# Steve Hsu

> Source: https://aiwiki.ai/wiki/steve_hsu
> Updated: 2026-10-29
> Fact-checked: 2026-10-29
> Categories: AI Research, AI for Science, People
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Steve Hsu." aiwiki.ai, 29 Oct 2026. https://aiwiki.ai/wiki/steve_hsu
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Steve Hsu** (published as **Stephen D. H. Hsu**) is an American theoretical physicist, startup founder and podcast host. He is Professor of Theoretical Physics and of Computational Mathematics, Science, and Engineering at Michigan State University (MSU), where he served as the university's senior research administrator from 2012 until his resignation from that post in 2020.[1][2][3] He founded the internet privacy company SafeWeb, which Symantec acquired in 2003, co-founded the embryo genetic testing company Genomic Prediction, and co-founded the AI startup Superfocus.ai.[1][4][5][6] Since 2019 he has hosted the Manifold podcast, and he writes the blog Information Processing.[7][3] In AI discussion he comments mainly on [large language models](https://aiwiki.ai/wiki/large_language_model) in mathematics and theoretical physics and on the Chinese AI industry, which he follows partly through visits to Chinese universities and AI labs.[8][9]

He uses "Steve Hsu" for his blog, podcast and X account (@hsu_steve), while his scientific papers carry the byline Stephen D. H. Hsu.[3][10]

## Education and academic career

Hsu received a bachelor's degree in physics from the California Institute of Technology in 1986, a master's degree in physics from the University of California, Berkeley in 1989, and a doctorate in physics from Berkeley in 1991. He was then named a Junior Fellow of the Harvard Society of Fellows.[1] He became an assistant professor at Yale University in 1995 and a full professor at the University of Oregon in 2005, and before moving to MSU he directed Oregon's Institute for Theoretical Science.[1] Reason magazine reported in its March 2002 issue that he was on a year-long hiatus from his Oregon physics post while running SafeWeb.[11]

The MSU Board of Trustees approved his appointment as vice president for research and graduate studies at a special meeting on July 23, 2012, effective August 20, 2012. He succeeded Ian Gray, who had held the post since 2004. MSU President Lou Anna K. Simon cited "the breadth of Stephen Hsu's experience as a scientist and scholar, as well as a Silicon Valley entrepreneur and founder of two companies."[1] By 2020 his title was senior vice president for research and innovation.[2] He left the administrative role on July 1, 2020 and returned to the tenured faculty (see [2020 resignation](#2020-resignation-as-vice-president)).[2]

| Period | Position | Source |
|---|---|---|
| 1986 | BS in physics, Caltech | [1] |
| 1991 | PhD in physics, UC Berkeley (MA 1989) | [1] |
| After 1991 | Junior Fellow, Harvard Society of Fellows | [1] |
| 1995 | Assistant professor, Yale University | [1] |
| 2005 | Full professor, University of Oregon (also director of its Institute for Theoretical Science by 2012) | [1] |
| 2012-2020 | Vice president for research and graduate studies at MSU, later senior vice president for research and innovation | [1][2] |
| Current (October 2026) | Professor of Theoretical Physics and of Computational Mathematics, Science, and Engineering, MSU | [3][12] |

## Physics research

MSU's 2012 announcement described Hsu's research as applications of quantum field theory, particularly to dark energy, black holes and particle physics.[1] His current biography lists quantum field theory, black holes, cosmology, quantum foundations, AI and machine learning, and computational genomics as research interests.[3]

| Year | Paper | Venue | Source |
|---|---|---|---|
| 2003 | "Quantum Production of Black Holes" | Physics Letters B | [13] |
| 2004 | "Entropy Bounds and Dark Energy" | Physics Letters B | [14] |
| 2004 | "Minimum Length from Quantum Mechanics and Classical General Relativity" (with Xavier Calmet and Michael Graesser) | Physical Review Letters | [15] |
| 2009 | "Black holes, information and decoherence" (with David Reeb) | Physical Review D | [16] |
| 2024 | "Replica Wormholes and Quantum Hair" (with Xavier Calmet) | arXiv preprint | [17] |
| 2025 | "Relativistic Covariance and Nonlinear Quantum Mechanics: Tomonaga-Schwinger Analysis" | Physics Letters B | [18] |

