Deep Learning

9 Min Read

onApril 24, 2026

Alibaba’s Qwen Hits 700 Million Downloads on Hugging Face—Overtakes Meta’s Llama as World’s Most Popular Open-Source AI Model

A Chinese AI model just quietly dethroned Meta’s Llama as the world’s most downloaded open-source system—and December’s…

8 Min Read

Artur Markus

onApril 17, 2026

Anthropic’s Constitutional Classifiers++ Cut Jailbreak Success Rate from 86% to 4.4%—Only 1 Universal Jailbreak Found in Bug Bounty Testing

Anthropic just compressed what should have been years of AI safety progress into one architecture update—blocking 95% of jailbreaks for 1%…

9 Min Read

Artur Markus

onApril 15, 2026

Meta Llama 4 Launches April 5 with 109B-Parameter Scout Model—17B Active Parameters Fit on Single H100, Outperform GPT-4.5 and Claude 3.7 on STEM Benchmarks

Meta just made frontier AI fit on one GPU—and simultaneously banned 450 million Europeans from using it.

10 Min Read

Artur Markus

onApril 11, 2026

OpenAI o3-mini Launches January 31 with 95% Cost Cut Over GPT-4—First Small Reasoning Model with Function Calling Outperforms o1 on Coding Benchmarks

OpenAI’s cheapest reasoning model now beats its most expensive one at writing code. That sentence should make every CTO reconsider their…

11 Min Read

Artur Markus

onApril 9, 2026

Anthropic’s Claude Mythos Hits 93.9% SWE-Bench but Won’t Be Released—$25/$125 Token Pricing Reserved for 40+ Whitelisted Security Teams

Anthropic just built an AI that solves 94% of real-world software engineering tasks—then decided nobody outside 40 security teams gets to use…

9 Min Read

Artur Markus

onApril 8, 2026

Ricursive Intelligence Raises $300M Series A at $4B Valuation for AI-Driven Chip Design—Ex-DeepMind Founders Build Platform That Automates Semiconductor Architecture

Two researchers just raised more than most companies exit for—to build AI that designs the chips that run AI. This is either the most circular…

9 Min Read

Artur Markus

onApril 6, 2026

Liquid AI Releases LFM2.5-1.2B-Thinking: 1.2B Parameter Reasoning Model Runs Fully Offline in 900 MB on Phones

A reasoning model smaller than a mobile game just outscored models twice its size—and it runs entirely on your phone with no internet…

9 Min Read

Artur Markus

onApril 5, 2026

Rocket Software Launches EVA AI Assistant on January 27—Traces Mainframe Issues From Symptoms to Code Lines Using Model Context Protocol

The AI industry spent 2025 building chatbots for email summarization while ignoring the infrastructure processing $3 trillion in daily…

9 Min Read

Artur Markus

onApril 3, 2026

Microsoft Maia 200 Cuts AI Token Costs 30% with 140 Billion Transistors—3nm Chip Deployed January 27 in US Data Centers, Outperforms Amazon Trainium and Google TPU on Inference

Microsoft just made your AI inference bill 30% cheaper—not through clever prompt engineering or model distillation, but by building the most…

11 Min Read

Artur Markus

onMarch 26, 2026

Moonshot AI’s Attention Residuals (AttnRes) Achieves Same Training Loss with 1.25× Less Compute—New Transformer Architecture Replaces Fixed Residuals with Softmax Attention

Residual connections have been sacred architecture for nine years—and Moonshot AI just proved they were holding us back. Their new AttnRes…

Deep Learning

Subscribe to my Blog