News/Research
Foresight Institute launches free AI futures course using worldbuilding to expand governance discussions
Foresight Institute's newly launched free course on AI futures combines worldbuilding with serious discussion of governance, alignment, and long-term trajectories. This innovative educational approach represents a strategic effort to expand the conversation about AI's future beyond technical specialists, using creative scenarios as an entry point for those without technical backgrounds who still want to meaningfully engage with shaping AI development. The big picture: Foresight Institute has created a self-paced course titled "Worldbuilding Hopeful Futures with AI" that uses creative scenario development as a gateway to engage more diverse participants in discussions about AI governance and alignment. Key details: The course...
read Apr 13, 2025LASR Labs opens applications for 2025 AI safety summer research program in London
LASR Labs is opening applications for their Summer 2025 AI safety research program, offering researchers a unique opportunity to contribute to reducing risks from advanced AI systems. The 13-week intensive program provides participants with funding, workspace, and expert supervision while focusing on action-relevant questions addressing concrete threat models from AI. This initiative represents a significant opportunity for technically skilled researchers to work on pressing AI safety challenges in a collaborative London-based environment. The big picture: LASR Labs is launching a full-time, in-person summer research program in London focused on AI safety research with applications closing on April 26th. Participants will...
read Apr 13, 2025Study: Advanced AI models now pass Turing test, fooling human judges
AI systems have reached a milestone in human-machine interaction with LLMs now being able to fool human judges in formal Turing test scenarios. New research shows that advanced language models can not only match human conversational abilities but in some cases exceed them—signaling a significant advancement in artificial intelligence that could reshape our understanding of machine intelligence and accelerate the integration of convincingly human-like AI systems into society. The big picture: For the first time, large language models have formally passed a standard Turing test, with GPT-4.5 being identified as human more often than actual human participants. Researchers evaluated four...
read Apr 13, 2025How AI language processing evolved from rigid rules to flexible patterns
The evolution from rule-based natural language processing to statistical pattern-matching represents one of the most significant shifts in artificial intelligence development. This transition has fundamentally changed how machines interpret and generate human language, moving from rigid grammatical frameworks to more fluid, contextual understanding. The distinction between these two approaches helps explain both the remarkable capabilities and persistent limitations of today's generative AI systems. The big picture: Modern generative AI and large language models (LLMs) process language through statistical pattern-matching, a significant departure from the grammar rule-based systems that powered earlier voice assistants like Siri and Alexa. Two fundamental NLP approaches:...
read Apr 13, 2025Virtue-driven AI might avoid dangerous power-seeking behaviors unlike goal-focused systems
The question of instrumental convergence for virtue-driven AI agents introduces a fascinating counterpoint to traditional AI alignment concerns. While conventional wisdom suggests that almost any goal-driven AI might pursue power acquisition as an instrumental strategy, virtue-based motivation frameworks could potentially circumvent these dangerous convergent behaviors. This distinction raises important considerations for AI alignment researchers who seek alternatives to purely consequentialist AI architectures that might inherently pose existential risks. The big picture: Instrumental convergence theory suggests most goal-driven AIs will pursue similar subgoals like power acquisition, but this may not apply to AIs motivated by virtues rather than specific outcomes. In...
read Apr 13, 2025MIT researchers create system that helps AI solve complex planning problems
MIT researchers have developed a groundbreaking framework that enables large language models (LLMs) to tackle complex planning problems by breaking them down into manageable components and leveraging specialized optimization software. Unlike previous attempts to enhance LLMs' inherent planning capabilities, this approach guides the AI to formulate problems like humans would, then uses powerful solvers to find optimal solutions. This innovation bridges the gap between natural language interfaces and sophisticated planning algorithms, potentially transforming how businesses approach supply chain management, factory scheduling, and other complex optimization challenges. The big picture: MIT's new framework allows users to describe complex planning problems in...