The 2025 paper is the one Hsu has described as, in his view, the first theoretical physics paper whose main idea came from an AI model (see [Using LLMs in physics research](#using-llms-in-physics-research)).[19]

## Computational genomics

From the 2010s Hsu applied methods from high-dimensional statistics to human genetics. In a 2014 arXiv review on the genetic architecture of intelligence and other quantitative traits, he used results from compressed sensing ([L1-penalized regression](https://aiwiki.ai/wiki/lasso_regression)) to estimate the sample sizes needed to build genomic predictors, which he put at roughly a million people for cognitive ability.[20] A 2018 paper in *Genetics* by Hsu's group, using almost 500,000 genotypes from the UK Biobank, built predictors for height, heel bone density and educational attainment that captured about 40, 20 and 9 percent of total variance respectively in data not used for training. Predicted heights correlated at about 0.65 with actual heights.[10] The paper lists Hsu's affiliations as MSU and the Cognitive Genomics Laboratory at BGI-Shenzhen.[10]

MIT Technology Review reported in 2017 that Hsu had long been interested in genetic selection for intelligence. In a 2014 essay titled "Super-Intelligent Humans Are Coming" he argued that selecting embryos for intelligence could raise a child's IQ by 15 points.[5] The same article reported that he had been closely involved with a project in China that aimed to sequence the genomes of "mathematical geniuses." It also noted that polygenic scoring algorithms of the time were built mostly on data from people of Northern European ancestry, which could limit their use for other populations.[5]

## Companies

| Company | Hsu's role | Field | Notes | Source |
|---|---|---|---|---|
| SafeWeb | Founder | Anonymous web browsing, later SSL VPN appliances | Based in Emeryville, California; acquired by Symantec for $26 million in cash in October 2003 | [11][4] |
| Robot Genius | Founder | Information security | Oakland, California | [1] |
| Genomic Prediction | Co-founder | Genetic testing of IVF embryos | Founded 2017 with Laurent Tellier (CEO) and Nathan Treff (chief scientific officer) | [5] |
| Othram | Founder (per Hsu's biography) | Forensic DNA and genetic genealogy | Listed among his startups in his own biography | [3] |
| Superfocus.ai | Co-founder | AI applications | Co-founded with CEO Tushar Sheth | [6][21] |

### SafeWeb

SafeWeb offered anonymous web access. Its free Triangle Boy software let volunteers turn their computers into entry points to the SafeWeb network, so that users in countries that blocked the SafeWeb site could still reach it.[22][11] In a 2001 Voice of America report, Hsu, then the company's president, said the service had users in China who could read Taiwanese newspapers and The New York Times, which were blocked there.[22] Reason reported in 2002 that the company had seed money from several venture capital firms and from the Central Intelligence Agency.[11] Symantec bought SafeWeb, by then a maker of clientless SSL VPN appliances, in a deal that closed on October 15, 2003.[4]

### Genomic Prediction

Genomic Prediction was founded in 2017 to offer polygenic risk scores for embryos created through in vitro fertilization. At launch, the company said it would report only disease risks, identifying embryos expected to develop serious medical problems.[5] As of October 2026, Hsu's biography describes the company's work as "DNA embryo screening for IVF" and states that more than 250,000 embryos have been genotyped.[3]

### Superfocus.ai

Superfocus.ai describes itself as building "AIs that see and hear, understand, and act for themselves." Its website lists Tushar Sheth as co-founder and CEO and Hsu as co-founder.[6] A 2026 Embedded Vision Summit speaker profile describes it as a seed-stage startup backed by Cercano Management, Trust Ventures, Metaplanet and other investors. At that May 2026 conference Sheth presented a session on running multimodal AI on low-cost chips, based on tests of 17 systems-on-chip.[21] Hsu's own biography describes the company's work as "AI applications, including V-L models on edge hardware," meaning [vision-language models](https://aiwiki.ai/wiki/vision_language_model) running on [edge devices](https://aiwiki.ai/wiki/edge_ai).[3]

