A new essay argues that rational humans and AIs should not operate based on goals, but instead align actions to "practices"—networks of actions, dispositions, evaluation criteria, and resources. The a…
Researchers argue that current multimodal AI approaches, which combine separate models for language, vision, and other data, will not achieve true Artificial General Intelligence (AGI) because they la…
Mathematics, once the primary driver of machine learning research, is shifting from providing theoretical guarantees and guiding granular architectural design to offering post-hoc explanations for emp…
LLM chatbot benchmarks like MMLU and HumanEval are becoming saturated, yet user experience is not improving proportionally because these tests measure one-shot performance rather than multi-turn, goal…
A new essay argues that the AI industry must develop concrete, positive visions for artificial intelligence grounded in human wellbeing, rather than falling into polarized debates of optimism or pessi…
Large language models (LLMs) like ChatGPT, which predict sequences of tokens, are being explored by quantitative traders for financial market predictions by modeling sequences of prices or trades inst…
AI models trained on human-created data reflect and often amplify real-world gender biases, as demonstrated by research showing sexist analogies in word embeddings and lower accuracy for darker-skinne…
Researchers Albert Gu and Tri Dao introduced Mamba, a State Space Model (SSM) that rivals Transformer performance while overcoming the quadratic bottleneck in attention mechanisms, enabling efficient …
Large language models (LLMs) could provide the breakthrough needed to make self-driving cars a reality, following a pattern similar to Alexander Fleming's accidental discovery of penicillin. While tra…
Text embeddings, the vector representations used to determine document relevance in Retrieval Augmented Generation (RAG) systems, can be inverted to recover the original text, according to a new study…
Machine learning models frequently fail in real-world applications due to overfitting on misleading data, hidden variables, and spurious correlations. Hundreds of COVID-19 prediction models proved use…
Researchers have developed deep learning methods to analyze single-cell sequencing data, enabling scientists to examine the genetic activity of individual cells and understand cellular diversity. This…
Complex sociotechnical challenge of counting fish passing through hydroelectric dams, a task the author encountered as a consultant in environmental science. It explains that dam operators must comply…
Researchers are exploring how to replicate classical algorithms—such as shortest path-finding and sorting—within deep neural networks to address key weaknesses in modern AI, including lack of provable…
Credulous mainstream coverage of "AI existential risk" has conflated speculative future dangers with real present-day harms, confusing large language models with algorithmic decision-making systems. T…