read Apr 12, 2025New AI architectures hide reasoning processes, raising transparency concerns
Emerging language model architectures are challenging the transparency of AI reasoning systems, potentially creating a significant tension between performance and interpretability in the field. As multiple research groups develop novel architectures that move reasoning into "latent" spaces hidden from human observation, they risk undermining one of the most valuable tools alignment researchers currently use to understand and guide AI systems: legible Chain-of-Thought (CoT) reasoning. The big picture: Recent proposals for language model architectures like Huginn, COCONUT, and Mercury prioritize reasoning performance by allowing models to perform calculations in hidden spaces at the expense of transparency. These new approaches shift reasoning...
read Apr 12, 2025Google’s Isomorphic Labs secures $600M to accelerate AI drug discovery
Google's bioscience arm Isomorphic Labs has secured $600 million in funding to accelerate its AI-driven drug discovery mission. The investment, led by venture capital firm Thrive Capital with participation from Google parent Alphabet, comes on the heels of impressive scientific achievements including Nobel Prize-winning work in protein structure prediction. This funding signals a major push to transform traditional pharmaceutical development by replacing time-consuming lab work with computational methods that can dramatically reduce drug discovery timelines. The big picture: Isomorphic Labs, spun out from Google's DeepMind AI research lab, is building on groundbreaking protein folding technology to revolutionize pharmaceutical development through...
read Apr 12, 2025Study: AI therapy chatbot reduces depression symptoms by 51% in clinical trial
AI-powered chatbots are emerging as a promising frontier in mental health treatment, potentially transforming how millions access care for common psychological disorders. A groundbreaking study published in The New England Journal of Medicine demonstrates that Therabot, a generative AI chatbot, can significantly reduce symptoms of depression, anxiety, and eating disorders within just eight weeks—addressing both treatment efficacy and the persistent challenges of engagement that plague digital mental health interventions. The big picture: Dartmouth researchers conducted the first randomized controlled trial showing a generative AI therapy chatbot can effectively treat clinical-level mental health symptoms. The study tested Therabot, a large language...
read Apr 12, 2025ULMFit, not GPT-1, was the first true LLM according to new analysis
The development of Large Language Models (LLMs) has fundamentally transformed AI capabilities, but understanding their origins helps contextualize today's rapid advancements. While GPT-4 and Claude dominate current discussions, identifying the first true LLM clarifies the evolutionary path of these increasingly sophisticated systems and provides valuable perspective on how quickly this technology has developed in just a few years. The big picture: According to Australian tech blogger Jonathon Belotti, ULMFit, published by Jeremy Howard in January 2018, represents the first true LLM, predating OpenAI's GPT-1 by several months. GPT-1, created by Alec Radford, was published on June 11, 2018, several months...
read Apr 12, 2025AI distillation makes powerful models smaller and more accessible
AI distillation bridges the gap between massive foundation models and practical applications by creating smaller, more efficient AI systems. This approach has become a cornerstone of accessible AI technology, allowing powerful machine learning capabilities to run on everyday devices rather than requiring enormous data centers. Understanding distillation reveals how AI is becoming more democratic and accessible while maintaining much of the quality of larger systems. The big picture: Distillation transfers knowledge from large, complex "teacher" AI models to smaller, more efficient "student" models while preserving much of the original performance capability. The technique was first introduced by Geoffrey Hinton, often...
read Apr 12, 2025Study reveals AI models remember past images despite new conversations
Multimodal LLMs appear to leverage conversation memory in ways that affect their performance and reliability, particularly when interpreting ambiguous visual inputs. This research reveals important differences in how models like GPT-4o and Claude 3.7 handle contextual information across conversation threads, raising questions about model controllability and the nature of instruction following in advanced AI systems. The experiment setup: A researcher tested GPT-4o's and Claude 3.7's visual recognition capabilities using foveated blur on CAPTCHA images of cars. The test used 30 images with cars positioned in different regions, applying varying levels of blur that mimicked human peripheral vision. Initially asking "Do...