In earlier public talks Hsu linked Superfocus to the problem of LLM [hallucination](https://aiwiki.ai/wiki/hallucination). A 2023 lecture he gave to theoretical physicists at Oxford, released as a Manifold episode, has a chapter on hallucination followed by one on "SuperFocus testing and examples."[23] In an as-told-to essay for Business Insider, published in Japanese translation on February 6, 2025, Hsu said Superfocus built enterprise AI for uses such as customer service, ordering, delivery, document analysis and travel scheduling, usually on models like GPT-4o. He said the company was testing [DeepSeek-V3](https://aiwiki.ai/wiki/deepseek_v3) and planned to move much of its work to open-weight models. His three reasons were cost (he estimated DeepSeek's models were about 30 times cheaper to run than comparable OpenAI models), speed for voice applications, and privacy, since open models can run inside a client's own cloud environment.[24]

## Information Processing and Manifold

Hsu wrote Information Processing on Google's Blogspot for many years. In May 2024 he moved it to Substack, where his older posts were imported.[25][3]

Manifold began on January 24, 2019, with an introductory episode in which Hsu and co-host Corey Washington introduced each other.[7] The first run reached episode 50 in June 2020.[26] When the show resumed in January 2022, episode numbering started again from 1.[27] The 123rd episode of the new numbering appeared on October 22, 2026.[28][8] The show's About page describes it as wide-ranging conversations with "writers, scientists, technologists, academics, entrepreneurs, investors, and more."[12]

AI became a recurring subject from 2023. In "ChatGPT, LLMs, and AI" (February 2023) Hsu explained how LLMs work and discussed hallucination.[29] In the Oxford lecture episode (May 2023) he covered deep learning history, word embeddings, the transformer architecture, hallucination and the possibility of AGI.[23] Guests have included [Misha Laskin](https://aiwiki.ai/wiki/misha_laskin) of [Reflection AI](https://aiwiki.ai/wiki/reflection_ai) in March 2025.[30] Episodes in 2026 covered [recursive self-improvement](https://aiwiki.ai/wiki/recursive_self-improvement), [effective accelerationism](https://aiwiki.ai/wiki/effective_accelerationism) and the US-China AI race.[31][28]

Hsu co-created *Machine God*, a feature-length documentary about AI, with John Greer and Lei Huang. A September 2026 Manifold episode with the three filmmakers covered the film's origins and its themes, including calls to pause or slow AI development, recursive self-improvement and gradual disempowerment.[32] In October 2026 Hsu described the film as being about AI existential risk, built mainly on interviews with people in the San Francisco Bay Area AI industry.[8]

## Using LLMs in physics research

In December 2025 Hsu wrote that he thought he had published "the first research article in theoretical physics in which the main idea came from an AI," in this case GPT-5. The paper, on the Tomonaga-Schwinger formulation of quantum field theory and state-dependent (nonlinear) modifications of quantum mechanics, appeared in *Physics Letters B*.[19][18] In a companion write-up he described a "Generator - Verifier" workflow in which separate model instances generate and check research steps. He said he used GPT-5, Gemini and Qwen-Max to do calculations, find errors and produce the finished paper. He argued that this role separation, with human oversight, makes results more reliable than single-shot inference, while noting that the models are "subject to confabulation."[19] In a Manifold episode on the work he compared frontier models to a "brilliant but unreliable genius colleague."[33]

The paper drew a critique. In a December 2025 preprint, UCL physicist Jonathan Oppenheim argued that the criterion "proposed by GPT-5 and published in Physics Letters B" actually tests locality of the Hamiltonian and does not detect whether a theory is nonlinear. He pointed instead to the criterion identified by Gisin and Polchinski about 35 years earlier.[34]

## Views on AI

Hsu's comments on AI are his own forecasts and assessments. They come mostly from his podcast and posts on X.

