Category: Machine Learning

9 Min Read

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…

14 Min Read

Artur Markus

onApril 4, 2026

Google Gemma 4 Ranks #3 on Arena AI Leaderboard—31B Open Model Hits 85.2% MMLU Pro and 89.2% AIME 2026, Outperforming Models 20× Larger

A 31-billion parameter model now outperforms systems with 600B+ parameters on reasoning and math benchmarks. Google just made 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…

8 Min Read

Artur Markus

onFebruary 22, 2026

Google Gemini 3.1 Pro Scores 77.1% on ARC-AGI-2—2.5x Jump Over Predecessor in Single Generation

Google just doubled AI reasoning capability in 90 days while keeping the price identical. The assumption that frontier AI improves linearly…

10 Min Read

Artur Markus

onJanuary 15, 2026

TII’s Falcon-H1R 7B Outperforms 47B Models on Math Reasoning While Running on a 16GB Laptop

A 7-billion parameter model just scored 88.1% on AIME-24 math reasoning, crushing NVIDIA’s 47B Nemotron at 49.7%. The assumption that…

8 Min Read

Artur Markus

onJanuary 12, 2026

NVIDIA Alpamayo: 10B-Parameter VLA Model Reduces Autonomous Driving Validation Variance by 83% Across 310,895 Real-World Clips

NVIDIA just open-sourced a 10-billion-parameter autonomous driving brain trained on 1,700+ hours of real-world footage from 25 countries—and…

12 Min Read

Artur Markus

onJanuary 9, 2026

Inverse Scaling in Test-Time Compute: When More ML Reasoning Tokens Systematically Destroy Performance

The industry just spent billions convincing you that longer AI thinking equals better results. New research proves that’s…

12 Min Read

Artur Markus

onJanuary 8, 2026

The Compute-Optimal Blind Spot: Why Most ML Teams Are Still Wasting 40% of Training Budget on the Wrong Scaling Trade-Off

Your ML team probably burned through 40% of its training budget last quarter—and nobody noticed because the waste happened before a single GPU…

11 Min Read

Artur Markus

onDecember 27, 2025

The Critical Batch Size Paradox: Why Your GPU Cluster Is Wasting 40% of Its Compute on Inefficient Training Steps

Your infrastructure team just maxed out batch size on that H100 cluster. The GPU utilization graphs look beautiful. But you’re…

11 Min Read

Artur Markus

onDecember 6, 2025

The Test-Time Compute Paradox: Why Reasoning Models Like o1 and DeepSeek R1 Are Proving That More Inference Compute Can Destroy Accuracy

The entire AI industry just pivoted to “thinking longer” as the path to superintelligence—but January 2025 research reveals these…

Machine Learning

Subscribe to my Blog