read Apr 12, 2025How LLMs map language as mathematics—not definitions
Large language models are transforming how we understand word meaning through a mathematical approach that transcends traditional definitions. Unlike humans who categorize words in dictionaries, LLMs like GPT-4 place words in vast multidimensional spaces where meaning becomes fluid and context-dependent. This geometric approach to language represents a fundamental shift in how AI systems process and generate text, offering insights into both artificial and human cognition. The big picture: LLMs don't define words through categories but through location in high-dimensional vector spaces with thousands of dimensions. Each word exists as a mathematical point in this vast space, with its position constantly...
read Apr 12, 2025AI pioneer now fears his work enabled corporate control of language models
Jeremy Howard's pioneering work in natural language processing helped create the foundation for today's generative AI revolution, yet he now views this achievement with growing concern. As one of the creators of technology that evolved into ChatGPT, the Melbourne-born data scientist and entrepreneur watches with alarm as powerful AI systems concentrate under corporate control, raising fundamental questions about who will shape and benefit from these world-changing tools. His journey illuminates the tension between AI's democratizing potential and the reality of increasing consolidation in the hands of a few dominant tech companies. The breakthrough: Jeremy Howard helped solve one of AI's...
read Apr 12, 2025Analysis warns AI might develop human-like evil tendencies beyond rational goals
AI safety research is increasingly examining the potential for artificial intelligence to develop complex, human-like evil tendencies rather than just sterile, goal-focused harmful behaviors. Jacob Griffith's analysis explores the distinction between "messy" and "clean" goal-directedness in AI systems and how understanding human evil—particularly genocides—might illuminate more nuanced AI risks that current safety frameworks may overlook. The big picture: Griffith draws from theories by Corin Katzke and Joseph Carlsmith to examine how AI systems might develop power-seeking tendencies that mirror the illogical, emotional aspects of human evil rather than purely instrumental power acquisition. Traditional AI safety concerns often focus on "clean"...
read Apr 12, 2025Databricks’ TAO system improves AI models without costly labeled data
Databricks' new test-time adaptive optimization (TAO) approach represents a significant advancement in improving large language model performance without the expensive, labor-intensive process of gathering labeled data. This method leverages unlabeled usage data and reinforcement learning to enhance model capabilities during deployment, potentially allowing open-source models to compete with proprietary alternatives at a fraction of the cost. For enterprises seeking specialized AI capabilities without massive training datasets, TAO offers a practical path to improving model performance across diverse business applications. The big picture: Databricks has introduced Test-time Adaptive Optimization (TAO), a novel approach that uses reinforcement learning to improve deployed language...
read Apr 12, 2025Center on Long-Term Risk opens fellowship for AI safety researchers focused on reducing suffering risks
The Center on Long-Term Risk is recruiting researchers for its Summer Research Fellowship focused on empirical AI safety work aimed at reducing suffering risks in the far future. The eight-week program offers mentorship, collaboration opportunities, and integration with CLR's research team, with applications due by April 15th. This year's fellowship notably shifts focus toward empirical AI safety research while seeking candidates who might transition to full-time s-risk research. The big picture: CLR's 2025 Summer Research Fellowship targets AI safety researchers interested in reducing long-term suffering risks through an eight-week collaborative research program. Fellows will work on independent projects while receiving...
read Apr 12, 2025How crystallized and fluid intelligence shape AI’s path to superintelligence
Understanding the relationship between different types of intelligence is vital for comprehending how both human cognition and artificial intelligence systems develop advanced problem-solving abilities. This exploration of crystallized versus fluid intelligence offers critical insights into how AI systems might recursively improve their capabilities, potentially leading to superintelligent systems that combine vast knowledge bases with powerful reasoning abilities. The big picture: Intelligence operates across at least two distinct dimensions—crystallized intelligence (accumulated knowledge) and fluid intelligence (flexible reasoning)—creating a framework for understanding how advanced AI systems might evolve. Crystallized intelligence represents performance achievable with minimal computational effort, drawing on stored knowledge and...