| Date | Venue | Position (attributed to Hsu) | Source |
|---|---|---|---|
| February 2025 | Business Insider essay | Said he would be very surprised if there were no major shift toward open-source models in 2026. Called [DeepSeek](https://aiwiki.ai/wiki/deepseek) an impressive, open company that publishes clear papers, and said he did not believe online claims that DeepSeek's models were derived from OpenAI's or that its reported training cost was false | [24] |
| October 8, 2026 | Manifold #122 | Argued that AI revenue must reach trillions of dollars a year within a few years (more than a tenfold increase) or current capital spending will prove a large misallocation, and that [OpenAI](https://aiwiki.ai/wiki/openai) and [Anthropic](https://aiwiki.ai/wiki/anthropic) could be overvalued even if it does, because of competition from other US and Chinese labs | [35] |
| October 22, 2026 (UTC) | X | Said the inference efficiency behind DeepSeek's margins is "based on architecture innovation, not distillation"; US labs lead at the high end, but DeepSeek is "smart enough" for everyday tasks, including medium-hard software work | [36] |
| October 22, 2026 | Manifold #123 | Recorded in Beijing; see below | [8] |
| October 23, 2026 (UTC) | X | Called OpenAI's [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) and [GPT-6.1 Sol](https://aiwiki.ai/wiki/gpt_6_1_sol) the best models at math and theoretical physics, with Claude "a bit behind," and said Chinese labs, being smaller, had put modest effort into math after reaching IMO gold-medal level and prioritized coding and agentic tasks | [37] |
| October 28, 2026 (UTC) | X | Commenting on [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) Intelligence Index v5, see below | [9] |

### Manifold #123: AI and mathematics, from Beijing

Hsu recorded episode 123 in Beijing while visiting the Yau Mathematical Sciences Center at Tsinghua University, where he said he was working with researchers in mathematics, theoretical physics and AI.[8] He said frontier models such as Astra and Claude were "superhuman" at mathematics and theoretical physics in the sense that they can do things even the best humans find extremely hard, while they might still make mistakes that humans are unlikely to make. He predicted that progress in these fields would be "dominated by AI results," with humans following behind. He contrasted what he described as the view of [Terence Tao](https://aiwiki.ai/wiki/terence_tao) and many other mathematicians, that AI would discourage people from doing mathematics, with S. T. Yau's announced plan in China to build an alliance of a thousand mathematicians who would embrace AI and open-source models.[8]

Hsu said his collaborators had built an automated method for producing hard mathematical training data and new benchmarks. By his account, Astra and Claude succeed on roughly half of these problems and the open models on about a fifth, and the team had discussed supplying such data to several Chinese labs.[8] He described compute and capital as the main limits on Chinese labs. He reported that people he spoke with expected even pretraining to be feasible on domestic Chinese hardware such as [Huawei Ascend](https://aiwiki.ai/wiki/huawei_ascend) within roughly a year. He also said that Chinese AI researchers he met were less "AGI-pilled" than their US counterparts and focused on near-term safety problems such as cybersecurity, job loss and bioweapons rather than long-term existential risk.[8]

### Comments on Intelligence Index v5

Artificial Analysis announced version 5 of its Intelligence Index on October 28, 2026 (UTC), adding Cognition's FrontierCode 2.0 coding benchmark among other changes.[38] Hsu responded the same day. He called the result of [DeepSeek V4.1 Flash](https://aiwiki.ai/wiki/deepseek_v4_1_flash) 1017 on [FrontierCode](https://aiwiki.ai/wiki/frontiercode) 2.0 impressive and said it "can't be the result of benchmaxxing." He argued that weak science and math scores had pulled Chinese models down relative to US frontier models such as GPT-6 Astra and [Claude Opus 5.5](https://aiwiki.ai/wiki/claude_opus_5_5). He predicted that Chinese labs would reprioritize math and science, while doubting they would shift more than 10 percent of their specialized model-capability resources that way. He called the [Gemini 4 Carbon](https://aiwiki.ai/wiki/gemini_4_carbon) scores "amazing." He also wrote that "adversarial verification of math+phys results by the best models is much better than human peer review."[9]

## 2020 resignation as vice president

In June 2020 a petition circulated calling for Hsu's removal as MSU's senior vice president for research and innovation. Local broadcaster WILX reported that critics pointed to comments including a 2017 interview with Stefan Molyneux and a blog post in which Hsu cited a 2019 study concluding that there was "no widespread racial bias in police shootings."[39] Hsu's supporters organized their own petition and letters of support.[40]

