[{"content":"During my research on how to solve the Boltzmann equation, an integro-differential equation which forms the foundation of computational fluid dynamics (CFD), in certain scenarios interesting in spintronics*, I discovered that a discontinuous Galerkin (DG) method might be suited well in order to do so. Although I haven\u0026rsquo;t finished this research project yet, in the spirit of the incremental approach of software development and as the core functionality might be applicable in other scenarios like producing data for training neural operators, I decided to publish in an open-source PyPI package:\na solver applying a user-specified DG scheme to (systems of) partial differential equations, using a solver to compute numerical fluxes through the finite elements\u0026rsquo; boundaries via solving the associated Riemann problem. All in JAX. Built on the basis of diffrax.\nSee GitHub for more details.\n*(while there are a lot of software packages for solving the Boltzmann equation for phonons, there are none for the case of magnons)\n","permalink":"https://d-a-reiss.github.io/projects/dg-solvax/","summary":"Discontinuous Galerkin and Riemann solver in JAX.","title":"dg-solvax"},{"content":"Currently I\u0026rsquo;m reading the classic \u0026quot;Gödel, Escher, Bach\u0026quot; (GEB) by Douglas R. Hofstadter, years ago recommended to me by my stepfather after he read it when he was studying computer science besides working as a math and physics teacher (in retrospect, him doing so has impressed and inspired me, until today). Since then, I had GEB on my bookshelf, but never read it because I always thought: \u0026ldquo;This is not the right time, because\u0026hellip;\u0026rdquo; \u0026ldquo;I\u0026rsquo;ve got too many other things on my mind\u0026rdquo;, \u0026ldquo;there\u0026rsquo;s so much other stuff to do\u0026rdquo;, and so on\u0026hellip; (To be honest, I may also have gotten a little bit intimidated by the intellectual level of the book as he described it, but as it turned out, I actually find it quite approachable.)\nI guess the first chapters of GEB might feel to some readers like a mere intellectual exercise or play, something I\u0026rsquo;ve got mixed feelings about (I think being attracted to it is rooted in my character, while being repelled by it in conscience and this is one of the reasons which repelled me from academia, where I\u0026rsquo;ve got the feeling that a lot of professors I talked to are motivated by simply playing around in their intellectual world, not caring about the consequences, although at least in Germany quite some money to be able to do so comes from us taxpayers). Still I think GEB is a positive example of an intellectual journey!\nWhy? Recently, you might have come across several news articles about AI solving this or that long-standing math problem, where in order to solve it or to verify the produced proof, automated theorem proving via Lean has been used by the AI or humans, respectively. Lean in turn might again feel like a mere intellectual game*, but in my opinion it has some very interesting real-world applications, especially related to security, which might become even more important in an age of uncontrolled, unleashed AI agents. The obstacle: Lean is complicated, even for me after having done theoretical physics since 2011. An example: Lean\u0026rsquo;s type system contains the concept of a hierarchy of infinitely many so-called universes. What are they? And why does Lean need it to achieve its aims?\nIn German there\u0026rsquo;s this idiom \u0026ldquo;mir geht ein Licht auf\u0026rdquo; (its closest English equivalent is \u0026ldquo;the penny dropped\u0026rdquo;), which might be literally translated as \u0026ldquo;a light has risen for me\u0026rdquo;\u0026mdash;in the case of understanding Lean\u0026rsquo;s universes, a light sparked by GEB, which I hadn\u0026rsquo;t anticipated at all before I started to read it.\n*(Still I had some fun with it and can recommend it; it reminded me of the teaching approach and the first 3 lectures of the series \u0026ldquo;Geometrical Anatomy of Theoretical Physics\u0026rdquo; by Frederic P. Schuller. Back then in 2013, both he and me were at the University of Erlangen, but in that semester I was an Erasmus student at the University of Lund, Sweden, and followed these lectures by watching the video recordings which I asked the university for, which was exceptional and felt like a novelty at that time.)