AI Research

Abacus.AI released three open-weight Smaug models — Smaug Flash on DeepSeek V4 Flash, Smaug Mini on Qwen3.8 27B, and Smaug Agentic on Kimi K3 — tuned for long-running agent loops and published on Hugging Face. Smaug Flas…

Researchers introduced Simple Attention Sparsification (SAS), a gated sparse attention mechanism that optimizes context ranking end-to-end with the language modeling loss by injecting the selector's continuous scores int…

A new evaluation called Duplex Cue finds that full-duplex voice agents adapt to listener contributions mid-turn far less often than humans, according to a single-model case study of 300 human-confirmed cues from unscript…

A technical handbook on FlashAttention explains that the algorithm achieves wall-clock speedups by reducing data movement between GPU high-bandwidth memory (HBM) and on-chip SRAM rather than by approximating attention, u…

Anthropic CEO Dario Amodei, OpenAI CEO Sam Altman, Elon Musk and Google DeepMind CEO Demis Hassabis all endorsed slowing the pace of frontier AI development on Saturday, hours after Amodei published a proposal titled "We…

Researchers from ByteDance, Tsinghua University and the Shanghai Artificial Intelligence Laboratory published a joint paper on Thursday titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement", out…

Nvidia added CUDA-Q Logical to its open-source CUDA-Q platform, giving researchers and developers a toolkit for working with logical qubits and quantum error correction. On December 10, 2024, quantum hardware company Inf…

A technical overview of co-packaged optics (CPO) for scale-up AI compute details the packaging architectures, photonic integrated circuit (PIC) materials, laser options, and electrical IC (EIC) designs needed for mid-dis…

An engineer running two small LLM judges — "brief" for AI feeds and "scout" for Reddit threads — documented five evaluation exercises showing that three runs of the same 21-case brief suite on September 3 with identical …

A blog post argues that large language models have exactly zero mathematical capabilities, rejecting a claim attributed to twenty-five Fields Medalists that LLMs can now solve major outstanding problems in many fields of…

A prospective PhD student planning to study Graph ML, spanning graph signal processing and graph deep learning, is seeking advice on branding their doctoral research to land a research scientist role at big tech or a com…

NVIDIA announced an expansion of its open source CUDA-Q platform with CUDA-Q Logical, an orchestration layer for designing and testing fault-tolerant quantum computing applications. Fermilab used CUDA-Q Logical to cut fa…

Cognichip unveiled its AI Chip Intelligence (ACI) system, which uses a physics-informed AI approach to speed chip design, according to Forbes contributor Karl Freund. The company claims the system could change how chips …

Nvidia announced an expansion of its open source CUDA-Q platform for fault-tolerant quantum computing, launching CUDA-Q Logical, an orchestration layer for developing applications on logical qubits. Fermilab used CUDA-Q …

Jacob Coxon resigned from Anthropic on September 8, about two months before his equity would have vested, and publicly accused Anthropic and OpenAI of "racing straight to self-improving superintelligence" and "gambling w…

A developer released a from-scratch Sudoku solver built on a Looped MLP-Mixer architecture, trained on the sapientinc/sudoku-extreme dataset and runnable on an NVIDIA Jetson Nano via JetPack Docker. The model reuses a sh…

Model distillation, a technique in which a smaller "student" model learns to replicate the behavior of a much larger "teacher" model, has become one of the most contested topics in the AI industry as unauthorized distill…

An AI system trained on more than 20,000 protein structures from pharmaceutical companies outperforms AlphaFold-like models that rely only on public data, according to a Nature report by Ewen Callaway published September…

Researchers introduced DataFlex-RL, an evaluation platform for comparing data policies in reinforcement learning with verifiable rewards (RLVR) under a common GRPO recipe. The platform targets how RLVR data policies deci…

Elon Musk announced on September 14 that xAI's Grok 4.8, a 2.5 trillion-parameter AI model, is finishing its initial training run this week and entering the reinforcement learning phase, following Grok 4.6's August 12 re…