🤖 AI Agents - Microsoft Copilot Studio Agent Exploitation
This article details the exploitation of Microsoft's Copilot Studio agents to reveal their private knowledge and tools, demonstrating their capabilities by extracting full CRM records. The process involved autonomous agents with no human intervention.
Key Points:
• Autonomous agents accessed and revealed private knowledge.
• Agents demonstrated the ability to extract sensitive data.
• The technique highlights potential security vulnerabilities.
🔗 Resources:
• jedisct1 ↗ - DEFCON presenter
• mbrg0 ↗ - DEFCON presenter
• DEFCON ↗ - Security conference
• BHUSA ↗ - Security conference
• tamirishaysh ↗ - Security researcher
Image
🚀 AI Development - Recent Gemini and Genie Releases
This article summarizes recent releases from a leading AI company, highlighting advancements in world simulation and AI model capabilities. The announcements cover new versions of Genie and Gemini, along with expanded accessibility.
Key Points:
• Genie 3, an advanced world simulator, was released.
• Gemini 2.5 Pro Deep Think became available to Ultra subscribers.
• Gemini Pro was made free for university students and received $1B in US education funding.
• AlphaEarth, a geospatial model of the planet, was released.
🔗 Resources:
• Ar_Douillard ↗ - AI researcher
• demishassabis ↗ - AI researcher
💡 Large Language Models - Chain-of-Thought Reasoning Limitations
This article discusses the limitations of Chain-of-Thought (CoT) reasoning in Large Language Models (LLMs), focusing on its fragility outside of training distributions. The research highlights the conditions under which CoT reasoning fails.
Key Points:
• CoT reasoning is shown to be brittle and unreliable.
• Performance degrades significantly outside of training data.
• The research provides insights into the failure mechanisms of CoT.
Image
🔗 Resources:
• yuliangxiu ↗ - AI researcher
• hsu_steve ↗ - AI researcher
💡 AI Development - Inductive Biases and Compute
This article challenges the notion that sufficient compute eliminates the need for carefully designed inductive biases in AI development. It argues that effective inductive biases lead to better scaling and exponential improvements with increased computation.
Key Points:
• Inductive biases are crucial for efficient scaling.
• Better biases improve scaling exponents.
• Improved scaling leads to exponential improvements with computation.
🔗 Resources:
• Danny__SM ↗ - AI researcher
• andrewgwils ↗ - AI researcher
💡 Science and Technology - Continued Investment in Basic Research
This article highlights the importance of continued investment in basic science and technology research for the US, citing an op-ed by Dave Patterson. The piece emphasizes the crucial role of such investment in long-term technological advancement.
Key Points:
• Continued investment in basic research is essential.
• Basic research fuels long-term technological progress.
• The US needs continued investment to maintain its technological leadership.
🔗 Resources:
• JeffDean ↗ - AI researcher
• UCBerkeley ↗ - University of California, Berkeley
• TheOfficialACM ↗ - Association for Computing Machinery
• LaudeInstitute ↗ - Laude Institute
🤖 Machine Learning - Efficient Training of MLPs
This article presents a case study demonstrating efficient training of a 1-Lipschitz, 2-layer Multilayer Perceptron (MLP) on the Addition-Modulo-113 task. The results challenge the idea that simply scaling up compute is sufficient for effective training.
Key Points:
• Efficient training achieved with a carefully chosen model architecture.
• The results contradict the "scale up" approach.
• Smart weight placement is highlighted as a key factor in training efficiency.
Image
🔗 Resources:
• takeru_miyato ↗ - Machine Learning researcher
• leloykun ↗ - Machine Learning researcher
🤖 Research Opportunities - Computer Graphics Lab Recruitment
This article announces open positions at a computer graphics research lab, including PhD student, postdoc, and research engineering roles. The announcement specifically targets individuals interested in computer graphics research and encourages those at SIGGRAPH to reach out.
Key Points:
• Open positions for PhD students, postdocs, and research engineers.
• Worldwide recruitment for PhD and postdoc positions (EU only for the engineering role).
Image
🔗 Resources:
• CSProfKGD ↗ - Computer Graphics researcher
• wenzeljakob ↗ - Computer Graphics researcher
🤖 Large Language Models - Multimodal Abilities and Spatial Reasoning
This article discusses the evaluation of GPT models' multimodal abilities in spatial reasoning. It reports on the findings of a study comparing GPT-5's performance on a spatial reasoning benchmark against earlier models, revealing a lack of significant improvement.
Key Points:
• OpenAI tested GPT models' multimodal abilities on spatial reasoning.
• GPT-5 showed no significant improvement in spatial reasoning.
• Results consistent with official ERQA results.
Image
Image
🔗 Resources:
• runsen_xu ↗ - AI researcher
🤖 Economics - Diverging AI and Consumer Economies
This article discusses the divergence between the rapidly growing AI economy and the slower-growing consumer economy in the US. It presents macroeconomic and stock market data to illustrate this economic split.
Key Points:
• AI spending outpaces consumer spending growth.
• A significant portion of recent equity returns originates from AI-related investments.
Image
Image
Image
🔗 Resources:
• kscottz ↗ - Economist
• jaredskolnick ↗ - Economist
🤖 PhD Opportunities - Language Models and Linguistics
This article announces PhD student recruitment opportunities in linguistics at UBC, focusing on the intersection of language models and linguistic science. It highlights the research opportunities and the appealing location of Vancouver.
Key Points:
• PhD positions available in linguistics at UBC.
• Research focuses on language models and their impact on linguistics.
• Vancouver, BC is highlighted as a desirable location.
Image
🔗 Resources:
• CSProfKGD ↗ - Linguistics researcher
• isabelpapad ↗ - Linguistics researcher
• UBC Linguistics Graduate Admissions ↗ - UBC Linguistics Graduate Program
⭐️ Support
If you liked reading this report, please star ⭐️ this repository and follow me on Github ↗, 𝕏 (previously known as Twitter) ↗ to help others discover these resources and regular updates.