\n","permalink":"https://d-a-reiss.github.io/writings/geb-and-lean/","summary":"\u003cp\u003eCurrently I\u0026rsquo;m reading the classic \u003ca href=\"https://en.wikipedia.org/wiki/G%C3%B6del,_Escher,_Bach\"\u003e\u0026quot;Gödel, Escher, Bach\u0026quot;\u003c/a\u003e (GEB) by Douglas R. Hofstadter, years ago recommended to me by my stepfather after he read it when he was studying computer science besides working as a math and physics teacher (in retrospect, him doing so has impressed and inspired me, until today).\nSince then, I had GEB on my bookshelf, but never read it because I always thought: \u0026ldquo;This is not the right time, because\u0026hellip;\u0026rdquo; \u0026ldquo;I\u0026rsquo;ve got too many other things on my mind\u0026rdquo;, \u0026ldquo;there\u0026rsquo;s so much other stuff to do\u0026rdquo;, and so on\u0026hellip;\n(To be honest, I may also have gotten a little bit intimidated by the intellectual level of the book as he described it, but as it turned out, I actually find it quite approachable.)\u003c/p\u003e","title":"First thoughts on \"Gödel, Escher, Bach\" and Lean"},{"content":"\u0026hellip;descriptions will follow.\n","permalink":"https://d-a-reiss.github.io/projects/project-placeholder/","summary":"\u0026hellip;summaries will follow.","title":"Further project descriptions'"},{"content":"I\u0026rsquo;m excited to announce that I\u0026rsquo;ll be able to attend the AI conference NVIDIA GTC Berlin 2026 on October 21–22, 2026 (which is part of NVIDIA\u0026rsquo;s global GPU Technology Conference series) thanks to my employer TNG Technology Consulting GmbH! It\u0026rsquo;s hard to tell which aspect of it excites me most, but probably: to get to know the people there.\nSo if you\u0026rsquo;ll be at the NVIDIA GTC Berlin as well, feel free to drop me a note, I\u0026rsquo;d be happy to meet you!\n","permalink":"https://d-a-reiss.github.io/writings/nvidia-gtc-berlin/","summary":"\u003cp\u003eI\u0026rsquo;m excited to announce that I\u0026rsquo;ll be able to attend the AI conference  \u003ca href=\"https://www.nvidia.com/en-eu/gtc/\"\u003eNVIDIA GTC Berlin 2026\u003c/a\u003e on October 21–22, 2026 (which is part of NVIDIA\u0026rsquo;s global GPU Technology Conference series) thanks to my employer \u003ca href=\"https://www.tngtech.com\"\u003eTNG Technology Consulting GmbH\u003c/a\u003e!\nIt\u0026rsquo;s hard to tell which aspect of it excites me most, but probably: to get to know the people there.\u003c/p\u003e\n\u003cp\u003eSo if you\u0026rsquo;ll be at the NVIDIA GTC Berlin as well, feel free to \u003ca href=\"https://de.linkedin.com/in/david-a-reiss-9582292a7\"\u003edrop me a note\u003c/a\u003e, I\u0026rsquo;d be happy to meet you!\u003c/p\u003e","title":"NVIDIA GTC Berlin 2026"},{"content":"I\u0026rsquo;m David, a theoretical physicist turned AI researcher, working as a software consultant for TNG Technology Consulting GmbH (formerly as a researcher at Freie Universität Berlin).\nLoving deep dives, both technical \u0026amp; personal, and spending time with my son. Passionate about science \u0026amp; technology, books \u0026amp; writing, art \u0026amp; theatre (great to live near Berlin!).\nThere would be so much more to say about me\u0026hellip; But these pages are mainly about my work and thoughts\u0026mdash;I hope you find them inspiring!\nI\u0026rsquo;m always looking for partners for intellectual discussions and research cooperations.\nFeel free to contact me via LinkedIn or email.\nMy current main personal research interests:\nAI interpretability and alignment reinforcement learning neural network operators numerical methods for partial and integro-differential equations (like the Boltzmann equation) ","permalink":"https://d-a-reiss.github.io/about/","summary":"\u003cp\u003eI\u0026rsquo;m David, a theoretical physicist turned AI researcher, working as a software consultant for \u003ca href=\"https://www.tngtech.com\"\u003eTNG Technology Consulting GmbH\u003c/a\u003e (formerly as a researcher at Freie Universität Berlin).\u003cbr\u003e\nLoving deep dives, both technical \u0026amp; personal, and spending time with my son.\nPassionate about science \u0026amp; technology, books \u0026amp; writing, art \u0026amp; theatre (great to live near Berlin!).\u003cbr\u003e\nThere would be so much more to say about me\u0026hellip; But these pages are mainly about my work and thoughts\u0026mdash;I hope you find them inspiring!\u003c/p\u003e","title":"About me"},{"content":"Diese Datenschutzerklärung informiert Sie gemäß Art. 13 DSGVO über die Verarbeitung personenbezogener Daten auf dieser Website.