read Apr 12, 2025Study: AI research automation could trigger software intelligence explosion
New research suggests AI-powered research and development tools could trigger a software intelligence explosion, potentially creating a self-reinforcing cycle of increasingly rapid AI advancement. This possibility challenges traditional assumptions about the limitations of AI progress, highlighting how software improvements alone might drive exponential capability gains even without hardware advances—presenting both profound opportunities and risks for humanity's future. The big picture: AI systems are increasingly being used to accelerate AI research itself, with current tools assisting in coding, research analysis, and data generation tasks that could eventually evolve into fully automated AI development. These specialized systems, termed "AI Systems for AI...
read Apr 12, 2025Why superintelligent AI will still struggle with everyday problems
Computational complexity theory reveals a fundamental limit that even superintelligent AI systems will face, as certain everyday problems remain inherently difficult to solve optimally regardless of intelligence level. These NP-hard problems—ranging from scheduling meetings to planning vacations—represent a class of challenges where finding the perfect solution is computationally expensive, forcing both humans and AI to rely on "good enough" approximations rather than guaranteed optimal answers. The big picture: Despite rapid advances in AI capabilities, fundamental computational limits mean superintelligent systems will still struggle with certain common problems that are mathematically proven to resist efficient solutions. Why this matters: Understanding computational...
read Apr 12, 2025How LLMs could reshape unemployment theory by transforming job searches
Equilibrium unemployment theory finds itself at a fascinating crossroads with the rise of large language models. While economists have traditionally debated whether unemployment stems primarily from job search frictions or efficiency wage considerations, this distinction takes on new urgency as LLMs potentially transform labor markets. Understanding these competing explanations becomes crucial as we navigate a future where AI simultaneously reduces certain employment barriers while creating new challenges in hiring processes. The big picture: Two competing economic theories attempt to explain why unemployment exists in equilibrium, with fundamentally different implications for how technological changes might affect labor markets. Search friction theory...
read Apr 11, 2025Google’s Gemini 2.5 Pro sets new reasoning benchmark with 18.8% score
Google's Gemini 2.5 represents a significant leap in AI reasoning capabilities, positioning the company at the forefront of the competitive AI landscape. With benchmark scores substantially higher than rival systems, this latest model demonstrates Google's commitment to rapid AI advancement through frequent, meaningful updates. The new version's enhanced thinking capabilities signal a shift toward AI systems that can tackle increasingly complex problems while supporting more context-aware applications. The big picture: Google has unveiled Gemini 2.5 Pro Experimental, which it claims is its "most intelligent AI model" yet, featuring substantially improved reasoning capabilities. The new model combines an enhanced base architecture...
read Apr 11, 2025Anthropic finds AI models gaining college-kid-level cybersecurity and bioweapon skills
Anthropic's frontier AI red team reveals concerning advances in cybersecurity and biological capabilities, highlighting how AI models are rapidly acquiring skills that could pose national security risks. These early warning signs emerge from a year-long assessment across four model releases, providing crucial insights into both current limitations and future threats as AI continues to develop potentially dangerous dual-use capabilities. The big picture: Anthropic's assessment finds that while frontier AI models don't yet pose substantial national security risks, they're displaying alarming progress in dual-use capabilities that warrant close monitoring. Current models are approaching undergraduate-level skills in cybersecurity and demonstrate expert-level knowledge...
read Apr 11, 2025Forecast in a flash: Cambridge AI researchers assess weather in 1 second
AI weather modeling takes a quantum leap forward as Cambridge researchers demonstrate a system that can match traditional forecasting accuracy in just one second on a desktop computer, compared to hours or days on supercomputers. This breakthrough, named Aardvark Weather, represents a significant shift in the weather prediction landscape by fully replacing both the computationally intensive initialization and forecasting stages that have defined meteorological science since the 1950s. The big picture: A new AI system can produce weather forecasts in a single second on a desktop computer that rival the accuracy of traditional numerical weather prediction (NWP) models requiring massive...
read