I designed my own fast game streaming video codec – PyroWave
PyroWave is a custom video codec designed for ultra-low latency game streaming, achieving encoding times as low as 0.13 ms on a RX 9070 XT, significantly outperforming traditional codecs like H.264 and HEVC.
My 2.5 year old laptop can write Space Invaders in JavaScript now (GLM-4.5 Air)
A 2.5-year-old laptop can now run the GLM-4.5 Air model, which boasts 106 billion parameters, enabling it to generate functional code like a Space Invaders game in JavaScript with no edits required.
Built RL for long-horizon agents – tested on 32x H100s but too poor to train
Terminal-Bench-RL achieves a significant milestone by developing a stable RL training infrastructure that utilizes 32x H100 GPUs across 4 nodes, enabling the training of long-horizon terminal-based coding agents, specifically the Qwen3-32B, which ranks as the highest scoring Qwen3 agent on the terminal-bench leaderboard without actual training.
Anthropic faces a class action lawsuit that could impose damages exceeding $1 billion due to its use of pirated books for AI model training, marking a significant legal precedent in the generative AI sector.
2025 Stack Overflow Developer Survey Results
The 2025 Stack Overflow Developer Survey reveals that over 49,000 developers from 177 countries participated, highlighting a significant focus on AI tools and community platforms, with 45% of respondents coding for less than 10 years.
SRAM Has No Chill: Exploiting Power Domain Separation to Steal On-Chip Secrets
Volt Boot is a novel attack that exploits power domain separation in modern SoCs to retain data in on-chip SRAM across power cycles, achieving 100% accuracy in data retrieval without the need for low temperatures or complex post-processing.
[2507.19457] GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
GEPA (Genetic-Pareto) leverages natural language reflection to optimize prompts, enabling LLMs to learn high-level rules more effectively than traditional reinforcement learning methods like GRPO, which often require extensive rollouts.
Do variable names matter for AI code completion? (2025)
Descriptive variable names significantly enhance AI code completion, yielding a 34.2% exact match rate compared to just 16.6% for obfuscated names, indicating that clarity in naming aids model performance.
SmallThinker: A Family of Efficient Large Language Models Natively Trained for Local Deployment
SmallThinker represents a breakthrough in local deployment of large language models (LLMs), designed from the ground up to operate efficiently on devices with limited resources rather than merely compressing existing cloud models.
The Making of Dario Amodei
Dario Amodei, CEO of Anthropic, has transformed the company into a $61 billion powerhouse, emphasizing AI's rapid advancement and its potential to disrupt 50% of entry-level jobs while advocating for responsible development and regulation.
Supervised Fine Tuning on Curated Data is Reinforcement Learning
Supervised Fine Tuning (SFT) on curated data can be reinterpreted as a method for maximizing a lower bound on the Reinforcement Learning (RL) objective, enhancing its effectiveness in sparse reward scenarios.
[D] First research project – feedback on "Ano", a new optimizer designed for noisy deep RL (also looking for arXiv endorsement)
The proposed optimizer, Ano, aims to decouple gradient magnitude from momentum direction, enhancing stability and speed in noisy deep RL environments, which are prevalent in this field.
[R] Misuse of ML for a cortical pain biomarker?
The critique in JAMA Neurology highlights methodological flaws in a previously published ML-based pain biomarker, specifically citing an incorrect validation set and an unrepresentative test set.
State of the Art SISR [R]
Extreme single-image super-resolution (SISR) techniques can achieve magnification factors of up to 100x, leveraging domain-specific texture synthesis tailored for materials.
[R] Multi-View Contrastive Learning: Principled Framework for 3+ Views and Modalities
MV-InfoNCE and MV-DHEL are novel loss functions that enhance multi-view contrastive learning by effectively managing interactions among multiple views, addressing the limitations of existing methods like SwAV and DINO.