\n1. Verantwortlicher Verantwortlicher im Sinne von Art. 4 Nr. 7 DSGVO ist der Betreiber dieser Website:\nDavid A. Reiss\nZum Kahleberg 28\n14478 Potsdam\nDeutschland\nE-Mail: dav.reiss[at]t-online.de\n2. Grundsätzliches: Worauf diese Website bewusst verzichtet Die Website ist bewusst datenschutzfreundlich gebaut. Insbesondere gilt:\nKein Einsatz von Cookies. Es werden keine Cookies gesetzt und keine Cookies gelesen. Kein Tracking und keine Analyse. Es werden weder Analyse- noch Tracking-Dienste eingesetzt. Es werden keine Nutzungsprofile gebildet. Kein Speichern auf Ihrem Endgerät. Es werden keine Informationen in Cookies, im Local Storage, im Session Storage oder in vergleichbaren Browser-Speichern abgelegt oder ausgelesen. Ein Fingerprinting findet nicht statt. Keine eingebetteten Inhalte von Drittanbietern. Es werden keine iframes, keine Videoeinbindungen, keine Social-Media-Plugins, keine Zählpixel und keine Webfonts von Servern Dritter geladen. Sämtliche Skripte, Stile und Symbole stammen von dem Server dieser Domain. Keine Formulare. Es gibt kein Kontaktformular, keinen Newsletter und keine Kommentarfunktion. Lokale Suche. Die Suchfunktion läuft vollständig im Browser. Bei der Suche werden keine Daten an Dritte übertragen. Verweise auf andere Dienste (GitHub, LinkedIn, Google Scholar) und auf externe Publikationen (beispielsweise arXiv) sind einfache Hyperlinks. Beim Aufruf dieser Website werden an diese Dienste keine Daten übermittelt. Erst der Klick auf einen Verweis ruft das jeweilige Ziel auf; ab dann gilt die Datenschutzerklärung des jeweiligen Anbieters. Ausgehende Verweise enthalten das Attribut rel=\u0026quot;noopener noreferrer\u0026quot;, sodass die zuvor gesehene Seite (Referrer) nicht an das Ziel übergeben wird.\n3. Hosting und Server-Protokolle („Logfiles“) Die Website wird als statisches Angebot über GitHub Pages ausgeliefert. Dienstleister ist GitHub, Inc., 88 Colin P. Kelly Jr. Street, 94107 San Francisco, CA, USA.\nBeim Abruf der Website verarbeitet GitHub als technischer Dienstleister neben den Inhalten der Website insbesondere folgende Daten:\ndie IP-Adresse zugreifender Endgeräte, Datum und Uhrzeit des Zugriffs, die aufgerufene Seite oder Datei sowie die übertragene Datenmenge, die Adresse der zuvor besuchten Seite (Referrer), Browser und Browserversion sowie das Betriebssystem des Endgeräts, Statusinformationen zur Anfrage (beispielsweise erfolgreicher Abruf oder Fehlercode). Zweck und Rechtsgrundlage Zweck dieser Verarbeitung ist die Auslieferung und Darstellung der Website sowie die Gewährleistung ihrer Sicherheit, Stabilität und Funktionsfähigkeit (beispielsweise Fehleranalyse und Abwehr von Angriffen).\nRechtsgrundlage ist Art. 6 Abs. 1 Satz 1 Buchstabe f DSGVO. Der Betreiber verfolgt damit ein berechtigtes Interesse an einem technisch zuverlässigen und sicheren Betrieb der Website.\nSpeicherdauer Die Protokolldaten werden gemäß den von GitHub dokumentierten Löschkonzepten gelöscht, sobald sie für die genannten Zwecke nicht mehr erforderlich sind. Eine dauerhafte Speicherung findet nicht statt.\nAuftragsverarbeitung und Datenübermittlung in die USA GitHub, Inc. ist ein Auftragsverarbeiter im Sinne des Art. 28 DSGVO und verarbeitet die Daten ausschließlich nach den Weisungen des Betreibers. Der Vertrag zur Auftragsverarbeitung ist als GitHub Data Protection Addendum in den GitHub Customer Terms verankert.\nDie Website wird von Servern in den USA ausgeliefert; damit findet eine Übermittlung personenbezogener Daten in ein Drittland statt. Diese Übermittlung stützt sich auf:\ndie Zertifizierung von GitHub, Inc. unter dem EU-US Data Privacy Framework (Adäquanzbeschluss nach Art. 45 DSGVO; Teilnehmerliste unter https://www.dataprivacyframework.gov/list), und ergänzend Standardvertragsklauseln der EU-Kommission nach Art. 46 Abs. 2 Buchstabe c DSGVO, soweit die Zertifizierung im Einzelfall nicht greift. Weitere Informationen:\nDatenschutzerklärung von GitHub: https://docs.github.com/en/site-policy/privacy-policies/github-general-privacy-statement Data Protection Addendum von GitHub: https://docs.github.com/en/site-policy/privacy-policies/github-data-protection-addendum 4. Ihre Rechte als betroffene Person Sie haben das Recht auf:\nAuskunft (Art. 15 DSGVO), Berichtigung unrichtiger Daten (Art. 16 DSGVO), Löschung gespeicherter Daten (Art. 17 DSGVO), Einschränkung der Verarbeitung (Art. 18 DSGVO), Datenübertragbarkeit (Art. 20 DSGVO), Widerspruch gegen die Verarbeitung auf Grundlage berechtigter Interessen (Art. 21 DSGVO). Einwilligungen — falls künftig erteilt — können Sie jederzeit mit Wirkung für die Zukunft widerrufen (Art. 7 Abs. 3 DSGVO). Derzeit besteht keine Einwilligung.\nZur Geltendmachung der Rechte genügt eine formlose E-Mail an die im Abschnitt Verantwortlicher genannte Adresse.