LLMs & Models
Large language models: GPT, Claude, Gemini, Mistral and open weights.
34 published articles
Open Weights Analysis
Gemma 4 is not a chatbot, and that's the point
Google DeepMind's Gemma 4 is an open-weight model designed for self-hosting and customization, not consumer chat. This analysis compares it to ChatGPT, Claude, and Qwen-3.5 across licensing, privacy, and deployment flexibility, revealing why it matters for regulated industries and on-device AI.
2026-07-10
AI Inference Acceleration
JetSpec Breaks the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting
JetSpec trains a causal parallel draft head over fused hidden states from the target model, producing candidate trees aligned with autoregressive factorization. It consistently outperforms baselines across dense and MoE Qwen3 models on math, coding, and chat tasks.
2026-07-10
Artificial Intelligence
Nvidia's new audio model does five jobs at once and beats the specialists at their own game
Nvidia's Audex unifies audio understanding, generation, and text reasoning in a single model, matching or beating task-specific systems on speech and audio benchmarks without sacrificing text performance.
2026-07-09
Alibaba Cloud
Alibaba Cloud's EMR Serverless Spark now processes images and video in plain SQL, no Python needed
Alibaba Cloud’s EMR Serverless Spark now supports images and video frames directly in SQL, letting data engineers skip Python overhead. A case study on autonomous driving data preprocessing shows automated ETL pipelines powered by Qwen vision models replacing manual annotation.
2026-07-09
Coding Agents
Cognition's new coding agent scores near frontier results for pocket change
Cognition's SWE-1.7 coding model narrows the gap to frontier systems at a fraction of the cost, scoring 42.3% on FrontierCode and running at 1,000 tok/s. The model was trained on an improved reinforcement learning pipeline using Kimi K2.7 as the base.
2026-07-09
Software development
The hidden tax on vibe-coded projects that shows up when you least expect it
An in-depth analysis of the promise and peril of vibe coding: why AI-generated code accelerates prototyping but introduces security risks, performance limits, and technical debt that demand careful human oversight.
2026-07-09
Open-source AI
China's MiniMax just open-sourced a 1M-token model that outruns GPT-5.5 on real coding tasks
MiniMax M3 is the first Chinese open-source model to combine native multimodality, 1M-token context, and advanced agentic coding. Its structured approach to scale, including a new attention architecture called MSA, challenges the assumption that open models must trail behind proprietary systems.
2026-07-09
DeepSeek's DSpark framework rearms speculative decoding for high-concurrency serving
Semi-autoregressive decoding just broke the 85% speed barrier in production AI inference
A new speculative decoding framework from DeepSeek tackles the two bottlenecks that have limited parallel drafters: suffix decay and wasteful verification. DSpark achieves 60, 85% faster generation speeds in production by coupling a semi-autoregressive architecture with a confidence-scheduled scheduler that prunes low-value tokens before the target model verifies them.
2026-07-08
AI Model Release
MiniMax launches M2.7 model with strong software engineering and office productivity skills
MiniMax's M2.7 model delivers strong results in software engineering benchmarks and professional office tasks, with a 97% skill adherence rate on complex instructions and an ELO score of 1495 on GDPval-AA.
2026-07-08
AI research
Your AI agent passed by accident. SkillCoach grades the process, not the answer.
SkillCoach is a self-evolving rubric framework that evaluates and improves agentic skill-use by analyzing skill selection, following, composition, and reflection processes, providing better supervision than outcome-only metrics.
2026-07-06
Model Evaluation
Ai2's olmo-eval gives LLM developers a microscope for every checkpoint
Ai2's olmo-eval brings per-question diffs and modular benchmarks to active LLM development, helping researchers tell real progress from statistical noise.
2026-07-06
Deep Learning
M3D and Real-Guidance Bring Dataset Distillation to High-Resolution Realms
Dataset distillation has long been stuck on low-res benchmarks, but a new approach called M3D changes that. By combining multi-scale matching, a data manifold prior, and a Real-guidance strategy, it scales to ImageNet-1K at 128×128 resolution, achieving 68.5% top-1 accuracy with just one image per class and cutting memory usage by ten times.
2026-07-05