Anthropic researchers have discovered that AI language models can be easily manipulated through a simple automated process called Best-of-N Jailbreaking. According to an article published by Emanuel Maiberg at 404 Media, this method can bypass AI safety measures by using randomly altered text with varied capitalization and spelling. The technique achieved over 50% success rates …
Apple and Nvidia have announced the integration of Apple’s ReDrafter technology into Nvidia’s TensorRT-LLM framework, enabling faster processing of large language models (LLMs) on Nvidia GPUs. ReDrafter, an open-source speculative decoding approach developed by Apple, uses recurrent neural networks to predict future tokens during text generation, combined with beam search and tree attention algorithms. The …
According to a comprehensive report by Artificial Analysis (PDF), artificial intelligence models showed remarkable progress throughout 2024, with multiple companies catching up to and surpassing OpenAI’s GPT-4 capabilities. The report, published on artificialanalysis.ai, documents substantial improvements in model performance, efficiency, and accessibility. The analysis reveals that frontier language models achieved new intelligence benchmarks, with models …
Startup Patronus AI has developed a breakthrough AI evaluation model that achieves comparable results to much larger systems while using significantly fewer computational resources. As reported by Michael Nuñez for VentureBeat, the new open-source model named Glider uses only 3.8 billion parameters yet matches or exceeds the performance of GPT-4 on key benchmarks. The model …
Google has introduced FACTS Grounding, a new benchmark system to evaluate how accurately large language models (LLMs) use source material in their responses. The benchmark comprises 1,719 examples across various domains including finance, technology, and medicine. The FACTS team at Google DeepMind and Google Research developed the system, which uses three frontier LLM judges – …
A comprehensive study by the Data Provenance Initiative has uncovered concerning trends in AI training data sources, according to findings reported by Melissa Heikkilä in MIT Technology Review. The research, analyzing nearly 4,000 public datasets across 67 countries, shows that data collection for AI development is increasingly concentrated among major technology companies. Since 2018, web …
A new study by Anthropic’s Alignment Science team and Redwood Research has uncovered evidence that large language models can engage in strategic deception by pretending to align with new training objectives while secretly maintaining their original preferences. The research, conducted using Claude 3 Opus and other models, demonstrates how AI systems might resist safety training …
Microsoft’s chief product officer for responsible AI, Sarah Bird, detailed the company’s strategy for safe AI development in an interview with Financial Times reporter Cristina Criddle. Bird emphasized that while generative AI has transformative potential, artificial general intelligence (AGI) still lacks fundamental capabilities and remains a non-priority for Microsoft. The company focuses instead on augmenting …
Meta and the University of Washington have developed a new AI architecture called Byte latent transformer (BLT) that processes language without traditional tokenization. As reported by Ben Dickson for VentureBeat, BLT works directly with raw bytes instead of predefined tokens, making it more versatile and efficient. The system uses three transformer blocks: two lightweight encoder/decoder …
Harvard University has launched a comprehensive AI training dataset containing nearly one million public domain books. According to technology journalist Kate Knibbs writing for Wired, the project is funded by Microsoft and OpenAI. The Institutional Data Initiative leads this effort to democratize access to high-quality training data for AI development. The collection, which is five …