\nBeschwerderecht Unabhängig davon haben Sie das Recht, bei einer Datenschutz-Aufsichtsbehörde Beschwerde einzulegen (Art. 77 DSGVO). Die Beschwerde kann bei der Aufsichtsbehörde erhoben werden, zu deren Aufgabenbereich der Wohnsitz oder der Arbeitsplatz des Beschwerdeführers gehört oder der Ort des mutmaßlichen Verstoßes liegt.\nFür den Betreiber gilt: Die für ihn berufene Aufsichtsbehörde richtet sich nach dem Bundesland, in dem er seinen Wohnsitz hat. In diesem Fall ist das (siehe https://www.lda.brandenburg.de/lda/de/):\nDie Landesbeauftragte für den Datenschutz und für das Recht auf Akteneinsicht Brandenburg\nDagmar Hartge\nStahnsdorfer Damm 77\n14532 Kleinmachnow\n5. Datenschutzbeauftragter Der Betreiber ist zur Bestellung eines Datenschutzbeauftragten nicht verpflichtet; es wurde kein Datenschutzbeauftragter bestellt.\n6. Sicherheit Die Website wird ausschließlich verschlüsselt über HTTPS mit modernen TLS-Verfahren ausgeliefert (aktuell: TLS 1.3). Unverschlüsselte HTTP-Aufrufe werden automatisch zur verschlüsselten Adresse umgeleitet.\n7. Änderungen dieser Datenschutzerklärung Diese Datenschutzerklärung wird bei einer Änderung des Umfangs der Datenverarbeitung oder der rechtlichen Rahmenbedingungen angepasst.\nStand: 02.09.2026\n","permalink":"https://d-a-reiss.github.io/datenschutz/","summary":"\u003cp\u003eDiese Datenschutzerklärung informiert Sie gemäß Art. 13 DSGVO über die\nVerarbeitung personenbezogener Daten auf dieser Website.\u003c/p\u003e\n\u003ch3 id=\"1-verantwortlicher\"\u003e1. Verantwortlicher\u003c/h3\u003e\n\u003cp\u003eVerantwortlicher im Sinne von Art. 4 Nr. 7 DSGVO ist der Betreiber dieser\nWebsite:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDavid A. Reiss\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZum Kahleberg 28\u003cbr\u003e\n14478 Potsdam\u003cbr\u003e\nDeutschland\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eE-Mail:\u003c/strong\u003e dav.reiss[at]t-online.de\u003c/p\u003e\n\u003ch3 id=\"2-grundsätzliches-worauf-diese-website-bewusst-verzichtet\"\u003e2. Grundsätzliches: Worauf diese Website bewusst verzichtet\u003c/h3\u003e\n\u003cp\u003eDie Website ist bewusst datenschutzfreundlich gebaut. Insbesondere gilt:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eKein Einsatz von Cookies.\u003c/strong\u003e Es werden keine Cookies gesetzt und keine\nCookies gelesen.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKein Tracking und keine Analyse.\u003c/strong\u003e Es werden weder Analyse- noch\nTracking-Dienste eingesetzt. Es werden keine Nutzungsprofile gebildet.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKein Speichern auf Ihrem Endgerät.\u003c/strong\u003e Es werden keine Informationen in\nCookies, im Local Storage, im Session Storage oder in vergleichbaren\nBrowser-Speichern abgelegt oder ausgelesen. Ein Fingerprinting findet\nnicht statt.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKeine eingebetteten Inhalte von Drittanbietern.\u003c/strong\u003e Es werden keine\niframes, keine Videoeinbindungen, keine Social-Media-Plugins, keine\nZählpixel und keine Webfonts von Servern Dritter geladen. Sämtliche\nSkripte, Stile und Symbole stammen von dem Server dieser Domain.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKeine Formulare.\u003c/strong\u003e Es gibt kein Kontaktformular, keinen Newsletter und\nkeine Kommentarfunktion.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eLokale Suche.\u003c/strong\u003e Die Suchfunktion läuft vollständig im Browser. Bei der\nSuche werden keine Daten an Dritte übertragen.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eVerweise auf andere Dienste (GitHub, LinkedIn, Google Scholar) und auf\nexterne Publikationen (beispielsweise arXiv) sind einfache Hyperlinks. Beim\nAufruf dieser Website werden an diese Dienste keine Daten übermittelt. Erst\nder Klick auf einen Verweis ruft das jeweilige Ziel auf; ab dann gilt\ndie Datenschutzerklärung des jeweiligen Anbieters. Ausgehende Verweise\nenthalten das Attribut \u003ccode\u003erel=\u0026quot;noopener noreferrer\u0026quot;\u003c/code\u003e, sodass die zuvor\ngesehene Seite (Referrer) nicht an das Ziel übergeben wird.\u003c/p\u003e","title":"Datenschutzerklärung"},{"content":"Angaben gemäß § 5 DDG Website-Betreiber / Diensteanbieter:\nDavid A. Reiss\nZum Kahleberg 28\n14478 Potsdam\nDeutschland\nE-Mail: dav.reiss[at]t-online.de\nVerantwortlich für den Inhalt nach § 18 Abs. 2 MStV David A. Reiss (Anschrift wie oben)\n(Hinweis: Diese Angabe betrifft journalistisch-redaktionelle Inhalte, z. B. die Bereiche Writings und Projects.)