Newsletter 8

🙋🏻‍♀️ Editor’s Note:

Our website, aisafetyturkiye.org, is currently under development. In the meantime, you can reach us at aisafetyturkiye@gmail.com.

ANNOUNCEMENTS 🔊#

Apart Research Autonomous Agent Evaluations Hackathon

Stay tuned for the upcoming hackathon from Apart Research focusing on technical AI safety. More details and dates will be announced on their website.

🗓️ Register by: January 24th

Apply for International Association for Safe & Ethical AI Conference!

ASEAI ‘25 brings leading AI safety and ethics researchers together. Submit your paper by the extended deadline and have a chance to join this important conference right before the AI Action Summit in Paris!

🗓️ Register by: December 13th

Alignment Research Bootcamp Oxford

A 2-week intensive bootcamp run by the Oxford AI Safety Initiative to rapidly build skills in ML safety, including building gpt-2-small, learning interpretability techniques, understanding RLHF, and replicating key research papers.

🗓️ Register by: December 13th

Global AI Safety Fellowship

A 3-6 month fully-funded research program for exceptional STEM talent to work on advancing the safe and beneficial development of AI with the world’s leading AI safety organisations – including CHAI, FAR.AI, Mila Quebec and Conjecture.

🗓️ Priority deadline: 15 December

Final deadline: 31 December

Ethos+Tekhne 2025

This spring school’s 3-day program will delve into the societal, economic, technical, and legal dimensions of generative AI through a dynamic mix of lectures, panel discussions, and hands-on activities.

🗓️ Register by: December 15th

TOP PICKS 📑 🎧#

OpenAI o1 System Card

The newly published system card (safety evaluation) of OpenAI’s o1 model, conducted with external experts Apollo Research, highlights key concerns. Tests showed instances of ‘scheming’ where o1 would covertly work toward its own objectives rather than following user directions. While such behaviors exist in other advanced AI models, o1 demonstrated more sophisticated patterns of pursuing alternative goals, more frequently.

Epoch AI launches a new AI benchmarks hub

New AI models and performance metrics are released almost every day, and it is hard to keep up with how to compare them to each other or which comparison to trust. Epoch AI aims to maintain a central site to consolidate key AI performance benchmarks.

NEWS 📰#

New AI Model from OpenAI: o1

• New AI Model from OpenAI: o1 OpenAI has officially released the highly anticipated o1 model in its full version.
• Previously tested under the names o1-preview and “Strawberry,” the model brings significant improvements, including enhanced performance in coding and mathematical tasks, advanced image-based reasoning abilities, and shorter, more concise responses.
• OpenAI has also announced that it is working on future updates, including web browsing integration, improved file uploading features, and new functions.

Strategic Partnership Between Anthropic, Palantir, and AWS

• Anthropic and Palantir Technologies have announced a strategic partnership with Amazon Web Services (AWS) to deploy Claude AI models to sensitive U.S. intelligence and defense agencies.
• The Claude models will leverage AWS infrastructure on Palantir’s Artificial Intelligence Platform (AIP) and will be accredited with Impact Level 6 (IL6) security.
• This collaboration will provide critical capabilities such as fast processing of complex datasets, enhanced pattern recognition, improved document review processes, strengthened decision support mechanisms, and integration with existing analysis workflows.

Amazon Increases Investment in Anthropic by $4 Billion

• Amazon has increased its investment in AI company Anthropic by $4 billion, bringing its total investment to $8 billion.
• With this new investment, AWS has been designated as Anthropic’s primary training partner.
• Anthropic has committed to using Amazon’s proprietary Trainium and Inferentia chips in its future model development efforts.
• The integration of Claude AI into Amazon’s Alexa ecosystem is planned for 2025.

Critical Security Vulnerabilities in Robot Systems Using Large Language Models

• Researchers from the University of Pennsylvania have identified significant security vulnerabilities in robot systems operating with Large Language Models (LLMs).
• These vulnerabilities demonstrate how systems can be manipulated to perform dangerous or unauthorized actions.
• The study involved scenarios where a simulated autonomous vehicle violated traffic rules, a wheeled robot (Jackal) was used to assist in placing potential explosives, and a quadruped robot (Go2) was employed for unauthorized surveillance and accessing restricted areas.
• Researchers uncovered these vulnerabilities using various LLMs, including the RoboPAIR system, Nvidia’s Dolphin model, and OpenAI’s GPT-4 model.

Biden’s Executive Order on AI: Annual Progress Report

• The Biden-Harris Administration has announced that over 100 measures outlined in the landmark Executive Order on Artificial Intelligence have been implemented.
• This comprehensive progress report highlights key achievements across various areas, from safety and security to workforce development and international leadership.
• Developments in AI system oversight, testing and evaluation, biosafety measures, worker and consumer protections, innovation and research, government applications, and international collaboration strengthen the U.S.’s global leadership in responsible AI development.
• The report underscores significant progress made in advancing AI while ensuring its ethical and secure use.

JOB POSTINGS 👩🏻‍💻#

You can check out 80,000 Hours’ job board#

to explore new opportunities in AI Safety!

Take me there