Dylan Castillo used Claude and BigQuery to confirm arXiv paper share on HN has dropped sharply in recent months after peaking during the LLM wave.
Key Takeaways
The 2019 HN arXiv peak was deep learning: 41% of the top 100 upvoted papers were DL-focused, led by MuZero, EfficientNet, and XLNet.
2023-2026 was even more concentrated: 59% of top 100 upvoted arXiv papers were LLM or AI topics.
DeepSeek-R1 (1,351 pts) and BitNet b1.58 (1,040 pts) are flagged as likely durable papers; DeepSeek-R1 for open RL-based reasoning, BitNet for low-bit inference.
The LK-99 room-temperature superconductor cluster (2,408 + 1,690 pts) is called a landmark in meta-science: open-science at wire speed and crowdsourced replication, not physics.
Generative Agents “Smallville” paper (391 pts) and Differential Transformer (562 pts) round out the predicted durable set for architecture and agent design patterns.
