01 — NVIDIA Developer
NVIDIA uses AI agents to rewrite Apache Spark UDFs for 22x faster GPU acceleration.
NVIDIA highlights how AI agents can automatically rewrite user-defined functions (UDFs) in Apache Spark jobs into GPU-accelerated equivalents using their cuDF plugin, achieving over 22x speedups. The approach includes converting UDFs into Spark SQL or RapidsUDF forms and, for complex cases, custom CUDA kernels. A major company applied this workflow to 30+ UDFs, reducing fallback overhead and accelerating thousands of jobs while preserving correctness.
Why it matters
Rewriting Spark UDFs for GPU acceleration traditionally demands heavy engineering, but AI agents streamline this, making significant speedups more accessible and reliable.
Read the source →02 — Hugging Face
HiRO-ACEv2 simulates global 3 km weather data at 1/300th the cost of traditional models.
HiRO-ACEv2, an update to the HiRO-ACE climate emulator, accurately simulates extreme rain, wind, and pressure events with fine 3 km resolution, closely matching the high-resolution X-SHiELD atmospheric model while reducing computational costs significantly. Combining AI models ACE and HiROv2, it produces year-long global datasets, capturing rare extreme weather and hurricanes realistically, though some wind and pressure asymmetries remain near storm eyes. Future plans include expanding climate scenarios and adding temperature and snow data outputs.
Why it matters
HiRO-ACEv2 offers researchers an affordable, flexible tool for long-term high-resolution extreme weather simulations, potentially accelerating climate study and disaster preparedness.
Read the source →03 — Apple ML Research
Apple ML Research introduces the first tri-modal Masked Diffusion Models pretrained from scratch.
Apple ML Research has developed tri-modal Masked Diffusion Models (MDMs) pretrained on text, image-text, and audio-text data, departing from prior bimodal approaches. Their study analyzes multimodal scaling, modality mixing, noise schedules, batch-size effects, and offers optimized sampling defaults. They present a novel stochastic differential equation reparameterization to separate physical and logical batch sizes, improving training efficiency. A 3-billion parameter model pretrained on 6.4 trillion tokens demonstrates strong text generation, text-to-image, and text-to-speech capabilities.
Why it matters
This systematic, large-scale research advances understanding of multimodal discrete diffusion models, informing future AI model design and optimization across multiple modalities.
Read the source →04 — Hugging Face
LFM2.5-24B-A2B runs fast, private AI agents on-device, beating larger MoE models on tool-use tasks.
Liquid AI's LFM2.5-24B-A2B is a 24 billion parameter MoE-based reasoning agent with only 2B active parameters per token. It supports extremely long context windows (131K tokens), doubles the token vocabulary, and outperforms larger models on five of eight tool-use benchmarks and all instruction-following tests. It achieves inference speeds up to 184 tokens/s on Apple M5 Max and 21K tokens/s on NVIDIA B300 GPU, with open-weight availability on Hugging Face for on-device deployment and fine-tuning.
Why it matters
LFM2.5-24B-A2B enables fast, private, and efficient on-device AI agents capable of complex reasoning and tool use without relying on cloud services.
Read the source →05 — MIT AI
New book chronicles rise and decline of tech worker labor movement since 2017.
MIT PhD student JS Tan and Clarissa Redwine document how tech workers organized to oppose policies and contract work they found objectionable, initially achieving successes like Google's 2018 Project Maven contract cancellation. However, market downturns, layoffs, and AI-driven deskilling since 2022 have weakened their leverage, leading to a current state of organizing paralysis. Their 2026 book "Against Tech Oligarchy" analyzes these shifts and the political economy of tech labor.
Why it matters
Understanding the dynamics of tech worker organization reveals challenges in labor influence amidst industry changes, impacting technology's social and political impact.
Read the source →