Ax Xinzi Cao, Jianyang Zhai, Pengfei Li, Zhiheng Hu, Cen Yan, Bingxu Mu, Guanghuan Fang, Bin She, Jiayu Li, Yihan Su, Dongyang Tao, Xiansong Huang, Fan Xu, Feidiao Yang, Yao Lu, Chang-Dong Wang, Yutong Lu, Weicheng Xue, Bin Zhou, Yonghong Tian 4/20/2026

AscendKernelGen: A Systematic Study of LLM-Based Kernel Generation for Neural Processing Units

AscendKernelGen uses LLMs to automate high-performance kernel generation for Neural Processing Units, addressing vendor-specific DSL expertise bottlenecks in AI hardware optimization.

Ax Kyle Waters, Lucas Nuzzi, Tadhg Looram, Alessandro Tomasiello, Ariel Ghislain Kemogne Kamdoum, Bikun Li, Damien Sileo, Egor Kretov, Francesco Fournier-Facio, Georgios Soloupis, Haile Kassahun, Hew Wolff, Jiaqi Cai, Lianghui Li, Marc Roth, Mohinder Naiya, Naixu Guo, Qicheng Tang, Richard Wheeler, Samuele Sala, Serguei Popov, Steven Dillmann, Yuqi Li 4/20/2026

COMPOSITE-Stem

COMPOSITE-STEM: 70 expert-written benchmark tasks in physics, biology, chemistry for evaluating AI agent reasoning on scientific discovery.

HN doener 4/19/2026

Swiss AI Initiative

Swiss government funding initiative for foundation model research and open science artifacts. €10M GPU hours available for core ML and applications.

HN rdslw 4/19/2026

Headless Everything for Personal AI

Essay on headless APIs for personal AI agents to interact with services (passports, hotels, banking, shopping). Practical agent applications.

HN simianwords 4/19/2026

The LLM costs are not going up

Analysis arguing LLM operational costs are sustainable and declining, countering common claims about economic viability.

HN utkarsh_apoorva 4/19/2026

The Missing Human Half of AI

AbleCredit case study on building human-centered AI systems for non-technical loan officers in India using WhatsApp/Instagram interfaces.

HN vignesh_146 4/19/2026

Model Council

Debate-style app orchestrating multiple LLMs (GPT, Grok, Claude) with scoring and verification

HN yoeven 4/19/2026

CNNs + VLM outperforms pure VLMs on OCR

Study comparing CNN-based OCR with vision language models, showing hybrid approach combining both architectures outperforms either alone.