\nHinweise Dies ist eine persönliche und nicht-kommerzielle Website. Sie dient der Vorstellung von Person, Werdegang, Arbeiten und Veröffentlichungen; es besteht kein Geschäftsverkehr und es werden keine Waren oder Dienstleistungen angeboten. Auf dieser Website werden keine werblichen Inhalte vertrieben. Bei Fragen zur Website, zu den Inhalten oder zum Datenschutz wenden Sie sich bitte an die oben genannte Kontaktmöglichkeit. Hinweise zum Datenschutz finden Sie in der Datenschutzerklärung. ","permalink":"https://d-a-reiss.github.io/impressum/","summary":"\u003ch3 id=\"angaben-gemäß--5-ddg\"\u003eAngaben gemäß § 5 DDG\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003eWebsite-Betreiber / Diensteanbieter:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDavid A. Reiss\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZum Kahleberg 28\u003cbr\u003e\n14478 Potsdam\u003cbr\u003e\nDeutschland\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eE-Mail:\u003c/strong\u003e dav.reiss[at]t-online.de\u003c/p\u003e\n\u003ch3 id=\"verantwortlich-für-den-inhalt-nach--18-abs-2-mstv\"\u003eVerantwortlich für den Inhalt nach § 18 Abs. 2 MStV\u003c/h3\u003e\n\u003cp\u003eDavid A. Reiss (Anschrift wie oben)\u003c/p\u003e\n\u003cp\u003e(Hinweis: Diese Angabe betrifft journalistisch-redaktionelle Inhalte, z. B.\ndie Bereiche \u003ca href=\"https://d-a-reiss.github.io/writings/\"\u003eWritings\u003c/a\u003e und \u003ca href=\"https://d-a-reiss.github.io/projects/\"\u003eProjects\u003c/a\u003e.)\u003c/p\u003e\n\u003ch3 id=\"hinweise\"\u003eHinweise\u003c/h3\u003e\n\u003cul\u003e\n\u003cli\u003eDies ist eine persönliche und nicht-kommerzielle Website. Sie dient der\nVorstellung von Person, Werdegang, Arbeiten und Veröffentlichungen; es\nbesteht kein Geschäftsverkehr und es werden keine Waren oder\nDienstleistungen angeboten.\u003c/li\u003e\n\u003cli\u003eAuf dieser Website werden keine werblichen Inhalte vertrieben.\u003c/li\u003e\n\u003cli\u003eBei Fragen zur Website, zu den Inhalten oder zum Datenschutz wenden Sie sich\nbitte an die oben genannte Kontaktmöglichkeit. Hinweise zum Datenschutz\nfinden Sie in der \u003ca href=\"/datenschutz/\"\u003eDatenschutzerklärung\u003c/a\u003e.\u003c/li\u003e\n\u003c/ul\u003e","title":"Impressum"},{"content":"In several iterations of this workshop at TNG (more are planned for the future), my TNG colleagues Johannes Flommersfeld, Jonathan du Mesnil de Rochement, Daniel Hirsch, and me gave a theoretical introduction to the concepts behind reinforcement learning (RL) and contrasted that with implementing and training own RL agents with Stable Baselines3 and Gymnasium. By the end of the workshop, participants had a solid understanding of the theoretical concepts underlying RL and the practical skills to implement an RL training suite in real-world applications.\n","permalink":"https://d-a-reiss.github.io/talks/workshops-rl-fundamentals/","summary":"In these workshops, TNG colleagues gained a foundational understanding of reinforcement learning.","title":"Workshop series \"Reinforcement learning fundamentals\""},{"content":"Way beyond this X-post, the R1T and R1T2 Chimera models published by TNG gained severe attention with a daily usage of more than 10 billion tokens via OpenRouter. So, what kind of \u0026quot;Frankensteining\u0026quot; is going on there? How can a small software consultancy such as TNG produce its own models?\nAt TNG, we, an internal AI research group, have been experimenting and publishing results with a focus on Mixture-of-Expert Large Language Models and adaptions of the DeepSeek model family. First, we manipulated the way experts work within a model under the name \u0026quot;Mixture of Tunable Experts\u0026quot;. We then continued with the \u0026quot;Assembly-of-Experts\u0026quot; merging process resulting in the Chimera models.\nThis talk by my TNG colleague Fabian Klemm and me at TNG\u0026rsquo;s Big Techday on October 24, 2025, aimed to give some technical insights into the work of TNG AI research. We recalled the theoretical basics and the most important results. Furthermore, we provided technical details on how things were done as well as some anecdotes about the successes and losses along this journey.\nYou can find a video recording of the talk on YouTube.