Usman Salis

371 articles

Staff author

Usman Salis is an engineer by training and a long-time follower of innovation in the tech industry. Active in the crypto and blockchain space since 2011, he has contributed to hundreds of online and print media outlets worldwide.

Usman currently collaborates with several leading blockchain platforms and continues to write passionately about technology, helping bring blockchain and AI into the global conversation.

Contact us if you have any questions or remarks.

Usman Salis Usman Salis Industry
8 hours ago

ACT Trains AI Agents to Judge Their Own Decisions - With 5-Point Accuracy Gains

A new reinforcement learning framework teaches AI agents to evaluate their own choices, not just copy expert behavior - delivering measurable gains over existing baselines.

Usman Salis Usman Salis Industry
1 day ago

ReMix Fixes LoRA Routing Collapse with Reinforcement Learning

ReMix uses reinforcement-based routing to fix LoRA adapter collapse in MoE fine-tuning, boosting LLM efficiency.

Usman Salis Usman Salis AI models
1 day ago

Cursor Benchmarks GPT-5.4 Coding Efficiency: 60%+ Score on CursorBench

Cursor launched CursorBench to score AI models on agentic coding tasks, measuring both intelligence and token efficiency across 10+ leading models.

Usman Salis Usman Salis Industry
1 day ago

Meta Study: Reasoning LLM Judges Reach 89.6% Score, Yet Reward Hacking Persists

Meta and Yale researchers show reasoning-based judges cut reward hacking in RL alignment pipelines, but challenges remain in non-verifiable training settings.

Usman Salis Usman Salis AI models
1 day ago

NVIDIA's AutoGaze Cuts 99% of Video Tokens, Speeds Up AI Models 19x

Researchers from NVIDIA, MIT, UC Berkeley, and Clarifai built AutoGaze - an AI filter that slashes video data by up to 99% without hurting model accuracy.

Usman Salis Usman Salis Industry
2 days ago

Nemotron 3 Super Hits 484 Tokens/Sec with Multi-Token Prediction

Nemotron 3 Super tops AI speed benchmarks at 484 tokens/sec using multi-token prediction, outpacing dozens of rival models across providers.

Usman Salis Usman Salis Industry
2 days ago

OpenClaw-RL: 145K Stars Framework Trains AI Agents from Live Conversations

OpenClaw-RL trains AI agents from real-world interactions using async RL across 5 agent types.

Usman Salis Usman Salis AI models
3 days ago

xAI Grok 4.20 API Launches With 2M Token Context and 3 Model Variants

xAI quietly launched the Grok 4.20 API with three model variants and a 2-million token context window, opening large-context AI to developers at tiered pricing.

Usman Salis Usman Salis Industry
3 days ago

Stanford-Tsinghua VLAW Framework Lifts Robot Task Success by 39.2%

Stanford and Tsinghua's VLAW framework lifts robot task success 39.2% by merging real-world data with AI-generated simulations.

Usman Salis Usman Salis Benchmarks
4 days ago

ChatGPT Leads AI Subscription Race as Gemini and Claude Hit 258% Growth

ChatGPT holds 13-14M paid users while Gemini grows 258% and Claude surges 200% year-over-year in 2026.

Usman Salis Usman Salis Industry
4 days ago

AI Agent Memory: 4-Layer Infrastructure Now Drives Next-Gen Systems

From chat logs to lifecycle systems - why AI memory management is becoming the next major infrastructure challenge for developers.

Usman Salis Usman Salis AI models
4 days ago

PinchTab Launch: 13x Token Cut for AI Browser Agents

A 12MB open-source Go binary letting AI agents control Chrome via HTTP - cutting token usage by up to 13x through accessibility tree parsing.

Usman Salis Usman Salis Industry
5 days ago

2 Diverse AI Agents Beat 16 Identical Ones, New Study Finds

New research shows diverse AI agent teams outperform larger identical ones, with just 2 varied agents matching 16 clones.

Usman Salis Usman Salis Industry
6 days ago

LightMem Memory System Cuts LLM Token Usage by 117x and API Calls by 310x

Researchers introduced LightMem, a memory architecture designed to make large language models faster and more efficient. Early results show improved accuracy while dramatically reducing token usage and API calls.

Usman Salis Usman Salis Industry
6 days ago

PaddlePaddle FlashMaskV4 Delivers 2.9x Speed Boost for Long-Context AI Workloads

PaddlePaddle's FlashMaskV4 is a new attention masking framework built on FlashAttention-4 architecture. It targets transformer efficiency and flexible masking for large-scale, long-context AI workloads.

Posts feed
Study: #1-Ranked Google Brands Appear in 31% of AI Answers
5 hours ago
MIT Study Questions How LLMs Learn Reasoning: 2 Key Findings That Could Cut Training Costs
5 hours ago
ACT Trains AI Agents to Judge Their Own Decisions - With 5-Point Accuracy Gains
8 hours ago
Anthropic Researchers Launch $1B Startup Mirendil to Speed Up Scientific Discovery
8 hours ago
AI Humanoid Robots Move Toward Homes as Reflex Shows 80 kg Robot Handling Daily Tasks
8 hours ago
Claude Opus 4.6 Scores 78.3% on MRCR v2 With 1M Token Context Window
9 hours ago
Grok 4.20-Beta1 Claims #2 Spot on Search Arena With Score of 1225
9 hours ago
New JTok Scaling Method Boosts LLM Efficiency With 35% Less Compute
10 hours ago
ReMix Fixes LoRA Routing Collapse with Reinforcement Learning
1 day ago
Cursor Benchmarks GPT-5.4 Coding Efficiency: 60%+ Score on CursorBench
1 day ago
Follow us on Twitter

Disclaimer: Aigazine provides content for informational and educational purposes only. We do not offer financial, investment, or legal advice. Any decisions made based on the information published on this website are solely at the reader’s own discretion and risk. We encourage all readers to conduct their own research and seek professional guidance when necessary.

Aigazine is a news platform focused on artificial intelligence, covering global AI trends, technology, and innovation. Aigazine is based in Tbilisi (0179, Georgia, Tbilisi City, Vake District, 49 Besarion Zhghenti Street, VAT 305786600).

© 2025 aigazine.com

We use cookies to improve your experience on our site and to show you relevant advertising. To find our more, read our privacy policy and cookie policy