI in February 2025. In June 2025 he joined Morph Labs as Chief Scientist, leading the development of verified superintelligence via autoformalization; he had previously been a seed investor in the firm. The Trinity autoformalization system was announced shortly after, formalizing in Lean a classical de Bruijn theorem related to the abc conjecture.

On September 11, 2025 Szegedy announced on X that he was starting Math Inc, dedicated to verified superintelligence via autoformalization, built on the reinforcement-learning infrastructure originally developed at Morph Labs. The announcement reported that the Math Inc agent Gauss had completed the formalization of Terence Tao and Alex Kontorovich's Strong Prime Number Theorem project in approximately three weeks, after eighteen months of stalled human progress.

Affiliations

  • Eötvös Loránd University: Mathematics student, 1990 to 1992.
  • University of Bonn, Research Institute for Discrete Mathematics: Doctoral student, 1998 to 2005, supervised by Bernhard Korte.
  • Cadence Design Systems: Research Scientist, 2005 to 2010.
  • Google Research: Software Engineer through Staff Research Scientist, 2010 to 2023. Lead of the N2Formal team.
  • xAI: Founding team member, March 2023 to February 2025.
  • Morph Labs: Chief Scientist, June to September 2025.
  • Math Inc: Founder, September 2025 to present.

Notable contributions

Szegedy's published record concentrates in deep convolutional architectures, neural-network training methods, adversarial robustness, and formal mathematics. His Google Scholar profile lists Going Deeper with Convolutions and Batch Normalization as the most-cited papers, each above 65,000 citations as of May 2026.

  • "Going Deeper with Convolutions" (Szegedy et al., CVPR 2015). First-author paper introducing the Inception module and GoogLeNet, winner of the ImageNet 2014 classification challenge. About 74,000 citations.
  • "Batch Normalization" (Ioffe and Szegedy, ICML 2015). Co-author paper introducing the batch-normalization layer; ICML 2025 Test-of-Time Award. About 68,000 citations.
  • "Rethinking the Inception Architecture for Computer Vision" (Szegedy et al., CVPR 2016). First-author paper introducing Inception v2 and v3. About 44,000 citations.
  • "Intriguing Properties of Neural Networks" (Szegedy et al., 2013). First-author paper introducing adversarial examples, a research line later extended by Ian Goodfellow, Aleksander Madry, and others.
  • "SSD: Single Shot MultiBox Detector" (Liu et al., ECCV 2016). Co-author paper on single-stage object detection.
  • "HOList" (Bansal, Loos, Rabe, Szegedy, Wilcox, ICML 2019). Co-author paper on a benchmark for neural theorem proving in HOL Light.
  • "A Promising Path Towards Autoformalization and General Artificial Intelligence" (Szegedy, CICM 2020). Sole-author position paper framing autoformalization as a route to general reasoning.
  • xAI founding-team research contributions (March 2023 to February 2025). Press coverage credited Szegedy as reasoning lead on Grok 3.
  • Gauss autoformalization agent (Math Inc, September 2025 to present). Reported to have autonomously completed the Strong Prime Number Theorem formalization project of Terence Tao and Alex Kontorovich.

Investments and boards

  • Morph Labs (AI / Software): Seed investor, prior to June 2025; subsequently Chief Scientist, June to September 2025.

Beyond the Morph Labs seed position, no separate public personal investor activity in AI, semiconductors, datacenters, software, or energy is on record.

Network

Szegedy's longest-running professional relationships fall in three cohorts. The first is the Google Research computer-vision and deep-learning cohort that produced the Inception, Batch Normalization, and SSD papers, including Sergey Ioffe, Vincent Vanhoucke, Jonathon Shlens, Wei Liu, and Dragomir Anguelov, plus the adversarial-examples co-authors Wojciech Zaremba, Ilya Sutskever, Ian Goodfellow, and Rob Fergus.

The second is the N2Formal team and the formal-mathematics community, including HOList co-authors Kshitij Bansal, Sarah Loos, Markus Rabe, and Bryan Wilcox. The autoformalization line links to Yuhuai (Tony) Wu, who worked with Szegedy at Google before they overlapped at xAI.

The third is the xAI founding team. Beyond Elon Musk, the cohort included Igor Babuschkin (departed August 2025 to launch Babuschkin Ventures), Greg Yang (informal advisor since January 2026), Wu (departed February 10, 2026), Jimmy Ba (departed February 10, 2026), Manuel Kroiss, Toby Pohlen, Ross Nordeen, Kyle Kosic, Guodong Zhang, and Zihang Dai.

Position in the field

As of May 2026, Szegedy occupies a structurally distinctive position among machine-learning researchers. The combination of first authorship on the GoogLeNet Inception paper, second authorship on Batch Normalization, and first authorship on the original adversarial-examples paper places him among the small group of authors at the top of the most-cited deep-learning papers of the 2010s. The Google Scholar citation count above 330,000 is high for the staff-research-scientist career stage, and the ICML 2025 Test-of-Time Award formalized the historical reach of Batch Normalization.

His 2025 sequence, from xAI to Morph Labs to Math Inc, placed him among the senior departures from xAI in the year preceding the SpaceX acquisition of February 2026. Press coverage characterized differences in research focus as the proximate reason for the exit, with the autoformalization research line as the destination. His public-commentary cadence runs through @ChrSzegedy, seminars at the Rényi Institute, TWIML AI Podcast episode 745, and the GTC 2024 Fireside Chat on automated reasoning for software synthesis and verification.

Outlook

Open questions over the next 6 to 18 months:

  • Math Inc capability cadence. Whether Gauss and successor systems demonstrate further autoformalization milestones beyond the Strong Prime Number Theorem result, including formalizations of currently open results or production use by working mathematicians.
  • Autoformalization market. Whether autoformalization remains a research-and-tools play or develops a commercial customer base in semiconductors, security-critical software, and academic mathematics.
  • Lean and Mathlib engagement. Whether Math Inc contributes upstream to the Lean and Mathlib communities, and at what pace.
  • xAI relationship. Whether the post-departure relationship remains technically active, given the Grok reasoning lineage.
  • Publication cadence. Whether the Math Inc work produces peer-reviewed papers at the cadence of the Google-era Inception and Batch Normalization line, or remains a commercial research effort.

Sources

Person xAI
About the author
Nextomoro

Nextomoro

nextomoro tracks progress for AI research labs, models, and what's next.

Read next

Ramin Hasani

Daniela Rus

Laurent Solly

Michael Rabbat

Pascale Fung

AI Research Lab Intelligence

nextomoro tracks progress for AI research labs, models, and what's next.

AI Research Lab Intelligence

Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to AI Research Lab Intelligence.

Your link has expired.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.