\n","permalink":"https://d-a-reiss.github.io/talks/talk-big-techday-25/","summary":"\u0026ldquo;The Germans just Frankensteined DeepSeek\u0026rsquo;s R1 and V3 into something called R1T Chimera\u0026rdquo;.","title":"Talk \"Tales from the machinery room - Customizing LLMs @ TNG AI research\""},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2025-klagges-assembly-of-experts/","summary":"Requiring 10^13-10^15 FLOPs to calculate one 8 bit weight in an LLM during pretraining is extremely expensive and seems inefficient. To better leverage the huge investments made into pretrained models, we develop the new \u0026ldquo;Assembly-of-Experts\u0026rdquo; (AoE) construction method to create capable child variants of existing Mixture-of-Experts parent models in linear time. Model weight tensors get interpolated individually, allowing to enhance or suppress semantic features of the parents. Varying the proportion of weights taken from the parent models, we observe some properties of the AoE child model changing gradually, while other behavioral traits emerge with a sharp transition. Surprisingly, nearly every generated model is functional and capable, which makes searching the model space straightforward. We construct the DeepSeek R1T \u0026ldquo;Chimera\u0026rdquo;, a 671B open-weights hybrid model combining DeepSeek\u0026rsquo;s V3-0324 and R1 model variants. The child inherits only the routed expert tensors of R1, but still achieves about R1-level intelligence. At the same time, it uses about 40% fewer output tokens, close to V3 speed. Constructed without any fine-tuning or distillation, the Chimera exhibits surprisingly compact, orderly reasoning compared to its parent models. The DeepSeek R1T Chimera model is available as open weights on \u003ca href=\"https://huggingface.co/tngtech/DeepSeek-R1T-Chimera\"\u003eHugging Face\u003c/a\u003e.","title":"Assembly of Experts: Linear-time construction of the Chimera LLM variants with emergent and adaptable behaviors"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2025-wilms-qrl-rare-dynamics/","summary":"Rare events are essential for understanding the behavior of non-equilibrium and industrial systems. It is of ongoing interest to develop methods for effectively searching for rare events. With the advent of quantum computing and its potential advantages over classical computing for applications like sampling certain probability distributions, the question arises whether quantum computers could also provide an advantage or inspire new methods for sampling the statistics of rare events. In this work, we propose a quantum reinforcement learning (QRL) method for studying rare dynamics, and we investigate their benefits over classical approaches based on neural networks. As a proof-of-concept example, we demonstrate that our QRL agents can learn and generate the rare dynamics of random walks, and we are able to explain this success as well as the different contributing factors to it via the intrinsic Fourier features of the parameterized quantum circuit. Furthermore, we show better learning behavior with fewer parameters compared to classical approaches. This is the first investigation of QRL applied to generating rare events and suggests that QRL is a promising method to study their dynamics and statistics.","title":"Quantum reinforcement learning of classical rare dynamics: Enhancement by intrinsic Fourier features"},{"content":"\u0026hellip;will follow soon.\n","permalink":"https://d-a-reiss.github.io/talks/poster-presentations-academia/","summary":"\u0026hellip;will follow soon.","title":"Past poster presentations in academia"},{"content":"\u0026hellip;will follow soon.\n","permalink":"https://d-a-reiss.github.io/talks/talks-academia/","summary":"\u0026hellip;will follow soon.","title":"Past talks in academia"},{"content":"\u0026hellip;will follow soon.\n","permalink":"https://d-a-reiss.github.io/talks/teaching-academia/","summary":"\u0026hellip;will follow soon.","title":"Past teaching in academia"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2024-franke-ac-magnetoelectric-transport/","summary":"(Note that cited here as the arXiv v1 preprint. The content was later reorganized (after disagreements about this I decided to be not included in the list of authors anymore) and published as \u003ca href=\"https://doi.org/10.1103/zpts-44w3\"\u003ePhys. Rev. B 113, 014419 (2026)\u003c/a\u003e.) Electron-magnon coupling at the interface between a normal metal and a magnetically ordered insulator modifies the electrical conductivity of the normal metal, an effect known as spin-Hall magnetoresistance. It can also facilitate magnon-mediated current drag, the nonlocal electric current response of two normal metal layers separated by a magnetic insulator. In this article, we present a theory of these two spintronic effects and their nonlinear counterparts for time-dependent applied electric fields $E(\\omega)$, with driving frequencies $\\omega$ up to the THz regime. We compare various mechanisms leading to a quadratic-in-$E$ response - Joule heating, spin-torque diode effect, and magnonic unidirectional