Yin Jun Phua, currently an assistant professor at Institute of Science Tokyo. Main research focus is on bridging the gap between symbolic AI and neural networks. https://yinjunphua.com/ 2026-07-03T00:00:00.000Z Yin Jun Phua https://yinjunphua.com/ https://yinjunphua.com/thoughts/2026-07-03-a-tabpfn-for-rules.html 2026-07-03T00:00:00.000Z 2026-07-03T00:00:00.000Z

NRI is a tabular foundation model in the TabPFN mold: pretrained once on synthetic tasks, applied zero-shot to datasets it has never seen. The difference is the output. Instead of a probability, NRI hands you a human-readable logical rule. How it works, what it costs, and when to use it.

https://yinjunphua.com/thoughts/2026-05-11-the-nri-research-diary.html 2026-05-11T00:00:00.000Z 2026-05-11T00:00:00.000Z

The story that didn't fit in the seven-page IJCAI 2026 paper on NRI, our neuro-symbolic foundation model for zero-shot logical rule induction: six months of chasing a bug that turned out to be the model doing exactly what we asked.

https://yinjunphua.com/thoughts/2025-12-08-participating-in-biohackathon-germany-2025.html 2025-12-08T00:00:00.000Z 2025-12-08T00:00:00.000Z

Notes from the 4th Biohackathon Germany, where I built NERVE — a Neuro-symbolic Evidence Review and Verification Engine that fact-checks biomedical LLM claims against ontologies and knowledge graphs through an MCP interface.

https://yinjunphua.com/thoughts/2024-01-18-cuda-unknown-error-with-podman.html 2024-01-18T00:00:00.000Z 2024-01-18T00:00:00.000Z

CUDA programs failed with a useless 'Unknown error' inside a rootless Podman container even though nvidia-smi worked. The fix: the missing /dev/nvidia-uvm device nodes, and how to create them by hand.

https://yinjunphua.com/thoughts/2023-11-29-agi-and-the-ever-moving-goalposts.html 2023-11-28T00:00:00.000Z 2023-11-28T00:00:00.000Z

For years, the advent of Artificial General Intelligence (AGI) has remained an elusive milestone. Despite the progress made with Large Language Models (LLMs), opinions are divided: some believe that AGI has finally been realized, while others strongly disagree.

https://yinjunphua.com/ja/thoughts/2023-11-14-ニューロシンボリックAI入門.html 2023-11-14T00:00:00.000Z 2023-11-14T00:00:00.000Z

深層機械学習モデルが優れた能力を発揮する一方で、隠れたバイアスや人間ではありえない誤りが生じる問題が明らかになっている。このため、解釈可能性や説明可能性の確保が一段と重要視されており、その結果、深層機械学習と記号推論の融合についての研究が近年注目を集めている。

https://yinjunphua.com/thoughts/2023-11-08-explainability-might-not-be-all-you-need.html 2023-11-08T00:00:00.000Z 2023-11-08T00:00:00.000Z

Interpretability and post-hoc XAI both fall short. What we really want from AI under delegation is predictable behaviour — the way we trust a well-trained shepherd dog whose cognition we can't interpret either — not just explanations.

https://yinjunphua.com/thoughts/2023-10-31-source-code-availability-is-not-reproducibility.html 2023-10-31T00:00:00.000Z 2023-10-31T00:00:00.000Z

Publicly available source code is not the same as reproducibility. On dependency rot, Python packaging defaults, the limits of Docker, and how fragile builds push AI research toward a narrow set of well-maintained models.

https://yinjunphua.com/thoughts/2023-10-11-bhyve-cpu-lockup-issue.html 2023-10-11T00:00:00.000Z 2023-10-11T00:00:00.000Z

Linux guests (Debian and Rocky) on bhyve/FreeBSD kept freezing with soft-lockup and clocksource watchdog errors. The real cause turned out to be insufficient entropy — enabling the virt_random option stopped the lockups.

https://yinjunphua.com/thoughts/2023-09-21-japan-seems-eager-to-get-on-the-generative-ai-hype-train.html 2023-09-21T00:00:00.000Z 2023-09-21T00:00:00.000Z

Japan, after enduring twenty years of stagnation and inflation, is keen to make a resurgence. The country is heavily investing in the promising field of generative AI, a technology that has recently garnered significant excitement.

https://yinjunphua.com/thoughts/2023-08-23-reflections-on-my-first-lecture.html 2023-08-23T00:00:00.000Z 2023-08-23T00:00:00.000Z

Reflections after teaching my first university programming course solo: steering away from LeetCode-style drills toward decomposing real problems, teaching debugging in VSCode, and finding the sweet spot between rigour and accessibility.