-
Bias, Toxicity, and Jailbreaking Large Language Models (LLMs)
Recent research highlights concerns about Large Language Models (LLMs), such as biased outputs and environmental impacts. Further details are available on Towards Data Science.
-
Sam Altman’s firing not related to safety, says Microsoft’s Brad Smith
Microsoft President Brad Smith stated Sam Altman’s temporary departure from OpenAI was not due to AI safety issues. Amid speculation and internal concerns over Altman’s management style, Microsoft, a close partner, has secured a non-voting observer seat on OpenAI’s board. Altman has since been reinstated, pledging to advance OpenAI’s mission and safety.
-
Microsoft plans £2.5 billion investment in the UK AI industry
Microsoft plans to invest £2.5 billion in the UK tech industry, focusing on AI infrastructure and development. The investment will expand data centers, introduce 20,000 GPUs by 2026, and train over a million people in AI skills. This move aims to reinforce the UK as a leading science and AI hub.
-
Easily build semantic image search using Amazon Titan
Digital publishers use machine learning for faster content creation, ensuring relevant images match articles. Amazon’s Titan Multimodal Embeddings model generates image and text embeddings for semantic search. This streamlines finding appropriate images, without keywords, by comparing metadata similarity—enhancing media workflows while maintaining quality. Amazon Bedrock simplifies AI application development for various modalities.
-
What Algorithms can Transformers Learn? A Study in Length Generalization
The paper explores Transformers’ capabilities in length generalization on algorithmic tasks and proposes a framework to predict their performance in this area. Accepted at NeurIPS 2023’s MATH workshop, it addresses the paradox of language models’ emergent properties versus their struggles with simple reasoning.
-
Increasing Coverage and Precision of Textual Information in Multilingual Knowledge Graphs
Researchers use knowledge graphs to enhance neural models in Natural Language Processing (NLP) and Computer Vision, grounding them in organized data. However, non-English languages face a scarcity of quality textual data. A new task, automatic Knowledge Graph Enhancement (KGE), has been introduced to improve non-English textual data’s quantity and quality.
-
SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding
This study, presented at NeurIPS 2023’s UniReps Workshop, introduces an efficient approach to combine vision foundation models (VFMs) like CLIP and SAM into a single model that leverages their respective semantic and spatial understanding strengths through multi-task learning techniques.
-
Swap Agnostic Learning, or Characterizing Omniprediction via Multicalibration
This work confirms that multigroup fairness concepts yield strong omniprediction—loss minimization across diverse loss functions. The study establishes a reciprocal link, showing that multicalibration and omniprediction are equivalent. New definitions are proposed. (47 words)
-
Federated Learning for Speech Recognition: Revisiting Current Trends Towards Large-Scale ASR
This paper, accepted for the NeurIPS 2023 workshop, discusses the overlooked potential of automatic speech recognition (ASR) in federated learning (FL) and differential privacy (DP), highlighting ASR’s suitability as a benchmark due to its data distribution and real-world relevance.
-
Can AI solve your problem?
Daniel Bakkelund suggests three heuristics to evaluate AI project viability: First, ensure you can clearly articulate the problem in writing. Second, ascertain if an informed human could theoretically solve the problem, given unlimited resources and time. Third, confirm that all necessary context for the AI to learn and give answers is available. If all conditions…