spin-Hall magnetoresistance - and show how these can be disentangled by their characteristic dependences on $\\omega$ and the magnetization direction.","title":"Theory of ac magnetoelectric transport in normal-metal - magnetic-insulator heterostructures"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2022-reiss-finite-frequency-spin-conductance/","summary":"The interface between a ferro-/ferrimagnetic insulator and a normal metal can support spin currents polarized collinear with and perpendicular to the magnetization direction. The flow of angular momentum perpendicular to the magnetization direction (\u0026ldquo;transverse\u0026rdquo; spin current) takes place via spin torque and spin pumping. The flow of angular momentum collinear with the magnetization (\u0026ldquo;longitudinal\u0026rdquo; spin current) requires the excitation of magnons. In this article we extend the existing theory of longitudinal spin transport in the zero-frequency weak-coupling limit in two directions. We calculate the longitudinal spin conductance non-perturbatively (but in the low-frequency limit) and at finite frequency (but in the limit of low interface transparency). For the paradigmatic spintronic material system YIG|Pt, we find that non-perturbative effects lead to a longitudinal spin conductance that is ca. 40% smaller than the perturbative limit, whereas finite-frequency corrections are relevant at low temperatures \u0026lt; 100 K only, when only few magnon modes are thermally occupied.","title":"Finite-frequency spin conductance of the interface between a ferro- or ferrimagnetic insulator and a normal metal"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2021-chekhov-ultrafast-demagnetization-iron/","summary":"We study ultrafast magnetization quenching of ferromagnetic iron following excitation by an optical vs a terahertz pump pulse. While the optical pump (photon energy of 3.1 eV) induces a strongly nonthermal electron distribution, terahertz excitation (~4 meV) results in a quasi-thermal perturbation of the electron population. The pump-induced spin and electron dynamics are interrogated by the magneto-optic Kerr effect (MOKE). A deconvolution procedure allows us to push the time resolution down to 130 fs, even though the driving terahertz pulse is more than 0.5 ps long. Remarkably, the MOKE signals exhibit an almost identical time evolution for both optical and terahertz pump pulses, despite the three orders of magnitude different number of excited electrons. We are able to quantitatively explain our results using a model based on quasi-elastic spin-flip scattering. It shows that in the small-perturbation limit, the rate of demagnetization of a metallic ferromagnet is proportional to the excess energy of the electrons, independent of the precise shape of their distribution. Our results reveal that the dynamics of ultrafast demagnetization and of the closely related terahertz spin transport do not depend on the pump photon energy.","title":"Ultrafast demagnetization of iron induced by optical versus terahertz pulses"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2021-reiss-spin-hall-magnetoresistance-thz/","summary":"In bilayers consisting of a normal metal (N) with spin-orbit coupling and a ferromagnet (F), the combination of the spin-Hall effect, the spin-transfer torque, and the inverse spin-Hall effect gives a small correction to the in-plane conductivity of N, which is referred to as spin-Hall magnetoresistance (SMR). We here present a theory of the SMR and the associated off-diagonal conductivity corrections for frequencies up to the terahertz regime. We show that the SMR signal has pronounced singularities at the spin-wave frequencies of F, which identifies it as a potential tool for all-electric spectroscopy of magnon modes. A systematic change of the magnitude of the SMR at lower frequencies is associated with the onset of a longitudinal magnonic contribution to spin transport across the F-N interface.","title":"Theory of spin-Hall magnetoresistance in the ac terahertz regime"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2022-rouzegar-thz-spin-transport/","summary":"Laser-induced terahertz spin transport (TST) and ultrafast demagnetization (UDM) are central but so far disconnected phenomena in femtomagnetism and terahertz spintronics. Here, we use broadband terahertz emission spectroscopy to reliably measure both processes in one setup. We find that the rate of UDM of a single ferromagnetic metal film F has the same time evolution as the flux of TST from F into an adjacent normal-metal layer N. This remarkable agreement shows that UDM and TST are