AI News

A New York City AI chatbot intended to help business owners has been sharing inaccurate and misleading information, particularly about legal matters. Launched in October, it is still in pilot mode, with the government acknowledging its shortcomings and urging users to verify information through other sources.

OpenAI has introduced the "Voice Engine," a voice cloning AI model enabling realistic speech generation from a 15-second audio clip. Currently in limited testing, it assists various partners and addresses therapeutic needs. OpenAI emphasizes ethical use, implementing safety measures against misuse.

Elon Musk's xAI will launch Grok-1.5 next week, an upgraded language model enhancing reasoning and problem-solving, with improved metrics like 50.6% on MATH and 81.3% on MMLU. It features a context window of 128,000 tokens. Grok-1.5 will be available to Premium users on the X platform.

Google will integrate its Gemini language model into Pixel 8 smartphones using Gemini Nano, which operates locally. This follows its debut on Pixel 8 Pro and Samsung Galaxy S24. Notably, the Pixel 8 will gain features like enhanced summarization in the Recorder app and improved Smart Reply in Gboard.

A study by Google’s DeepMind found that its AI system, the Search-Augmented Factuality Evaluator (SAFE), outperforms human fact-checkers in verifying claims from large language models, achieving 76% accuracy in discrepancies. SAFE is more cost-effective, being 20 times cheaper than human checks. Despite claims of "superhuman" performance, experts urge benchmarking against expert human raters. Transparency in methodologies and qualifications is essential for assessing effectiveness. Innovations like SAFE are crucial for improving trust in AI-generated content, but must involve diverse stakeholder input.

SambaNova Systems has unveiled its Samba-CoE v0.2 Large Language Model, achieving 330 tokens per second while requiring only 8 sockets, outperforming competitors like Databricks and MistralAI. This efficiency breakthrough supports ongoing innovations and positions SambaNova as a key player in the AI landscape.

SydeLabs, a California startup, is tackling cybersecurity challenges posed by generative AI with its real-time, intent-based firewall and three key products: SydeBox, SydeGuard, and SydeComply. Recently funded with $2.5 million, the company aims to enhance vulnerability detection in AI applications while ensuring their security and compliance.

MineOS has launched an AI Asset Discovery and Risk Assessment module, enhancing visibility into AI tools for effective governance amid rising privacy concerns. This tool helps organizations identify AI systems and assess risks, ensuring compliance with regulations like the EU's AI Act, while positioning MineOS as a leader in the data privacy sector.

AI21 Labs has introduced “Jamba,” a hybrid generative AI framework combining Structured State Space models and transformer architecture. While not replacing transformers, Jamba aims to enhance generative reasoning tasks and optimize context management. It features a 256K context window and a Mixture of Experts model for improved efficiency. Jamba is set to enter beta on the AI21 Platform soon.

Microsoft has launched new Azure AI tools to enhance the safety and reliability of generative AI applications. These offerings address issues like automatic hallucinations and prompt injection attacks through features such as Prompt Shields and Groundedness Detection, promoting secure development practices and real-time monitoring.

Hume AI has raised $50 million in Series B funding to advance its emotionally intelligent AI, focusing on understanding user emotions through voice. Its Empathic Voice Interface (EVI) aims to recognize over 50 nuanced emotions, enhancing user interactions. Hume AI promotes ethical use of its technology and has received positive industry feedback.

Lightning AI, in partnership with Nvidia, launched Thunder, an open-source compiler for PyTorch that enhances AI model training efficiency by up to 40% using multiple GPUs. It aims to optimize GPU utilization without needing more hardware. Thunder's development is led by PyTorch expert Thomas Viehmann.

Inflection is set to limit free access to its AI chatbot, Pi, in the upcoming months, as announced by the startup. In a significant update, users will now have the ability to export their Pi conversations, reflecting the company's shift under new CEO Sean White toward enterprise-focused products.

Unity has launched its Muse AI platform as an early access subscription service to simplify game development, featuring AI tools like a chatbot for coding assistance and sprite creation tools for 2D projects. Future features include AI character animation. Also announced was Sentis, an advanced AI engine, and new cloud tools for collaboration.

Google has launched Project Astra and the upgraded Gemini AI model, showcasing enhanced multimodal capabilities at the Google I/O conference. Gemini enhances AI search, supports real-time voice, and boasts a 1 million token context window. The new text-to-video model, Veo, aims to rival OpenAI's Sora. Google's TPU Trillium chip offers significant performance boosts.