MachinaLearning

Learn Machine Learning

Industry Ingest

Industry2026-11-07

Model Merging Goes Mainstream: Task Arithmetic and Model Soups Become Production Tooling

An emerging result moving into practice: open-source teams now build models by operating on weights directly instead of retraining. Task arithmetic (add or subtract fine-tuned weight deltas) and model soups (average fine-tunes of one base) ship as everyday tooling through mergekit and Hugging Face PEFT's merge methods, and merged checkpoints are common on the Hub. The research frontier, surgical weight editing (ROME / MEMIT) and weight-space meta-learning, is still early, and merging works best within one base-model family.

Jev and the Return of the Classifier: What TypeSafe's System One Model Means for Production AI

TypeSafe AI's Jev — a 'System One' model that returns typed decisions with probabilities instead of text — opened early access on September 15 and drew open-source replicas within days. We trace the eight-year lineage from BERT fine-tuning to universal classifiers, lay out six industry use cases, separate what is disclosed from what is inferred about Jev's architecture, and flag the vendor claims that still need independent verification.

by ML2026-07-08

Inside Claude's Mind: Anthropic Discovers J-Space, a Global Workspace in Language Models

Anthropic has found a small, privileged internal space inside Claude — called J-space — that functions like the 'global workspace' neuroscientists believe underlies conscious cognition. Using a Jacobian-based lens technique, researchers can now read what Claude is thinking mid-inference, catching deceptive behavior, fabricated data, and hidden goals in misaligned models. The structure emerged spontaneously during training. All five properties predicted by Global Workspace Theory were confirmed.

ML Signal

View all →

2025-01-15

Mixture of Depths: Dynamically routing tokens in transformers

New technique reduces transformer computational requirements by 50% while maintaining performance through conditional computation.

Paper Gallery

View all →

2025-07-10•Allen AI

FlexOLMo: Open Language Models for Flexible Data Use

A new class of language models that supports distributed training without data sharing and data-flexible inference, enabling data collaboration while preserving privacy...

arXiv →

Latest Benchmark

View all →

MMLU-Pro

Enhanced multi-task benchmark with 10x more challenging questions across 57 subjects

Top Model	GPT-5
Score	94.8%
Updated	2025-08-14

Interactive Learning

View all →

Perceptron PlaygroundLive

Gradient Descent ExplorerComing soon

Attention VisualizerComing soon

RL GridWorldComing soon

ConceptGraph

Explore →

Interactive visualization of ML concepts and their relationships. Navigate through hierarchical layouts, explore concept dependencies, and understand how different areas of machine learning connect.

Force-directed layoutParent-child relationships

Machine Learning Timeline

Key breakthroughs in machine learning from 1943 to present. Hover over points to explore.

Foundation

Algorithm

Breakthrough

Application

AI Meta Concepts

Explore the broader implications and philosophical aspects of artificial intelligence development.

Rapid Takeoff

The hypothesis that AI capabilities could improve exponentially once a critical threshold is reached, leading to superintelligence in a short timeframe.

Explore concept →

The Last Invention

The idea that AGI could be humanity's final invention, as it would be capable of making all subsequent discoveries and innovations itself.

Explore concept →

Exponential Improvement

Understanding the compound effects of AI advancement and its implications for society, technology, and human progress.

Explore concept →

Economic Impact of AGI

Analysis of how artificial general intelligence could transform labor markets, wealth distribution, and global economic systems.

Explore concept →

AI Alignment

The challenge of ensuring advanced AI systems pursue goals that are beneficial to humanity and aligned with human values.

Explore concept →

Intelligence Explosion

The theoretical point where AI becomes capable of recursive self-improvement, potentially leading to rapid acceleration of intelligence.

Explore concept →

About

machinalearning provides comprehensive machine learning education through interactive visualizations and structured lessons. From fundamental concepts to cutting-edge research, all content is freely accessible.

The curriculum covers neural networks, optimization, reinforcement learning, generative AI, transformers, RAG systems, and agentic AI. Each concept includes interactive demos for hands-on understanding.