On June 19, 2020, MSU President Samuel L. Stanley Jr. announced that Hsu was resigning from his leadership position and returning to a tenured faculty position effective July 1. Stanley said that the university supported faculty members' academic freedom to "address the most difficult and controversial issues," but that "when senior administrators at MSU choose to speak out on any issue, they are viewed as speaking for the university as a whole."[2] Hsu wrote the same day that Stanley had asked for his resignation, that he disagreed with the decision because "serious issues of academic freedom and freedom of inquiry are at stake," and that, since he served at the president's pleasure, he had agreed to resign.[40]

## References

1. Michigan State University. ["Stephen Hsu named new MSU research vice president"](https://web.archive.org/web/2013/http://msutoday.msu.edu/news/2012/stephen-hsu-named-new-msu-research-vice-president/). MSUToday, July 2012 (archived copy).
2. Michigan State University. ["Statement on resignation of Stephen Hsu"](https://msutoday.msu.edu/news/2020/06/2020-06-19-statement-hsu-resignation). MSUToday, June 19, 2020.
3. Steve Hsu. ["About"](https://stevehsu.substack.com/about). Information Processing (Substack). Accessed October 30, 2026.
4. Computerworld. ["Symantec purchases SSL VPN maker SafeWeb"](https://www.computerworld.com/article/1321782/symantec-purchases-ssl-vpn-maker-safeweb.html). October 20, 2003.
5. Antonio Regalado. ["Eugenics 2.0: We're at the dawn of choosing embryos by health, height, and more"](https://www.technologyreview.com/2017/11/01/105176/eugenics-20-were-at-the-dawn-of-choosing-embryos-by-health-height-and-more/). MIT Technology Review, November 1, 2017.
6. Superfocus. [Superfocus.ai home page](https://superfocus.ai/). Accessed October 30, 2026.
7. Manifold. ["Introductions, Episode #0"](https://share.transistor.fm/s/b5baa956). January 24, 2019.
8. Steve Hsu. ["AI and Mathematics, from Beijing, #123" (transcript)](https://www.manifold1.com/episodes/ai-and-mathematics-from-beijing-123/transcript). Manifold, October 22, 2026.
9. Steve Hsu (@hsu_steve). [Post on X](https://x.com/hsu_steve/status/2115589096772919414). October 28, 2026 (UTC).
10. Louis Lello, Steven G. Avery, Laurent Tellier, Ana I. Vazquez, Gustavo de los Campos, Stephen D. H. Hsu. ["Accurate Genomic Prediction of Human Height"](https://doi.org/10.1534/genetics.118.301267). Genetics 210(2): 477-497, October 2018. arXiv:1709.06489.
11. Sara Rimensnyder. ["Triangle Boy Howdy"](https://reason.com/2002/03/01/triangle-boy-howdy-2/). Reason, March 2002 issue.
12. Steve Hsu. ["About"](https://www.manifold1.com/about). Manifold podcast. Accessed October 30, 2026.
13. Stephen D. H. Hsu. ["Quantum Production of Black Holes"](https://arxiv.org/abs/hep-ph/0203154). Physics Letters B 555: 92-98 (2003).
14. Stephen D. H. Hsu. ["Entropy Bounds and Dark Energy"](https://arxiv.org/abs/hep-th/0403052). Physics Letters B 594: 13-16 (2004).
15. Xavier Calmet, Michael Graesser, Stephen D. H. Hsu. ["Minimum Length from Quantum Mechanics and Classical General Relativity"](https://arxiv.org/abs/hep-th/0405033). Physical Review Letters 93, 211101 (2004).
16. Stephen D. H. Hsu, David Reeb. ["Black holes, information and decoherence"](https://arxiv.org/abs/0903.2258). Physical Review D 79, 124037 (2009).
17. Xavier Calmet, Stephen D. H. Hsu. ["Replica Wormholes and Quantum Hair"](https://arxiv.org/abs/2404.02331). arXiv:2404.02331, April 2024.
18. Stephen D. H. Hsu. ["Relativistic Covariance and Nonlinear Quantum Mechanics: Tomonaga-Schwinger Analysis"](https://www.sciencedirect.com/science/article/pii/S0370269325008111). Physics Letters B (2025). arXiv:[2511.15935](https://arxiv.org/abs/2511.15935).
