Microsoft Announces Orca-Math, Chatbots Continue To Be Racist, HuggingFace Starts Robotic Project, Huawei Beats Nvidia A100, & More!
Welcome to the AI Search newsletter. Here are the top updates in AI this week.
Microsoft's small language model outperforms larger models in math
Microsoft has developed a small language model called Orca-Math that has been found to outperform larger models on standardized math tests. The researchers conducted tests on the Grade School Math 8K benchmark, which consists of 8,500 grade-school math word problems requiring multistep reasoning to solve. Orca-Math, with its 7 billion parameters, achieved a score of 86.81% on the GSM8K, which is comparable to larger models such as GPT-4-0613. The key difference between Orca-Math and popular large language models like ChatGPT is the former's specialized focus on solving math problems rather than engaging in conversations or answering general questions.
This AI can actually generate text & words CORRECTLY
AI chatbots were racist even after anti-racism training
AI chatbots continue to use racist stereotypes despite anti-racism training, as discovered by researchers. The team conducted experiments on several large language models like GPT-4 and GPT-3.5, revealing that the chatbots displayed negative biases when prompted to comment on text documents in African American English compared to Standard American English. The study found that larger language models exhibited more negative bias towards authors of African American English texts, suggesting a deeper underlying issue within these AI systems.
Murf AI
Introducing Murf, the most versatile AI text-to-speech generator. Create studio-quality voice overs in minutes using lifelike AI voices suitable for podcasts, videos, presentations, and more. Choose from 120+ text to speech voices in 20+ languages, with the ability to add video, music, or images and sync them to the voiceover.
Hugging Face launches new open robotics project
Hugging Face is venturing into the field of robotics with a new open-source project led by Remi Cadene, a former scientist at Tesla. This initiative aims to integrate deep learning and embodied AI into low-cost robotic systems, bridging the gap between machine learning and robotics. This move demonstrates Hugging Face's dedication to pushing the boundaries of robotics and AI. This strategic expansion signals a shift in focus for Hugging Face, from software to hardware, at a time when interest and investment in robotics are on the rise.
Huawei's Ascend 910B outperforms Nvidia's A100 AI chip
Despite facing US sanctions that limit its semiconductor development, Huawei has been bolstering its chip business through partnerships with domestic suppliers. The Ascend 910B, manufactured by SMIC using the 7-nanometer process, reportedly offers AI processing power on par with or better than Nvidia's A100 GPU. Companies like Baidu have already ordered the chip, while iFlytek has launched a computing platform based on Huawei's Ascend chips.
Luminar NEO
Luminar Neo is a new AI-driven creative image editor developed by Skylum Software. It’s designed to make complex editing quick and easy for all levels of photographers, from beginners to pros. Luminar Neo leverages artificial intelligence and 3D depth mapping, offering innovative tools such as the new Portrait Background Removal AI, Mask AI, and the Relight AI tool with 3D Depth Mapping. It can function as a standalone application for macOS and Windows, and can also integrate with Lightroom Classic, Photoshop, Photos for macOS, and Microsoft Photos as a plugin and extension. With Luminar Neo, you can transform your photos into the images you imagined, making creative image editing accessible and fun.
Google Deepmind introduces Genie
Google Deepmind introduced a revolutionary AI called Genie, which is trained on Internet videos to generate interactive and playable virtual worlds from various inputs such as synthetic images, photographs, and sketches. This innovative model, known as a foundation world model, can create endless variations of playable environments without the need for action labels. Genie is a new paradigm in generative AI, allowing users to interact with imagined virtual worlds with just a single image prompt. It has the ability to learn fine-grained controls exclusively from Internet videos, enabling a new generation of creators to easily generate and step into virtual worlds.
People can’t distinguish AI deepfakes 40% of the time
Study finds that people struggle to differentiate between real images and AI-generated ones, with only 61% of participants able to accurately identify the difference. 260 participants were shown a mix of real images and AI-generated images and asked to label them accordingly. Surprisingly, the participants paid attention to details like fingers, teeth, and eyes to distinguish between the two types of images, but their assessments were often incorrect. Lead author Andreea Pocol highlighted the challenge of rapidly advancing AI technology and the potential misuse of AI-generated images for disinformation purposes, such as creating fake images of public figures for malicious intent.
Sony’s new AI bass generator
Sony introduced a new latent diffusion model for generating realistic bass accompaniments that align with the style and tonality of input music tracks. This model, trained on a dataset of bass guitar encodings, can create coherent basslines of any length and allows users to control the timbre and playing style of the generated bass through a technique called 'style grounding.' The researchers envision further development of this tool to include other instruments, such as drums, piano, guitar, strings, and sound effects, to help musicians worldwide enhance their compositions.