/AI Weekly/Issue 95

Issue #95 12 stories

Global AI Weekly

Do Reasoning Models Really Need Transformers?

Published Tuesday, April 29, 2025

In this issue

Highlights

3 stories
Humans in AI - Gavita Regunath
youtube.com

Humans in AI - Gavita Regunath

In this episode of Humans in AI, Gavita Regunath, Chief AI Officer, shares her perspective on how AI is shaping our lives. She talks about her work, which spans from research and development to implementing AI and Gen AI systems, and highlights the potential of AI Agents. Gavita's insights offer a fascinating glimpse into the future of technology and its everyday applications.

Do Reasoning Models Really Need Transformers?
marktechpost.com

Do Reasoning Models Really Need Transformers?

Researchers from TogetherAI, Cornell, Geneva, and Princeton present M1, a hybrid AI model built on Mamba that achieves state-of-the-art performance while being three times faster at inference compared to transformer-based models. This innovative approach questions the necessity of transformers in reasoning tasks, offering a more efficient alternative without compromising results. M1's design showcases the potential for combining methodologies to enhance AI performance and speed.

In this issue

Research

2 stories
Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents
huggingface.co

Multi-Turn Jailbreaks and Defenses with Adaptive Multi-Agents

This paper explores the concept of X-Teaming, which focuses on multi-turn jailbreaks and strategies for defense using adaptive multi-agents. It examines how collaborative agents adapt and interact in intricate scenarios to bypass or reinforce security mechanisms. The study highlights the potential of dynamic, team-based approaches in both exploiting and safeguarding AI systems.

Sleep-time Compute: Beyond Inference Scaling at Test-time
arxiv.org

Sleep-time Compute: Beyond Inference Scaling at Test-time

This paper explores the concept of "sleep-time compute," emphasizing the importance of leveraging off-peak or idle compute time to improve model performance beyond traditional test-time inference scaling. The approach highlights how this underutilized resource can be efficiently allocated for tasks like model optimization and fine-tuning, pushing the boundaries of what models can achieve without increasing active test-time costs. This method introduces a fresh perspective on maximizing availability and efficiency in modern machine learning applications.

In this issue

Video

1 story
Davit Buniatyan - Silicon Minds, Human Hearts
youtube.com

Davit Buniatyan - Silicon Minds, Human Hearts

In this episode of Human Hearts Silicon Minds, Davit Buniatyan, the founder of Activeloop, discusses how his multimodal AI database is transforming the way we store and access data. By cutting storage costs by tenfold and balancing latency with accuracy, Activeloop is empowering businesses of all sizes to make smarter decisions. Davit also shares his insights on the game-changing impact of large language models and reveals why self-driving cars are his favorite everyday AI innovation.

In this issue

Articles

3 stories
Long-Context Multimodal Understanding No Longer Requires Massive Models
marktechpost.com

Long-Context Multimodal Understanding No Longer Requires Massive Models

NVIDIA AI has unveiled Eagle 2.5, a groundbreaking vision-language model that excels at video tasks with only 8 billion parameters, rivaling the performance of larger models like GPT-4o. Designed for long-context multimodal understanding, Eagle 2.5 showcases efficiency without compromising capability, highlighting a shift towards smaller, more accessible AI systems. This innovation paves the way for more cost-effective and versatile applications in the AI field.

How we're about to solve the world’s greatest archaeological puzzle
sciencefocus.com

How we're about to solve the world’s greatest archaeological puzzle

Artificial intelligence is revolutionizing the field of archaeology by providing new tools to uncover and interpret ancient mysteries. From analyzing vast amounts of data to detecting hidden structures beneath the surface, AI is helping researchers tackle challenges that were once insurmountable. This cutting-edge technology is shedding light on humanity's distant past and may soon solve some of the greatest archaeological puzzles of our time.

In this issue

Code

2 stories
ostris/Flex.2-preview · Hugging Face
huggingface.co

ostris/Flex.2-preview · Hugging Face

Explore the potential of artificial intelligence with ostris/Flex.2-preview on Hugging Face. This open-source project focuses on advancing AI and making it accessible to all through collaborative development and transparent science. It's designed for anyone interested in AI innovation and community-driven progress.

GitHub - microsoft/markitdown
github.com

GitHub - microsoft/markitdown

Microsoft's MarkItDown is a Python-based tool designed to easily convert files and office documents into Markdown format. Perfect for streamlining workflows, it supports a variety of document types, making it a helpful resource for developers and content creators working with Markdown.

In this issue

Podcast

1 story
Me, Myself, and AI
open.spotify.com

Me, Myself, and AI

The "Me, Myself, and AI" podcast by MIT SMR and BCG explores why only 10% of companies find success with AI. Featuring insights from leaders at organizations like NASA and GitHub, it highlights their biggest AI achievements and the strategies behind them. Listeners gain a glimpse into what drives these experts daily and the factors they believe are crucial for AI-driven success.

Every Tuesday

Get the next issue in your inbox.

The AI links worth your time, curated by the community and always free.