19. Steve Hsu. ["Theoretical Physics with Generative AI"](https://stevehsu.substack.com/p/theoretical-physics-with-generative). Information Processing, December 2, 2025.
20. Stephen D. H. Hsu. ["On the genetic architecture of intelligence and other quantitative traits"](https://arxiv.org/abs/1408.3421). arXiv:1408.3421, August 2014.
21. Edge AI and Vision Alliance. ["Tushar Sheth"](https://embeddedvisionsummit.com/2026/speaker/tushar-sheth/). 2026 Embedded Vision Summit speaker page. Accessed October 30, 2026.
22. Voice of America. ["Travelling the Internet Invisibly"](https://www.voanews.com/amp/a-13-a-2001-05-30-2-travelling-66952532/377992.html). May 30, 2001.
23. Steve Hsu. ["Artificial Intelligence & Large Language Models: Oxford Lecture, #35"](https://www.manifold1.com/episodes/artificial-intelligence-large-language-models-oxford-lecture-35). Manifold, May 11, 2023.
24. Business Insider Japan. ["AIスタートアップの創業者として私が「DeepSeekに移行しよう」と思う3つの理由"](https://www.businessinsider.jp/article/deepseek-ai-openai-china/) (Japanese translation of a Business Insider as-told-to essay by Jane Zhang, based on a conversation with Steve Hsu). February 6, 2025.
25. Steve Hsu. ["Information Processing (this blog) has moved to Substack!"](https://infoproc.blogspot.com/). Information Processing (Blogspot), May 12, 2024.
26. Manifold. ["Warren Hatch on Seeing the Future in the Era of COVID-19, #50"](https://share.transistor.fm/s/e4e1d582). June 11, 2020.
27. Manifold. ["James Lee on Polygenic Prediction and Embryo Selection, #1"](https://share.transistor.fm/s/a1c33069). January 24, 2022.
28. Manifold. [Episode list](https://www.manifold1.com/episodes). Accessed October 30, 2026.
29. Steve Hsu. ["ChatGPT, LLMs, and AI, #29"](https://www.manifold1.com/episodes/chatgpt-llms-and-ai). Manifold, February 2, 2023.
30. Steve Hsu. ["Misha Laskin, Reflection.ai: From Physics to SuperIntelligence, Manifold Episode 81"](https://stevehsu.substack.com/p/misha-laskin-reflectionai-from-physics). Information Processing, March 13, 2025.
31. Steve Hsu. ["State of AI, Summer 2026, #116"](https://www.manifold1.com/episodes/state-of-ai-summer-2026-116). Manifold, July 16, 2026.
32. Steve Hsu. ["Machine God: Hsu, Greer, Huang, #120"](https://www.manifold1.com/episodes/machine-god-hsu-greer-huang-120). Manifold, September 10, 2026.
33. Steve Hsu. ["Theoretical Physics With Generative AI, #101"](https://www.manifold1.com/episodes/theoretical-physics-with-generative-ai-101). Manifold, December 18, 2025.
34. Jonathan Oppenheim. ["Nonlinear Quantum Mechanics and Artificial Intelligence"](https://arxiv.org/abs/2512.07809). arXiv:2512.07809, December 8, 2025.
35. Steve Hsu. ["AI Boom or Bust?, #122"](https://www.manifold1.com/episodes/ai-boom-or-bust-122). Manifold, October 8, 2026.
36. Steve Hsu (@hsu_steve). [Post on X](https://x.com/hsu_steve/status/2113065498212728858). October 22, 2026 (UTC).
37. Steve Hsu (@hsu_steve). [Post on X](https://x.com/hsu_steve/status/2113454060024270917). October 23, 2026 (UTC).
38. Artificial Analysis (@ArtificialAnlys). ["Announcing Artificial Analysis Intelligence Index v5"](https://x.com/ArtificialAnlys/status/2115265133467635877). Post on X, October 28, 2026 (UTC).
39. WILX News 10. ["MSU Vice President of research and innovation Stephen Hsu resigns"](https://www.wilx.com/content/news/MSU-Vice-President-of-research-and-innovation-Stephen-Hsu-resigns-571381341.html). June 19, 2020.
40. Steve Hsu. ["Resignation"](https://infoproc.blogspot.com/2020/06/resignation.html). Information Processing, June 19, 2020.