AI research | The Decoder

Maximilian Schreiner

Jan 20, 2025

AI research

Sakana AI's Transformer² is a new approach to help language models learn

Japanese company Sakana AI introduces Transformer², an approach for language models that can adapt more flexibly to new tasks using expert vectors trained with Singular Value Fine-Tuning (SVF). Transformer² learns optimal expert vectors through reinforcement learning and can combine them using different strategies to efficiently adapt to new tasks. In tests, Transformer² outperformed LoRAs by up to 16 percent on mathematical tasks while requiring significantly fewer parameters. But for now, SVF can only work with the skills already present in the pre-trained model.

Maximilian Schreiner

Jan 20, 2025

AI research

DeepSeek's latest R1 model matches OpenAI's o1 in reasoning benchmarks

Matthias Bastian

Jan 19, 2025

AI research

OpenAI quietly funded independent math benchmark before setting record with o3

Matthias Bastian

Jan 18, 2025

AI research

ELIZA, the world's first chatbot, is back online after six decades thanks to AI historians

Researchers revive OG ChatGPT ELIZA from the 1960s

Maximilian Schreiner

Jan 16, 2025

AI research

MatterGen: Microsoft presents AI tools for generating and simulating new materials

Maximilian Schreiner

Jan 16, 2025

AI research

Ndea's deep learning-guided program synthesis aims to create AI that learns like humans

Maximilian Schreiner

Jan 15, 2025

AI research

Frequent AI use linked to lower critical thinking scores, research shows

Jonathan Kemper

Jan 15, 2025

AI research

MiniMax introduces AI models with record context length for agents with 'long term memory'

Graphic design: "MiniMax-01" text on a red background with geometric shapes, minimalist style with yellow accents.

Maximilian Schreiner

Jan 14, 2025

AI research

Microsoft tested 100 AI tools and found hackers don't need complex math to break them

Matthias Bastian

Jan 12, 2025

AI research

Microsoft's rStar-Math enables small language models to ace complicated math problems

Microsoft Research Asia has introduced rStar-Math, a new method that enables small language models (SLMs) with only 1.5 to 7 billion parameters to match or exceed the performance of OpenAI's o1-preview in mathematical tasks.