Week 43, end of the week

Week 44, end of the week episode

Welcome to AI & Consciousness, where we dissect the latest AI breakthroughs and ethical dilemmas. Today, we’re navigating a whirlwind of news. We have groundbreaking academic research on AI reasoning and efficiency, alongside major industry moves like OpenAI’s new security agent and Nvidia’s staggering $5 trillion valuation. We'll explore the rise of autonomous AI agents, the high-stakes economics of the AI boom, and the urgent question that looms over it all: As these systems grow more powerful, who is held accountable? Let’s dive in.

Week 43, end of the week

This week on AI & Consciousness, we explore the rise of superhuman AI agents achieving visually-guided computer control and new frameworks measuring machine creativity. We investigate how AI is transforming healthcare with smarter reasoning and reshaping finance with personalized investment insights and fraud detection. We also tackle the dark side: AI-driven insurance fraud, the ethical dilemmas of cultural bias, and the "Verification-Value Paradox" facing legal professionals. Join us for a deep dive into the technology and ethics shaping our future.

Week 43, mid week episode

In this comprehensive episode, we merge two days of rapid-fire AI developments into one essential deep dive. We explore the groundbreaking new frameworks that are finally helping to define and measure Artificial General Intelligence (AGI) and revealing the surprising limits of "reflection" in today's most advanced models. We'll cover major breakthroughs in healthcare, where AIs like DeepSomatic are finding previously missed cancer variants in children, and in law, where new systems are redefining corporate accountability in the age of algorithms. From clever 'jailbreaking' attacks that expose critical vulnerabilities to the fun side of AI playing Dungeons & Dragons, this episode covers the innovations, the risks, and the societal shifts shaping our world. Sharing a lot of references from this episode. * References and Further Reading * A Definition of AGI: Proposes a quantifiable framework to define and measure Artificial General Intelligence based on human cognitive abilities. * DeepSomatic: Details Google's AI model that identified 10 previously missed genetic variants in pediatric leukemia cells. * Distractor Injection Attacks: Reveals how top LLMs can be distracted by irrelevant tasks, cutting task accuracy by up to 60%. * DTKG: A dual-track knowledge graph framework that improves complex multi-hop question answering in RAG systems. * From Local to Global (GISP): Introduces GISP, a structured pruning method making LLMs up to 50% smaller without losing performance. * FST.ai [http://FST.ai] 2.0: An explainable AI system to assist Taekwondo referees, reducing decision review times by 85%. * Illusions of reflection: Shows that frontier LLMs lack functional, goal-driven reflective reasoning, a key gap in current AI capabilities. * Is Multilingual LLM Watermarking Truly Multilingual? (STEAM): Presents STEAM, a method using back-translation to fix fairness issues and ensure watermarking works in low-resource languages. * Na Prática, qual IA Entende o Direito?: A study finding that a specialized legal AI (JusIA) significantly outperforms general models like ChatGPT on legal tasks. * Operationalising Extended Cognition: Proposes a legal framework for holding corporations accountable for decisions made by their AI systems. * The Right to Be Remembered: Argues for a digital right to combat the erasure of minority voices and cultural memory by LLMs. * Team-Phi: A multi-agent framework that automatically evaluates and selects models for anonymizing patient health data. * VERA-V: A framework that automates the discovery of 'jailbreak' vulnerabilities in multimodal AIs like GPT-4o. * What Limits Agentic Systems Efficiency? (SpecCache): Introduces SpecCache, a method to speed up web-based AI agents by up to 3.2x via intelligent caching.

Week 42, end of the week. arxiv.org special edition

This friday, Iv'e dived down into arxiv.org [http://arxiv.org] last 7 days of publication within the AI and ML realm. arxiv.org [http://arxiv.org] is a free, open-access repository where researchers can upload and share scientific papers. These papers are known as e-prints or preprints, meaning they are often the versions of articles before they have been formally peer-reviewed and published in a traditional academic journal. Scientists and researchers use arXiv to share their findings with the global community immediately. This allows others to see, discuss, and build upon new work months or even years before it appears in a journal. If you want to read up on a paper early, this is one of the places to be. In the end of the podcast we rush through the lates news around AI.

Week 42, mid week episode

Today, we’re diving into a whirlwind of innovation: self-improving language models, AI-driven consumer insights, and the race to redefine industries like healthcare, finance, and market research. Let’s unpack how these advancements are reshaping our world—and what they mean for the future of human-AI collaboration.