driven by the same force, which is fully determined by the state of the ferromagnet. An analytical model consistently and quantitatively explains our observations. It reveals that both UDM in F and TST in the F|N stack arise from a generalized spin voltage, which is defined for arbitrary, nonthermal electron distributions. We also conclude that contributions due to a possible temperature difference between F and N are minor and that the spin-current amplitude can, in principle, be increased by one order of magnitude. In general, our findings allow one to apply the vast knowledge of UDM to TST, thereby opening up new pathways toward large-amplitude terahertz spin currents and, thus, energy-efficient ultrafast spintronic devices.","title":"Laser-induced terahertz spin transport in magnetic nanostructures arises from the same force as ultrafast demagnetization"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2020-averchenko-time-bandwidth-product/","summary":"Two complementary properties, such as frequency and duration of a pulse, cannot be precisely measured simultaneously. The lower bound on the time-bandwidth product, known also as uncertainty relation, plays a very important role in quantum theory. In this work, we consider single-photon pulses with arbitrary temporal and spectral profiles, and derive a lower bound for the single-photon time-bandwidth product.","title":"Lower bounds for the time-bandwidth product of a single-photon pulse"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2020-muhlhauser-fracton-phases/","summary":"The quantum robustness of fracton phases is investigated by studying the influence of quantum fluctuations on the X-Cube model and Haah\u0026rsquo;s code, which realize a type-I and type-II fracton phase, respectively. To this end a finite uniform magnetic field is applied to induce quantum fluctuations in the fracton phase resulting in zero-temperature phase transitions between fracton phases and polarized phases. Using high-order series expansions and a variational approach, all phase transitions are classified as strongly first order, which turns out to be a consequence of the (partial) immobility of fracton excitations. Indeed, single fractons as well as few-fracton composites can hardly lower their excitation energy by delocalization due to the intriguing properties of fracton phases, as demonstrated in this work explicitly in terms of fracton quasi-particles.","title":"Quantum robustness of fracton phases"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2019-reiss-toric-code-field/","summary":"We study the robustness of 3D intrinsic topological order under external perturbations by investigating the paradigmatic microscopic model, the 3D toric code in an external magnetic field. Exact dualities as well as variational calculations reveal a ground-state phase diagram with first and second-order quantum phase transitions. The variational approach can be applied without further approximations only for certain field directions. In the general field case, an approximative scheme based on an expansion of the variational energy in orders of the variational parameters is developed. For the breakdown of the 3D intrinsic topological order, it is found that the (im-)mobility of the quasiparticle excitations is crucial in contrast to their fractional statistics.","title":"Quantum robustness and phase transitions of the 3D toric code in a field"},{"content":"","permalink":"https://d-a-reiss.github.io/publications/2017-adelhardt-haldane-phases/","summary":"We use linked-cluster expansions to analyze the quantum phase transitions between symmetry unbroken trivial and topological Haldane phases in two different spin-one chains. The first model is the spin-one Heisenberg chain in the presence of a single-ion anisotropy while the second one is the dimerized spin-one Heisenberg chain. For both models we determine the ground-state energy and the one-particle gap inside the non-topological phase as a high-order series using perturbative continuous unitary transformations. Extrapolations of the gap series are applied to locate the quantum critical point and to extract the associated critical exponent. We find that this approach works unsatisfactory for the anisotropic chain, since the quality of the extrapolation appears insufficient due to the large correlation length exponent. In contrast, extrapolation schemes display very good convergence for the gap closing in the case of the dimerized spin-one Heisenberg chain.","title":"Quantum phase transitions to topological Haldane phases in spin-one chains studied by linked-cluster expansions"}]