2 min read

Claude Web Search Preview and Sesame Voice AI Mimics Human Speech

AI Models and Paradigms

Chain of Draft (CoD) Proposed to Enhance Efficiency in AI Reasoning: A new AI model paradigm called Chain of Draft (CoD) has been proposed, which reduces verbosity and focuses on critical insights, matching or surpassing the accuracy of Chain-of-Thought (CoT) while using significantly fewer tokens, thus reducing cost and latency across various reasoning tasks.

GPT-4.5 Achieves Top Position in Elimination Game Benchmark: GPT-4.5 Preview has achieved the top position in the Elimination Game Benchmark, which evaluates social reasoning skills such as forming alliances, deception, and persuasion. The benchmark results highlight GPT-4.5's strong performance in social reasoning tasks, outperforming other models in this specific area. However, it underperformed in the reasoning-oriented Step Game benchmark, where reasoning models held all top spots.

Atom of Thoughts (AOT) Method Enhances Smaller Models' Reasoning: The Atom of Thoughts (AOT) method has been introduced, enhancing the performance of the gpt-4o-mini model to achieve an 80.6% F1 score on HotpotQA, surpassing other models like o3-mini and DeepSeek-R1. The method involves decomposing questions into a Directed Acyclic Graph (DAG) and iteratively contracting subquestions until an atomic question is reached.

Sesame's Voice AI Mimics Human Voices with Remarkable Accuracy: Sesame has released a new Voice AI model that is capable of mimicking human voices with remarkable accuracy, making it difficult to distinguish between human and AI-generated speech. The demo showcases the AI's ability to adapt its tone and pitch, even in confrontational scenarios.

AI Products and Services

Claude Web Search Feature Coming Soon: Anthropic is set to release a web search feature for their AI model, Claude, as a preview soon. This feature was announced by Dario Amodei, CEO of Anthropic, in a recent interview. The feature is expected to enhance the capabilities of Claude, especially for users who need to access real-time information.

AI Research and Development

NVIDIA's Sim-to-Real Reinforcement Learning for Humanoid Manipulation: NVIDIA researchers have developed a new approach to reinforcement learning for vision-based dexterous manipulation on humanoids. The method includes novel techniques such as an automated real-to-sim tuning module, a generalized reward design scheme, a divide-and-conquer distillation process, and a mixture of sparse and dense object representations. The research demonstrates robust generalization and high performance in humanoid dexterous manipulation tasks without the need for human demonstration.

Meta's Non-Invasive Brain-to-Text Decoding System: Meta has developed a non-invasive brain-to-text decoding system using a brain-computer interface (BCI) that translates brain signals into text with a character error rate of approximately 7%. The system employs a combination of convolutional, transformer, and language modules to process brain signals and predict sentences, potentially aiding individuals with communication impairments.